b9e876b2bcec19ad9413c7406a81140932a9ab818b77fb491c8f4ac74828cb4f

.gitattributes

@@ -41,3 +41,11 @@ examples/test16.mov filter=lfs diff=lfs merge=lfs -text
 examples/test17.mov filter=lfs diff=lfs merge=lfs -text
 examples/test18.mov filter=lfs diff=lfs merge=lfs -text
 examples/test19.mov filter=lfs diff=lfs merge=lfs -text
+third-party/DPVO/build/lib.win-amd64-3.9/lietorch_backends.cp39-win_amd64.pyd filter=lfs diff=lfs merge=lfs -text
+third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/altcorr/correlation.obj filter=lfs diff=lfs merge=lfs -text
+third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/altcorr/correlation_kernel.obj filter=lfs diff=lfs merge=lfs -text
+third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/fastba/ba.obj filter=lfs diff=lfs merge=lfs -text
+third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/fastba/ba_cuda.obj filter=lfs diff=lfs merge=lfs -text
+third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/lietorch/src/lietorch.obj filter=lfs diff=lfs merge=lfs -text
+third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/lietorch/src/lietorch_cpu.obj filter=lfs diff=lfs merge=lfs -text
+third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/lietorch/src/lietorch_gpu.obj filter=lfs diff=lfs merge=lfs -text

requirements.txt

@@ -0,0 +1,21 @@
+chumpy @ git+https://github.com/mattloper/chumpy
+numpy==1.22.3
+yacs
+joblib
+scikit-image
+opencv-python
+imageio[ffmpeg]
+matplotlib
+tensorboard
+smplx
+progress
+einops
+mmcv==1.3.9
+timm==0.4.9
+munkres
+xtcocotools>=1.8
+loguru
+setuptools==59.5.0
+tqdm
+ultralytics
+gdown==4.6.0

test.py

@@ -0,0 +1,230 @@
+import os
+import argparse
+import os.path as osp
+from glob import glob
+from collections import defaultdict
+
+import cv2
+import torch
+import joblib
+import numpy as np
+from loguru import logger
+from progress.bar import Bar
+
+from configs.config import get_cfg_defaults
+from lib.data.datasets import CustomDataset
+from lib.utils.imutils import avg_preds
+from lib.utils.transforms import matrix_to_axis_angle
+from lib.models import build_network, build_body_model
+from lib.models.preproc.detector import DetectionModel
+from lib.models.preproc.extractor import FeatureExtractor
+from lib.models.smplify import TemporalSMPLify
+
+try:
+    from lib.models.preproc.slam import SLAMModel
+
+    _run_global = True
+except:
+    logger.info('DPVO is not properly installed. Only estimate in local coordinates !')
+    _run_global = False
+
+
+def run(cfg,
+        video,
+        output_pth,
+        network,
+        calib=None,
+        run_global=True,
+        save_pkl=False,
+        visualize=False,
+        run_smplify=False):
+    cap = cv2.VideoCapture(video)
+    assert cap.isOpened(), f'Failed to load video file {video}'
+    fps = cap.get(cv2.CAP_PROP_FPS)
+    length = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    width, height = cap.get(cv2.CAP_PROP_FRAME_WIDTH), cap.get(cv2.CAP_PROP_FRAME_HEIGHT)
+
+    # Whether or not estimating motion in global coordinates
+    run_global = run_global and _run_global
+
+    # Preprocess
+    with torch.no_grad():
+        if not (osp.exists(osp.join(output_pth, 'tracking_results.pth')) and
+                osp.exists(osp.join(output_pth, 'slam_results.pth'))):
+
+            detector = DetectionModel(cfg.DEVICE.lower())
+            extractor = FeatureExtractor(cfg.DEVICE.lower(), cfg.FLIP_EVAL)
+
+            if run_global:
+                slam = SLAMModel(video, output_pth, width, height, calib)
+            else:
+                slam = None
+
+            bar = Bar('Preprocess: 2D detection and SLAM', fill='#', max=length)
+            while (cap.isOpened()):
+                flag, img = cap.read()
+                if not flag: break
+
+                # 2D detection and tracking
+                detector.track(img, fps, length)
+
+                # SLAM
+                if slam is not None:
+                    slam.track()
+
+                bar.next()
+
+            tracking_results = detector.process(fps)
+
+            if slam is not None:
+                slam_results = slam.process()
+            else:
+                slam_results = np.zeros((length, 7))
+                slam_results[:, 3] = 1.0  # Unit quaternion
+
+            # Extract image features
+            # TODO: Merge this into the previous while loop with an online bbox smoothing.
+            tracking_results = extractor.run(video, tracking_results)
+            logger.info('Complete Data preprocessing!')
+
+            # Save the processed data
+            joblib.dump(tracking_results, osp.join(output_pth, 'tracking_results.pth'))
+            joblib.dump(slam_results, osp.join(output_pth, 'slam_results.pth'))
+            logger.info(f'Save processed data at {output_pth}')
+
+        # If the processed data already exists, load the processed data
+        else:
+            tracking_results = joblib.load(osp.join(output_pth, 'tracking_results.pth'))
+            slam_results = joblib.load(osp.join(output_pth, 'slam_results.pth'))
+            logger.info(f'Already processed data exists at {output_pth} ! Load the data .')
+
+    # Build dataset
+    dataset = CustomDataset(cfg, tracking_results, slam_results, width, height, fps)
+
+    # run WHAM
+    results = defaultdict(dict)
+
+    n_subjs = len(dataset)
+    for subj in range(n_subjs):
+
+        with torch.no_grad():
+            if cfg.FLIP_EVAL:
+                # Forward pass with flipped input
+                flipped_batch = dataset.load_data(subj, True)
+                _id, x, inits, features, mask, init_root, cam_angvel, frame_id, kwargs = flipped_batch
+                flipped_pred = network(x, inits, features, mask=mask, init_root=init_root, cam_angvel=cam_angvel,
+                                       return_y_up=True, **kwargs)
+
+                # Forward pass with normal input
+                batch = dataset.load_data(subj)
+                _id, x, inits, features, mask, init_root, cam_angvel, frame_id, kwargs = batch
+                pred = network(x, inits, features, mask=mask, init_root=init_root, cam_angvel=cam_angvel,
+                               return_y_up=True, **kwargs)
+
+                # Merge two predictions
+                flipped_pose, flipped_shape = flipped_pred['pose'].squeeze(0), flipped_pred['betas'].squeeze(0)
+                pose, shape = pred['pose'].squeeze(0), pred['betas'].squeeze(0)
+                flipped_pose, pose = flipped_pose.reshape(-1, 24, 6), pose.reshape(-1, 24, 6)
+                avg_pose, avg_shape = avg_preds(pose, shape, flipped_pose, flipped_shape)
+                avg_pose = avg_pose.reshape(-1, 144)
+                avg_contact = (flipped_pred['contact'][..., [2, 3, 0, 1]] + pred['contact']) / 2
+
+                # Refine trajectory with merged prediction
+                network.pred_pose = avg_pose.view_as(network.pred_pose)
+                network.pred_shape = avg_shape.view_as(network.pred_shape)
+                network.pred_contact = avg_contact.view_as(network.pred_contact)
+                output = network.forward_smpl(**kwargs)
+                pred = network.refine_trajectory(output, cam_angvel, return_y_up=True)
+
+            else:
+                # data
+                batch = dataset.load_data(subj)
+                _id, x, inits, features, mask, init_root, cam_angvel, frame_id, kwargs = batch
+
+                # inference
+                pred = network(x, inits, features, mask=mask, init_root=init_root, cam_angvel=cam_angvel,
+                               return_y_up=True, **kwargs)
+
+        # if False:
+        if run_smplify:
+            smplify = TemporalSMPLify(smpl, img_w=width, img_h=height, device=cfg.DEVICE)
+            input_keypoints = dataset.tracking_results[_id]['keypoints']
+            pred = smplify.fit(pred, input_keypoints, **kwargs)
+
+            with torch.no_grad():
+                network.pred_pose = pred['pose']
+                network.pred_shape = pred['betas']
+                network.pred_cam = pred['cam']
+                output = network.forward_smpl(**kwargs)
+                pred = network.refine_trajectory(output, cam_angvel, return_y_up=True)
+
+        # ========= Store results ========= #
+        pred_body_pose = matrix_to_axis_angle(pred['poses_body']).cpu().numpy().reshape(-1, 69)
+        pred_root = matrix_to_axis_angle(pred['poses_root_cam']).cpu().numpy().reshape(-1, 3)
+        pred_root_world = matrix_to_axis_angle(pred['poses_root_world']).cpu().numpy().reshape(-1, 3)
+        pred_pose = np.concatenate((pred_root, pred_body_pose), axis=-1)
+        pred_pose_world = np.concatenate((pred_root_world, pred_body_pose), axis=-1)
+        pred_trans = (pred['trans_cam'] - network.output.offset).cpu().numpy()
+
+        results[_id]['pose'] = pred_pose
+        results[_id]['trans'] = pred_trans
+        results[_id]['pose_world'] = pred_pose_world
+        results[_id]['trans_world'] = pred['trans_world'].cpu().squeeze(0).numpy()
+        results[_id]['betas'] = pred['betas'].cpu().squeeze(0).numpy()
+        results[_id]['verts'] = (pred['verts_cam'] + pred['trans_cam'].unsqueeze(1)).cpu().numpy()
+        results[_id]['frame_ids'] = frame_id
+
+    if save_pkl:
+        joblib.dump(results, osp.join(output_pth, "wham_output.pkl"))
+
+    # Visualize
+    if visualize:
+        from lib.vis.run_vis import run_vis_on_demo
+        with torch.no_grad():
+            run_vis_on_demo(cfg, video, results, output_pth, network.smpl, vis_global=run_global)
+
+
+if __name__ == '__main__':
+
+    VIDEO_PATH = "examples/test19.mov"
+    OUTPUT_PATH = "output/demo"
+    CALIB_PATH = None
+    ESTIMATE_LOCAL_ONLY = False
+    VISUALIZE = True
+    SAVE_PKL = True
+    RUN_SMPLIFY = False
+    GENDER = 'male'
+
+
+    cfg = get_cfg_defaults()
+    cfg.merge_from_file('configs/yamls/demo.yaml')
+
+    logger.info(f'GPU name -> {torch.cuda.get_device_name()}')
+    logger.info(f'GPU feat -> {torch.cuda.get_device_properties("cuda")}')
+
+    # ========= Load WHAM ========= #
+    smpl_batch_size = cfg.TRAIN.BATCH_SIZE * cfg.DATASET.SEQLEN
+    smpl = build_body_model(device=cfg.DEVICE, gender=GENDER, batch_size=smpl_batch_size)
+    network = build_network(cfg, smpl)
+    network.eval()
+
+    # Output folder
+    sequence = '.'.join(VIDEO_PATH.split('/')[-1].split('.')[:-1])
+    output_pth = osp.join(OUTPUT_PATH, sequence)
+    os.makedirs(output_pth, exist_ok=True)
+
+    faces_np = network.smpl.get_faces()
+    np.save(osp.join(output_pth, f'faces_{GENDER}.npy'), faces_np)
+
+    run(cfg,
+        VIDEO_PATH,
+        output_pth,
+        network,
+        CALIB_PATH,
+        run_global=not ESTIMATE_LOCAL_ONLY,
+        save_pkl=SAVE_PKL,
+        visualize=VISUALIZE,
+        run_smplify=RUN_SMPLIFY)
+
+    print()
+    logger.info('Done !')

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/altcorr/__init__.py

@@ -0,0 +1 @@
+from .correlation import corr, patchify

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/altcorr/correlation.py

@@ -0,0 +1,74 @@
+import torch
+import cuda_corr
+
+class CorrLayer(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, fmap1, fmap2, coords, ii, jj, radius, dropout):
+        """ forward correlation """
+        ctx.save_for_backward(fmap1, fmap2, coords, ii, jj)
+        ctx.radius = radius
+        ctx.dropout = dropout
+        corr, = cuda_corr.forward(fmap1, fmap2, coords, ii, jj, radius)
+
+        return corr
+
+    @staticmethod
+    def backward(ctx, grad):
+        """ backward correlation """
+        fmap1, fmap2, coords, ii, jj = ctx.saved_tensors
+
+        if ctx.dropout < 1:
+            perm = torch.rand(len(ii), device="cuda") < ctx.dropout
+            coords = coords[:,perm]
+            grad = grad[:,perm]
+            ii = ii[perm]
+            jj = jj[perm]
+
+        fmap1_grad, fmap2_grad = \
+            cuda_corr.backward(fmap1, fmap2, coords, ii, jj, grad, ctx.radius)
+
+        return fmap1_grad, fmap2_grad, None, None, None, None, None
+
+
+class PatchLayer(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, net, coords, radius):
+        """ forward patchify """
+        ctx.radius = radius
+        ctx.save_for_backward(net, coords)
+        
+        patches, = cuda_corr.patchify_forward(net, coords, radius)
+        return patches
+
+    @staticmethod
+    def backward(ctx, grad):
+        """ backward patchify """
+        net, coords = ctx.saved_tensors
+        grad, = cuda_corr.patchify_backward(net, coords, grad, ctx.radius)
+
+        return grad, None, None
+
+def patchify(net, coords, radius, mode='bilinear'):
+    """ extract patches """
+
+    patches = PatchLayer.apply(net, coords, radius)
+
+    if mode == 'bilinear':
+        offset = (coords - coords.floor()).to(net.device)
+        dx, dy = offset[:,:,None,None,None].unbind(dim=-1)
+
+        d = 2 * radius + 1
+        x00 = (1-dy) * (1-dx) * patches[...,:d,:d]
+        x01 = (1-dy) * (  dx) * patches[...,:d,1:]
+        x10 = (  dy) * (1-dx) * patches[...,1:,:d]
+        x11 = (  dy) * (  dx) * patches[...,1:,1:]
+
+        return x00 + x01 + x10 + x11
+
+    return patches
+    
+
+def corr(fmap1, fmap2, coords, ii, jj, radius=1, dropout=1):
+    return CorrLayer.apply(fmap1, fmap2, coords, ii, jj, radius, dropout)
+
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/ba.py

@@ -0,0 +1,182 @@
+import torch
+from torch_scatter import scatter_sum
+
+from . import fastba
+from . import lietorch
+from .lietorch import SE3
+
+from .utils import Timer
+
+from . import projective_ops as pops
+
+class CholeskySolver(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, H, b):
+        # don't crash training if cholesky decomp fails
+        U, info = torch.linalg.cholesky_ex(H)
+
+        if torch.any(info):
+            ctx.failed = True
+            return torch.zeros_like(b)
+
+        xs = torch.cholesky_solve(b, U)
+        ctx.save_for_backward(U, xs)
+        ctx.failed = False
+
+        return xs
+
+    @staticmethod
+    def backward(ctx, grad_x):
+        if ctx.failed:
+            return None, None
+
+        U, xs = ctx.saved_tensors
+        dz = torch.cholesky_solve(grad_x, U)
+        dH = -torch.matmul(xs, dz.transpose(-1,-2))
+
+        return dH, dz
+
+# utility functions for scattering ops
+def safe_scatter_add_mat(A, ii, jj, n, m):
+    v = (ii >= 0) & (jj >= 0) & (ii < n) & (jj < m)
+    return scatter_sum(A[:,v], ii[v]*m + jj[v], dim=1, dim_size=n*m)
+
+def safe_scatter_add_vec(b, ii, n):
+    v = (ii >= 0) & (ii < n)
+    return scatter_sum(b[:,v], ii[v], dim=1, dim_size=n)
+
+# apply retraction operator to inv-depth maps
+def disp_retr(disps, dz, ii):
+    ii = ii.to(device=dz.device)
+    return disps + scatter_sum(dz, ii, dim=1, dim_size=disps.shape[1])
+
+# apply retraction operator to poses
+def pose_retr(poses, dx, ii):
+    ii = ii.to(device=dx.device)
+    return poses.retr(scatter_sum(dx, ii, dim=1, dim_size=poses.shape[1]))
+
+def block_matmul(A, B):
+    """ block matrix multiply """
+    b, n1, m1, p1, q1 = A.shape
+    b, n2, m2, p2, q2 = B.shape
+    A = A.permute(0, 1, 3, 2, 4).reshape(b, n1*p1, m1*q1)
+    B = B.permute(0, 1, 3, 2, 4).reshape(b, n2*p2, m2*q2)
+    return torch.matmul(A, B).reshape(b, n1, p1, m2, q2).permute(0, 1, 3, 2, 4)
+
+def block_solve(A, B, ep=1.0, lm=1e-4):
+    """ block matrix solve """
+    b, n1, m1, p1, q1 = A.shape
+    b, n2, m2, p2, q2 = B.shape
+    A = A.permute(0, 1, 3, 2, 4).reshape(b, n1*p1, m1*q1)
+    B = B.permute(0, 1, 3, 2, 4).reshape(b, n2*p2, m2*q2)
+
+    A = A + (ep + lm * A) * torch.eye(n1*p1, device=A.device)
+
+    X = CholeskySolver.apply(A, B)
+    return X.reshape(b, n1, p1, m2, q2).permute(0, 1, 3, 2, 4)
+
+
+def block_show(A):
+    import matplotlib.pyplot as plt
+    b, n1, m1, p1, q1 = A.shape
+    A = A.permute(0, 1, 3, 2, 4).reshape(b, n1*p1, m1*q1)
+    plt.imshow(A[0].detach().cpu().numpy())
+    plt.show()
+
+def BA(poses, patches, intrinsics, targets, weights, lmbda, ii, jj, kk, bounds, ep=100.0, PRINT=False, fixedp=1, structure_only=False):
+    """ bundle adjustment """
+
+    b = 1
+    n = max(ii.max().item(), jj.max().item()) + 1
+
+    coords, v, (Ji, Jj, Jz) = \
+        pops.transform(poses, patches, intrinsics, ii, jj, kk, jacobian=True)
+
+    p = coords.shape[3]
+    r = targets - coords[...,p//2,p//2,:]
+
+    v *= (r.norm(dim=-1) < 250).float()
+
+    in_bounds = \
+        (coords[...,p//2,p//2,0] > bounds[0]) & \
+        (coords[...,p//2,p//2,1] > bounds[1]) & \
+        (coords[...,p//2,p//2,0] < bounds[2]) & \
+        (coords[...,p//2,p//2,1] < bounds[3])
+
+    v *= in_bounds.float()
+
+    if PRINT:
+        print((r * v[...,None]).norm(dim=-1).mean().item())
+
+    r = (v[...,None] * r).unsqueeze(dim=-1)    
+    weights = (v[...,None] * weights).unsqueeze(dim=-1)
+
+    wJiT = (weights * Ji).transpose(2,3)
+    wJjT = (weights * Jj).transpose(2,3)
+    wJzT = (weights * Jz).transpose(2,3)
+
+    Bii = torch.matmul(wJiT, Ji)
+    Bij = torch.matmul(wJiT, Jj)
+    Bji = torch.matmul(wJjT, Ji)
+    Bjj = torch.matmul(wJjT, Jj)
+
+    Eik = torch.matmul(wJiT, Jz)
+    Ejk = torch.matmul(wJjT, Jz)
+
+    vi = torch.matmul(wJiT, r)
+    vj = torch.matmul(wJjT, r)
+
+    # fix first pose
+    ii = ii.clone()
+    jj = jj.clone()
+
+    n = n - fixedp
+    ii = ii - fixedp
+    jj = jj - fixedp
+
+    kx, kk = torch.unique(kk, return_inverse=True, sorted=True)
+    m = len(kx)
+
+    B = safe_scatter_add_mat(Bii, ii, ii, n, n).view(b, n, n, 6, 6) + \
+        safe_scatter_add_mat(Bij, ii, jj, n, n).view(b, n, n, 6, 6) + \
+        safe_scatter_add_mat(Bji, jj, ii, n, n).view(b, n, n, 6, 6) + \
+        safe_scatter_add_mat(Bjj, jj, jj, n, n).view(b, n, n, 6, 6)
+
+    E = safe_scatter_add_mat(Eik, ii, kk, n, m).view(b, n, m, 6, 1) + \
+        safe_scatter_add_mat(Ejk, jj, kk, n, m).view(b, n, m, 6, 1) 
+
+    C = safe_scatter_add_vec(torch.matmul(wJzT, Jz), kk, m)
+
+    v = safe_scatter_add_vec(vi, ii, n).view(b, n, 1, 6, 1) + \
+        safe_scatter_add_vec(vj, jj, n).view(b, n, 1, 6, 1)
+
+    w = safe_scatter_add_vec(torch.matmul(wJzT,  r), kk, m)
+
+    if isinstance(lmbda, torch.Tensor):
+        lmbda = lmbda.reshape(*C.shape)
+        
+    Q = 1.0 / (C + lmbda)
+    
+    ### solve w/ schur complement ###
+    EQ = E * Q[:,None]
+
+    if structure_only or n == 0:
+        dZ = (Q * w).view(b, -1, 1, 1)
+
+    else:
+        S = B - block_matmul(EQ, E.permute(0,2,1,4,3))
+        y = v - block_matmul(EQ, w.unsqueeze(dim=2))
+        dX = block_solve(S, y, ep=ep, lm=1e-4)
+
+        dZ = Q * (w - block_matmul(E.permute(0,2,1,4,3), dX).squeeze(dim=-1))
+        dX = dX.view(b, -1, 6)
+        dZ = dZ.view(b, -1, 1, 1)
+
+    x, y, disps = patches.unbind(dim=2)
+    disps = disp_retr(disps, dZ, kx).clamp(min=1e-3, max=10.0)
+    patches = torch.stack([x, y, disps], dim=2)
+
+    if not structure_only and n > 0:
+        poses = pose_retr(poses, dX, fixedp + torch.arange(n))
+
+    return poses, patches

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/blocks.py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import torch_scatter
+
+class LayerNorm1D(nn.Module):
+    def __init__(self, dim):
+        super(LayerNorm1D, self).__init__()
+        self.norm = nn.LayerNorm(dim, eps=1e-4)
+
+    def forward(self, x):
+        return self.norm(x.transpose(1,2)).transpose(1,2)
+
+class GatedResidual(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+
+        self.gate = nn.Sequential(
+            nn.Linear(dim, dim),
+            nn.Sigmoid())
+
+        self.res = nn.Sequential(
+            nn.Linear(dim, dim),
+            nn.ReLU(inplace=True),
+            nn.Linear(dim, dim))
+
+    def forward(self, x):
+        return x + self.gate(x) * self.res(x)
+
+class SoftAgg(nn.Module):
+    def __init__(self, dim=512, expand=True):
+        super(SoftAgg, self).__init__()
+        self.dim = dim
+        self.expand = expand
+        self.f = nn.Linear(self.dim, self.dim)
+        self.g = nn.Linear(self.dim, self.dim)
+        self.h = nn.Linear(self.dim, self.dim)
+
+    def forward(self, x, ix):
+        _, jx = torch.unique(ix, return_inverse=True)
+        w = torch_scatter.scatter_softmax(self.g(x), jx, dim=1)
+        y = torch_scatter.scatter_sum(self.f(x) * w, jx, dim=1)
+
+        if self.expand:
+            return self.h(y)[:,jx]
+            
+        return self.h(y)
+
+class SoftAggBasic(nn.Module):
+    def __init__(self, dim=512, expand=True):
+        super(SoftAggBasic, self).__init__()
+        self.dim = dim
+        self.expand = expand
+        self.f = nn.Linear(self.dim, self.dim)
+        self.g = nn.Linear(self.dim,        1)
+        self.h = nn.Linear(self.dim, self.dim)
+
+    def forward(self, x, ix):
+        _, jx = torch.unique(ix, return_inverse=True)
+        w = torch_scatter.scatter_softmax(self.g(x), jx, dim=1)
+        y = torch_scatter.scatter_sum(self.f(x) * w, jx, dim=1)
+
+        if self.expand:
+            return self.h(y)[:,jx]
+            
+        return self.h(y)
+
+
+### Gradient Clipping and Zeroing Operations ###
+
+GRAD_CLIP = 0.1
+
+class GradClip(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x):
+        return x
+
+    @staticmethod
+    def backward(ctx, grad_x):
+        grad_x = torch.where(torch.isnan(grad_x), torch.zeros_like(grad_x), grad_x)
+        return grad_x.clamp(min=-0.01, max=0.01)
+
+class GradientClip(nn.Module):
+    def __init__(self):
+        super(GradientClip, self).__init__()
+
+    def forward(self, x):
+        return GradClip.apply(x)
+
+class GradZero(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x):
+        return x
+
+    @staticmethod
+    def backward(ctx, grad_x):
+        grad_x = torch.where(torch.isnan(grad_x), torch.zeros_like(grad_x), grad_x)
+        grad_x = torch.where(torch.abs(grad_x) > GRAD_CLIP, torch.zeros_like(grad_x), grad_x)
+        return grad_x
+
+class GradientZero(nn.Module):
+    def __init__(self):
+        super(GradientZero, self).__init__()
+
+    def forward(self, x):
+        return GradZero.apply(x)
+
+
+class GradMag(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x):
+        return x
+
+    @staticmethod
+    def backward(ctx, grad_x):
+        print(grad_x.abs().mean())
+        return grad_x

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/config.py

@@ -0,0 +1,27 @@
+from yacs.config import CfgNode as CN
+
+_C = CN()
+
+# max number of keyframes
+_C.BUFFER_SIZE = 2048
+
+# bias patch selection towards high gradient regions?
+_C.GRADIENT_BIAS = True
+
+# VO config (increase for better accuracy)
+_C.PATCHES_PER_FRAME = 80
+_C.REMOVAL_WINDOW = 20
+_C.OPTIMIZATION_WINDOW = 12
+_C.PATCH_LIFETIME = 12
+
+# threshold for keyframe removal
+_C.KEYFRAME_INDEX = 4
+_C.KEYFRAME_THRESH = 12.5
+
+# camera motion model
+_C.MOTION_MODEL = 'DAMPED_LINEAR'
+_C.MOTION_DAMPING = 0.5
+
+_C.MIXED_PRECISION = True
+
+cfg = _C

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/data_readers/__init__.py

@@ -0,0 +1 @@
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/data_readers/augmentation.py

@@ -0,0 +1,66 @@
+import torch
+import torchvision.transforms as transforms
+import numpy as np
+import torch.nn.functional as F
+
+
+class RGBDAugmentor:
+    """ perform augmentation on RGB-D video """
+
+    def __init__(self, crop_size):
+        self.crop_size = crop_size
+        self.augcolor = transforms.Compose([
+            transforms.ToPILImage(),
+            transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.2/3.14),
+            transforms.RandomGrayscale(p=0.1),
+            transforms.RandomInvert(p=0.1),
+            transforms.ToTensor()])
+
+        self.max_scale = 0.5
+
+    def spatial_transform(self, images, depths, poses, intrinsics):
+        """ cropping and resizing """
+        ht, wd = images.shape[2:]
+
+        max_scale = self.max_scale
+        min_scale = np.log2(np.maximum(
+            (self.crop_size[0] + 1) / float(ht),
+            (self.crop_size[1] + 1) / float(wd)))
+
+        scale = 1
+        if np.random.rand() < 0.8:
+            scale = 2 ** np.random.uniform(0.0, max_scale)
+
+        intrinsics = scale * intrinsics
+
+        ht1 = int(scale * ht)
+        wd1 = int(scale * wd)
+
+        depths = depths.unsqueeze(dim=1)
+
+        images = F.interpolate(images, (ht1, wd1), mode='bicubic', align_corners=False)
+        depths = F.interpolate(depths, (ht1, wd1), recompute_scale_factor=False)
+
+        # always perform center crop (TODO: try non-center crops)
+        y0 = (images.shape[2] - self.crop_size[0]) // 2
+        x0 = (images.shape[3] - self.crop_size[1]) // 2
+
+        intrinsics = intrinsics - torch.tensor([0.0, 0.0, x0, y0])
+        images = images[:, :, y0:y0+self.crop_size[0], x0:x0+self.crop_size[1]]
+        depths = depths[:, :, y0:y0+self.crop_size[0], x0:x0+self.crop_size[1]]
+
+        depths = depths.squeeze(dim=1)
+        return images, poses, depths, intrinsics
+
+    def color_transform(self, images):
+        """ color jittering """
+        num, ch, ht, wd = images.shape
+        images = images.permute(1, 2, 3, 0).reshape(ch, ht, wd*num)
+        images = 255 * self.augcolor(images[[2,1,0]] / 255.0)
+        return images[[2,1,0]].reshape(ch, ht, wd, num).permute(3,0,1,2).contiguous()
+
+    def __call__(self, images, poses, depths, intrinsics):
+        if np.random.rand() < 0.5:
+            images = self.color_transform(images)
+
+        return self.spatial_transform(images, depths, poses, intrinsics)

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/data_readers/base.py

@@ -0,0 +1,176 @@
+import numpy as np
+import torch
+import torch.utils.data as data
+import torch.nn.functional as F
+
+import csv
+import os
+import cv2
+import math
+import random
+import json
+import pickle
+import os.path as osp
+
+from .augmentation import RGBDAugmentor
+from .rgbd_utils import *
+
+class RGBDDataset(data.Dataset):
+    def __init__(self, name, datapath, n_frames=4, crop_size=[480,640], fmin=10.0, fmax=75.0, aug=True, sample=True):
+        """ Base class for RGBD dataset """
+        self.aug = None
+        self.root = datapath
+        self.name = name
+
+        self.aug = aug
+        self.sample = sample
+
+        self.n_frames = n_frames
+        self.fmin = fmin # exclude very easy examples
+        self.fmax = fmax # exclude very hard examples
+        
+        if self.aug:
+            self.aug = RGBDAugmentor(crop_size=crop_size)
+
+        # building dataset is expensive, cache so only needs to be performed once
+        cur_path = osp.dirname(osp.abspath(__file__))
+        if not os.path.isdir(osp.join(cur_path, 'cache')):
+            os.mkdir(osp.join(cur_path, 'cache'))
+        
+        self.scene_info = \
+            pickle.load(open('datasets/TartanAir.pickle', 'rb'))[0]
+
+        self._build_dataset_index()
+                
+    def _build_dataset_index(self):
+        self.dataset_index = []
+        for scene in self.scene_info:
+            if not self.__class__.is_test_scene(scene):
+                graph = self.scene_info[scene]['graph']
+                for i in graph:
+                    if i < len(graph) - 65:
+                        self.dataset_index.append((scene, i))
+            else:
+                print("Reserving {} for validation".format(scene))
+
+    @staticmethod
+    def image_read(image_file):
+        return cv2.imread(image_file)
+
+    @staticmethod
+    def depth_read(depth_file):
+        return np.load(depth_file)
+
+    def build_frame_graph(self, poses, depths, intrinsics, f=16, max_flow=256):
+        """ compute optical flow distance between all pairs of frames """
+        def read_disp(fn):
+            depth = self.__class__.depth_read(fn)[f//2::f, f//2::f]
+            depth[depth < 0.01] = np.mean(depth)
+            return 1.0 / depth
+
+        poses = np.array(poses)
+        intrinsics = np.array(intrinsics) / f
+        
+        disps = np.stack(list(map(read_disp, depths)), 0)
+        d = f * compute_distance_matrix_flow(poses, disps, intrinsics)
+
+        graph = {}
+        for i in range(d.shape[0]):
+            j, = np.where(d[i] < max_flow)
+            graph[i] = (j, d[i,j])
+
+        return graph
+
+    def __getitem__(self, index):
+        """ return training video """
+
+        index = index % len(self.dataset_index)
+        scene_id, ix = self.dataset_index[index]
+
+        frame_graph = self.scene_info[scene_id]['graph']
+        images_list = self.scene_info[scene_id]['images']
+        depths_list = self.scene_info[scene_id]['depths']
+        poses_list = self.scene_info[scene_id]['poses']
+        intrinsics_list = self.scene_info[scene_id]['intrinsics']
+
+        # stride = np.random.choice([1,2,3])
+
+        d = np.random.uniform(self.fmin, self.fmax)
+        s = 1
+
+        inds = [ ix ]
+
+        while len(inds) < self.n_frames:
+            # get other frames within flow threshold
+
+            if self.sample:
+                k = (frame_graph[ix][1] > self.fmin) & (frame_graph[ix][1] < self.fmax)
+                frames = frame_graph[ix][0][k]
+
+                # prefer frames forward in time
+                if np.count_nonzero(frames[frames > ix]):
+                    ix = np.random.choice(frames[frames > ix])
+
+                elif ix + 1 < len(images_list):
+                    ix = ix + 1
+
+                elif np.count_nonzero(frames):
+                    ix = np.random.choice(frames)
+
+            else:
+                i = frame_graph[ix][0].copy()
+                g = frame_graph[ix][1].copy()
+
+                g[g > d] = -1
+                if s > 0:
+                    g[i <= ix] = -1
+                else:
+                    g[i >= ix] = -1
+
+                if len(g) > 0 and np.max(g) > 0:
+                    ix = i[np.argmax(g)]
+                else:
+                    if ix + s >= len(images_list) or ix + s < 0:
+                        s *= -1
+
+                    ix = ix + s
+            
+            inds += [ ix ]
+
+
+        images, depths, poses, intrinsics = [], [], [], []
+        for i in inds:
+            images.append(self.__class__.image_read(images_list[i]))
+            depths.append(self.__class__.depth_read(depths_list[i]))
+            poses.append(poses_list[i])
+            intrinsics.append(intrinsics_list[i])
+
+        images = np.stack(images).astype(np.float32)
+        depths = np.stack(depths).astype(np.float32)
+        poses = np.stack(poses).astype(np.float32)
+        intrinsics = np.stack(intrinsics).astype(np.float32)
+
+        images = torch.from_numpy(images).float()
+        images = images.permute(0, 3, 1, 2)
+
+        disps = torch.from_numpy(1.0 / depths)
+        poses = torch.from_numpy(poses)
+        intrinsics = torch.from_numpy(intrinsics)
+
+        if self.aug:
+            images, poses, disps, intrinsics = \
+                self.aug(images, poses, disps, intrinsics)
+
+        # normalize depth
+        s = .7 * torch.quantile(disps, .98)
+        disps = disps / s
+        poses[...,:3] *= s
+
+        return images, poses, disps, intrinsics 
+
+    def __len__(self):
+        return len(self.dataset_index)
+
+    def __imul__(self, x):
+        self.dataset_index *= x
+        return self

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/data_readers/factory.py

@@ -0,0 +1,26 @@
+
+import pickle
+import os
+import os.path as osp
+
+# RGBD-Dataset
+from .tartan import TartanAir
+
+def dataset_factory(dataset_list, **kwargs):
+    """ create a combined dataset """
+
+    from torch.utils.data import ConcatDataset
+
+    dataset_map = { 
+        'tartan': (TartanAir, ),
+    }
+
+    db_list = []
+    for key in dataset_list:
+        # cache datasets for faster future loading
+        db = dataset_map[key][0](**kwargs)
+
+        print("Dataset {} has {} images".format(key, len(db)))
+        db_list.append(db)
+
+    return ConcatDataset(db_list)

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/data_readers/frame_utils.py

@@ -0,0 +1,164 @@
+import numpy as np
+from PIL import Image
+from os.path import *
+import re
+import cv2
+cv2.setNumThreads(0)
+
+
+TAG_CHAR = np.array([202021.25], np.float32)
+
+def readFlowKITTI(filename):
+    flow = cv2.imread(filename, cv2.IMREAD_ANYDEPTH|cv2.IMREAD_COLOR)
+    flow = flow[:,:,::-1].astype(np.float32)
+    flow, valid = flow[:, :, :2], flow[:, :, 2]
+    flow = (flow - 2**15) / 64.0
+    return flow, valid
+
+def readFlow(fn):
+    """ Read .flo file in Middlebury format"""
+    # Code adapted from:
+    # http://stackoverflow.com/questions/28013200/reading-middlebury-flow-files-with-python-bytes-array-numpy
+
+    # WARNING: this will work on little-endian architectures (eg Intel x86) only!
+    # print 'fn = %s'%(fn)
+    with open(fn, 'rb') as f:
+        magic = np.fromfile(f, np.float32, count=1)
+        if 202021.25 != magic:
+            print('Magic number incorrect. Invalid .flo file')
+            return None
+        else:
+            w = np.fromfile(f, np.int32, count=1)
+            h = np.fromfile(f, np.int32, count=1)
+            # print 'Reading %d x %d flo file\n' % (w, h)
+            data = np.fromfile(f, np.float32, count=2*int(w)*int(h))
+            # Reshape data into 3D array (columns, rows, bands)
+            # The reshape here is for visualization, the original code is (w,h,2)
+            return np.resize(data, (int(h), int(w), 2))
+
+def readPFM(file):
+    file = open(file, 'rb')
+
+    color = None
+    width = None
+    height = None
+    scale = None
+    endian = None
+
+    header = file.readline().rstrip()
+    if header == b'PF':
+        color = True
+    elif header == b'Pf':
+        color = False
+    else:
+        raise Exception('Not a PFM file.')
+
+    try:
+        dim_match = re.match(rb'^(\d+)\s(\d+)\s$', file.readline())
+    except:
+        dim_match = re.match(r'^(\d+)\s(\d+)\s$', file.readline())
+
+    if dim_match:
+        width, height = map(int, dim_match.groups())
+    else:
+        raise Exception('Malformed PFM header.')
+
+    scale = float(file.readline().rstrip())
+    if scale < 0: # little-endian
+        endian = '<'
+        scale = -scale
+    else:
+        endian = '>' # big-endian
+
+    data = np.fromfile(file, endian + 'f')
+    shape = (height, width, 3) if color else (height, width)
+
+    data = np.reshape(data, shape)
+    data = np.flipud(data)
+    return data
+
+
+def writeFlow(filename,uv,v=None):
+    """ Write optical flow to file.
+    
+    If v is None, uv is assumed to contain both u and v channels,
+    stacked in depth.
+    Original code by Deqing Sun, adapted from Daniel Scharstein.
+    """
+    nBands = 2
+
+    if v is None:
+        assert(uv.ndim == 3)
+        assert(uv.shape[2] == 2)
+        u = uv[:,:,0]
+        v = uv[:,:,1]
+    else:
+        u = uv
+
+    assert(u.shape == v.shape)
+    height,width = u.shape
+    f = open(filename,'wb')
+    # write the header
+    f.write(TAG_CHAR)
+    np.array(width).astype(np.int32).tofile(f)
+    np.array(height).astype(np.int32).tofile(f)
+    # arrange into matrix form
+    tmp = np.zeros((height, width*nBands))
+    tmp[:,np.arange(width)*2] = u
+    tmp[:,np.arange(width)*2 + 1] = v
+    tmp.astype(np.float32).tofile(f)
+    f.close()
+
+
+def readDPT(filename):
+    """ Read depth data from file, return as numpy array. """
+    f = open(filename,'rb')
+    check = np.fromfile(f,dtype=np.float32,count=1)[0]
+    TAG_FLOAT = 202021.25
+    TAG_CHAR = 'PIEH'
+    assert check == TAG_FLOAT, ' depth_read:: Wrong tag in flow file (should be: {0}, is: {1}). Big-endian machine? '.format(TAG_FLOAT,check)
+    width = np.fromfile(f,dtype=np.int32,count=1)[0]
+    height = np.fromfile(f,dtype=np.int32,count=1)[0]
+    size = width*height
+    assert width > 0 and height > 0 and size > 1 and size < 100000000, ' depth_read:: Wrong input size (width = {0}, height = {1}).'.format(width,height)
+    depth = np.fromfile(f,dtype=np.float32,count=-1).reshape((height,width))
+    return depth
+
+def cam_read(filename):
+    """ Read camera data, return (M,N) tuple.
+    M is the intrinsic matrix, N is the extrinsic matrix, so that
+    x = M*N*X,
+    with x being a point in homogeneous image pixel coordinates, X being a
+    point in homogeneous world coordinates."""
+    f = open(filename,'rb')
+    check = np.fromfile(f,dtype=np.float32,count=1)[0]
+    M = np.fromfile(f,dtype='float64',count=9).reshape((3,3))
+    N = np.fromfile(f,dtype='float64',count=12).reshape((3,4))
+
+    E = np.eye(4)
+    E[0:3,:] = N
+
+    fx, fy, cx, cy = M[0,0], M[1,1], M[0,2], M[1,2]
+    kvec = np.array([fx, fy, cx, cy])
+
+    q = Rotation.from_matrix(E[:3,:3]).as_quat()
+    pvec = np.concatenate([E[:3,3], q], 0)
+
+    return pvec, kvec
+
+
+def read_gen(file_name, pil=False):
+    ext = splitext(file_name)[-1]
+    if ext == '.png' or ext == '.jpeg' or ext == '.ppm' or ext == '.jpg':
+        return Image.open(file_name)
+    elif ext == '.bin' or ext == '.raw':
+        return np.load(file_name)
+    elif ext == '.flo':
+        return readFlow(file_name).astype(np.float32)
+    elif ext == '.pfm':
+        return readPFM(file_name).astype(np.float32)
+    elif ext == '.dpt':
+        return readDPT(file_name).astype(np.float32)
+    elif ext == '.cam':
+        return cam_read(file_name)
+    return []

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/data_readers/rgbd_utils.py

@@ -0,0 +1,188 @@
+import numpy as np
+import os.path as osp
+
+import torch
+from ..lietorch import SE3
+
+from scipy.spatial.transform import Rotation
+
+def parse_list(filepath, skiprows=0):
+    """ read list data """
+    data = np.loadtxt(filepath, delimiter=' ', dtype=np.unicode_, skiprows=skiprows)
+    return data
+
+def associate_frames(tstamp_image, tstamp_depth, tstamp_pose, max_dt=1.0):
+    """ pair images, depths, and poses """
+    associations = []
+    for i, t in enumerate(tstamp_image):
+        if tstamp_pose is None:
+            j = np.argmin(np.abs(tstamp_depth - t))
+            if (np.abs(tstamp_depth[j] - t) < max_dt):
+                associations.append((i, j))
+
+        else:
+            j = np.argmin(np.abs(tstamp_depth - t))
+            k = np.argmin(np.abs(tstamp_pose - t))
+        
+            if (np.abs(tstamp_depth[j] - t) < max_dt) and \
+                    (np.abs(tstamp_pose[k] - t) < max_dt):
+                associations.append((i, j, k))
+            
+    return associations
+
+def loadtum(datapath, frame_rate=-1):
+    """ read video data in tum-rgbd format """
+    if osp.isfile(osp.join(datapath, 'groundtruth.txt')):
+        pose_list = osp.join(datapath, 'groundtruth.txt')
+    
+    elif osp.isfile(osp.join(datapath, 'pose.txt')):
+        pose_list = osp.join(datapath, 'pose.txt')
+
+    else:
+        return None, None, None, None
+
+    image_list = osp.join(datapath, 'rgb.txt')
+    depth_list = osp.join(datapath, 'depth.txt')
+
+    calib_path = osp.join(datapath, 'calibration.txt')
+    intrinsic = None
+    if osp.isfile(calib_path):
+        intrinsic = np.loadtxt(calib_path, delimiter=' ')
+        intrinsic = intrinsic.astype(np.float64)
+
+    image_data = parse_list(image_list)
+    depth_data = parse_list(depth_list)
+    pose_data = parse_list(pose_list, skiprows=1)
+    pose_vecs = pose_data[:,1:].astype(np.float64)
+
+    tstamp_image = image_data[:,0].astype(np.float64)
+    tstamp_depth = depth_data[:,0].astype(np.float64)
+    tstamp_pose = pose_data[:,0].astype(np.float64)
+    associations = associate_frames(tstamp_image, tstamp_depth, tstamp_pose)
+
+    # print(len(tstamp_image))
+    # print(len(associations))
+
+    indicies = range(len(associations))[::5]
+
+    # indicies = [ 0 ]
+    # for i in range(1, len(associations)):
+    #     t0 = tstamp_image[associations[indicies[-1]][0]]
+    #     t1 = tstamp_image[associations[i][0]]
+    #     if t1 - t0 > 1.0 / frame_rate:
+    #         indicies += [ i ]
+
+    images, poses, depths, intrinsics, tstamps = [], [], [], [], []
+    for ix in indicies:
+        (i, j, k) = associations[ix]
+        images += [ osp.join(datapath, image_data[i,1]) ]
+        depths += [ osp.join(datapath, depth_data[j,1]) ]
+        poses += [ pose_vecs[k] ]
+        tstamps += [ tstamp_image[i] ]
+        
+        if intrinsic is not None:
+            intrinsics += [ intrinsic ]
+
+    return images, depths, poses, intrinsics, tstamps
+
+
+def all_pairs_distance_matrix(poses, beta=2.5):
+    """ compute distance matrix between all pairs of poses """
+    poses = np.array(poses, dtype=np.float32)
+    poses[:,:3] *= beta # scale to balence rot + trans
+    poses = SE3(torch.from_numpy(poses))
+
+    r = (poses[:,None].inv() * poses[None,:]).log()
+    return r.norm(dim=-1).cpu().numpy()
+
+def pose_matrix_to_quaternion(pose):
+    """ convert 4x4 pose matrix to (t, q) """
+    q = Rotation.from_matrix(pose[:3, :3]).as_quat()
+    return np.concatenate([pose[:3, 3], q], axis=0)
+
+def compute_distance_matrix_flow(poses, disps, intrinsics):
+    """ compute flow magnitude between all pairs of frames """
+    if not isinstance(poses, SE3):
+        poses = torch.from_numpy(poses).float().cuda()[None]
+        poses = SE3(poses).inv()
+
+        disps = torch.from_numpy(disps).float().cuda()[None]
+        intrinsics = torch.from_numpy(intrinsics).float().cuda()[None]
+
+    N = poses.shape[1]
+    
+    ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))
+    ii = ii.reshape(-1).cuda()
+    jj = jj.reshape(-1).cuda()
+
+    MAX_FLOW = 100.0
+    matrix = np.zeros((N, N), dtype=np.float32)
+
+    s = 2048
+    for i in range(0, ii.shape[0], s):
+        flow1, val1 = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])
+        flow2, val2 = pops.induced_flow(poses, disps, intrinsics, jj[i:i+s], ii[i:i+s])
+        
+        flow = torch.stack([flow1, flow2], dim=2)
+        val = torch.stack([val1, val2], dim=2)
+        
+        mag = flow.norm(dim=-1).clamp(max=MAX_FLOW)
+        mag = mag.view(mag.shape[1], -1)
+        val = val.view(val.shape[1], -1)
+
+        mag = (mag * val).mean(-1) / val.mean(-1)
+        mag[val.mean(-1) < 0.7] = np.inf
+
+        i1 = ii[i:i+s].cpu().numpy()
+        j1 = jj[i:i+s].cpu().numpy()
+        matrix[i1, j1] = mag.cpu().numpy()
+
+    return matrix
+
+
+def compute_distance_matrix_flow2(poses, disps, intrinsics, beta=0.4):
+    """ compute flow magnitude between all pairs of frames """
+    # if not isinstance(poses, SE3):
+    #     poses = torch.from_numpy(poses).float().cuda()[None]
+    #     poses = SE3(poses).inv()
+
+    #     disps = torch.from_numpy(disps).float().cuda()[None]
+    #     intrinsics = torch.from_numpy(intrinsics).float().cuda()[None]
+
+    N = poses.shape[1]
+    
+    ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))
+    ii = ii.reshape(-1)
+    jj = jj.reshape(-1)
+
+    MAX_FLOW = 128.0
+    matrix = np.zeros((N, N), dtype=np.float32)
+
+    s = 2048
+    for i in range(0, ii.shape[0], s):
+        flow1a, val1a = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s], tonly=True)
+        flow1b, val1b = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])
+        flow2a, val2a = pops.induced_flow(poses, disps, intrinsics, jj[i:i+s], ii[i:i+s], tonly=True)
+        flow2b, val2b = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])
+
+        flow1 = flow1a + beta * flow1b
+        val1 = val1a * val2b
+
+        flow2 = flow2a + beta * flow2b
+        val2 = val2a * val2b
+        
+        flow = torch.stack([flow1, flow2], dim=2)
+        val = torch.stack([val1, val2], dim=2)
+        
+        mag = flow.norm(dim=-1).clamp(max=MAX_FLOW)
+        mag = mag.view(mag.shape[1], -1)
+        val = val.view(val.shape[1], -1)
+
+        mag = (mag * val).mean(-1) / val.mean(-1)
+        mag[val.mean(-1) < 0.8] = np.inf
+
+        i1 = ii[i:i+s].cpu().numpy()
+        j1 = jj[i:i+s].cpu().numpy()
+        matrix[i1, j1] = mag.cpu().numpy()
+
+    return matrix

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/data_readers/tartan.py

@@ -0,0 +1,110 @@
+
+import numpy as np
+import torch
+import glob
+import cv2
+import os
+import os.path as osp
+
+from ..lietorch import SE3
+from .base import RGBDDataset
+
+# cur_path = osp.dirname(osp.abspath(__file__))
+# test_split = osp.join(cur_path, 'tartan_test.txt')
+# test_split = open(test_split).read().split()
+
+
+test_split = [
+    "abandonedfactory/abandonedfactory/Easy/P011",
+    "abandonedfactory/abandonedfactory/Hard/P011",
+    "abandonedfactory_night/abandonedfactory_night/Easy/P013",
+    "abandonedfactory_night/abandonedfactory_night/Hard/P014",
+    "amusement/amusement/Easy/P008",
+    "amusement/amusement/Hard/P007",
+    "carwelding/carwelding/Easy/P007",
+    "endofworld/endofworld/Easy/P009",
+    "gascola/gascola/Easy/P008",
+    "gascola/gascola/Hard/P009",
+    "hospital/hospital/Easy/P036",
+    "hospital/hospital/Hard/P049",
+    "japanesealley/japanesealley/Easy/P007",
+    "japanesealley/japanesealley/Hard/P005",
+    "neighborhood/neighborhood/Easy/P021",
+    "neighborhood/neighborhood/Hard/P017",
+    "ocean/ocean/Easy/P013",
+    "ocean/ocean/Hard/P009",
+    "office2/office2/Easy/P011",
+    "office2/office2/Hard/P010",
+    "office/office/Hard/P007",
+    "oldtown/oldtown/Easy/P007",
+    "oldtown/oldtown/Hard/P008",
+    "seasidetown/seasidetown/Easy/P009",
+    "seasonsforest/seasonsforest/Easy/P011",
+    "seasonsforest/seasonsforest/Hard/P006",
+    "seasonsforest_winter/seasonsforest_winter/Easy/P009",
+    "seasonsforest_winter/seasonsforest_winter/Hard/P018",
+    "soulcity/soulcity/Easy/P012",
+    "soulcity/soulcity/Hard/P009",
+    "westerndesert/westerndesert/Easy/P013",
+    "westerndesert/westerndesert/Hard/P007",
+]
+
+
+class TartanAir(RGBDDataset):
+
+    # scale depths to balance rot & trans
+    DEPTH_SCALE = 5.0
+
+    def __init__(self, mode='training', **kwargs):
+        self.mode = mode
+        self.n_frames = 2
+        super(TartanAir, self).__init__(name='TartanAir', **kwargs)
+
+    @staticmethod 
+    def is_test_scene(scene):
+        # print(scene, any(x in scene for x in test_split))
+        return any(x in scene for x in test_split)
+
+    def _build_dataset(self):
+        from tqdm import tqdm
+        print("Building TartanAir dataset")
+
+        scene_info = {}
+        scenes = glob.glob(osp.join(self.root, '*/*/*/*'))
+        for scene in tqdm(sorted(scenes)):
+            images = sorted(glob.glob(osp.join(scene, 'image_left/*.png')))
+            depths = sorted(glob.glob(osp.join(scene, 'depth_left/*.npy')))
+
+            if len(images) != len(depths):
+                continue
+            
+            poses = np.loadtxt(osp.join(scene, 'pose_left.txt'), delimiter=' ')
+            poses = poses[:, [1, 2, 0, 4, 5, 3, 6]]
+            poses[:,:3] /= TartanAir.DEPTH_SCALE
+            intrinsics = [TartanAir.calib_read()] * len(images)
+
+            # graph of co-visible frames based on flow
+            graph = self.build_frame_graph(poses, depths, intrinsics)
+
+            scene = '/'.join(scene.split('/'))
+            scene_info[scene] = {'images': images, 'depths': depths, 
+                'poses': poses, 'intrinsics': intrinsics, 'graph': graph}
+
+        return scene_info
+
+    @staticmethod
+    def calib_read():
+        return np.array([320.0, 320.0, 320.0, 240.0])
+
+    @staticmethod
+    def image_read(image_file):
+        return cv2.imread(image_file)
+
+    @staticmethod
+    def depth_read(depth_file):
+        depth = np.load(depth_file) / TartanAir.DEPTH_SCALE
+        depth[depth==np.nan] = 1.0
+        depth[depth==np.inf] = 1.0
+        return depth
+
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/dpvo.py

@@ -0,0 +1,402 @@
+import torch
+import numpy as np
+import torch.nn.functional as F
+
+from . import fastba
+from . import altcorr
+from . import lietorch
+from .lietorch import SE3
+
+from .net import VONet
+from .utils import *
+from . import projective_ops as pops
+
+autocast = torch.cuda.amp.autocast
+Id = SE3.Identity(1, device="cuda")
+
+
+class DPVO:
+    def __init__(self, cfg, network, ht=480, wd=640, viz=False):
+        self.cfg = cfg
+        self.load_weights(network)
+        self.is_initialized = False
+        self.enable_timing = False
+        
+        self.n = 0      # number of frames
+        self.m = 0      # number of patches
+        self.M = self.cfg.PATCHES_PER_FRAME
+        self.N = self.cfg.BUFFER_SIZE
+
+        self.ht = ht    # image height
+        self.wd = wd    # image width
+
+        DIM = self.DIM
+        RES = self.RES
+
+        ### state attributes ###
+        self.tlist = []
+        self.counter = 0
+
+        # dummy image for visualization
+        self.image_ = torch.zeros(self.ht, self.wd, 3, dtype=torch.uint8, device="cpu")
+
+        self.tstamps_ = torch.zeros(self.N, dtype=torch.long, device="cuda")
+        self.poses_ = torch.zeros(self.N, 7, dtype=torch.float, device="cuda")
+        self.patches_ = torch.zeros(self.N, self.M, 3, self.P, self.P, dtype=torch.float, device="cuda")
+        self.intrinsics_ = torch.zeros(self.N, 4, dtype=torch.float, device="cuda")
+
+        self.points_ = torch.zeros(self.N * self.M, 3, dtype=torch.float, device="cuda")
+        self.colors_ = torch.zeros(self.N, self.M, 3, dtype=torch.uint8, device="cuda")
+
+        self.index_ = torch.zeros(self.N, self.M, dtype=torch.long, device="cuda")
+        self.index_map_ = torch.zeros(self.N, dtype=torch.long, device="cuda")
+
+        ### network attributes ###
+        self.mem = 32
+
+        if self.cfg.MIXED_PRECISION:
+            self.kwargs = kwargs = {"device": "cuda", "dtype": torch.half}
+        else:
+            self.kwargs = kwargs = {"device": "cuda", "dtype": torch.float}
+        
+        self.imap_ = torch.zeros(self.mem, self.M, DIM, **kwargs)
+        self.gmap_ = torch.zeros(self.mem, self.M, 128, self.P, self.P, **kwargs)
+
+        ht = ht // RES
+        wd = wd // RES
+
+        self.fmap1_ = torch.zeros(1, self.mem, 128, ht // 1, wd // 1, **kwargs)
+        self.fmap2_ = torch.zeros(1, self.mem, 128, ht // 4, wd // 4, **kwargs)
+
+        # feature pyramid
+        self.pyramid = (self.fmap1_, self.fmap2_)
+
+        self.net = torch.zeros(1, 0, DIM, **kwargs)
+        self.ii = torch.as_tensor([], dtype=torch.long, device="cuda")
+        self.jj = torch.as_tensor([], dtype=torch.long, device="cuda")
+        self.kk = torch.as_tensor([], dtype=torch.long, device="cuda")
+        
+        # initialize poses to identity matrix
+        self.poses_[:,6] = 1.0
+
+        # store relative poses for removed frames
+        self.delta = {}
+
+        self.viewer = None
+        if viz:
+            self.start_viewer()
+
+    def load_weights(self, network):
+        # load network from checkpoint file
+        if isinstance(network, str):
+            from collections import OrderedDict
+            state_dict = torch.load(network)
+            new_state_dict = OrderedDict()
+            for k, v in state_dict.items():
+                if "update.lmbda" not in k:
+                    new_state_dict[k.replace('module.', '')] = v
+            
+            self.network = VONet()
+            self.network.load_state_dict(new_state_dict)
+
+        else:
+            self.network = network
+
+        # steal network attributes
+        self.DIM = self.network.DIM
+        self.RES = self.network.RES
+        self.P = self.network.P
+
+        self.network.cuda()
+        self.network.eval()
+
+        # if self.cfg.MIXED_PRECISION:
+        #     self.network.half()
+
+
+    def start_viewer(self):
+        from dpviewer import Viewer
+
+        intrinsics_ = torch.zeros(1, 4, dtype=torch.float32, device="cuda")
+
+        self.viewer = Viewer(
+            self.image_,
+            self.poses_,
+            self.points_,
+            self.colors_,
+            intrinsics_)
+
+    @property
+    def poses(self):
+        return self.poses_.view(1, self.N, 7)
+
+    @property
+    def patches(self):
+        return self.patches_.view(1, self.N*self.M, 3, 3, 3)
+
+    @property
+    def intrinsics(self):
+        return self.intrinsics_.view(1, self.N, 4)
+
+    @property
+    def ix(self):
+        return self.index_.view(-1)
+
+    @property
+    def imap(self):
+        return self.imap_.view(1, self.mem * self.M, self.DIM)
+
+    @property
+    def gmap(self):
+        return self.gmap_.view(1, self.mem * self.M, 128, 3, 3)
+
+    def get_pose(self, t):
+        if t in self.traj:
+            return SE3(self.traj[t])
+
+        t0, dP = self.delta[t]
+        return dP * self.get_pose(t0)
+
+    def terminate(self):
+        """ interpolate missing poses """
+        self.traj = {}
+        for i in range(self.n):
+            self.traj[self.tstamps_[i].item()] = self.poses_[i]
+
+        poses = [self.get_pose(t) for t in range(self.counter)]
+        poses = lietorch.stack(poses, dim=0)
+        poses = poses.inv().data.cpu().numpy()
+        tstamps = np.array(self.tlist, dtype=np.float)
+
+        if self.viewer is not None:
+            self.viewer.join()
+
+        return poses, tstamps
+
+    def corr(self, coords, indicies=None):
+        """ local correlation volume """
+        ii, jj = indicies if indicies is not None else (self.kk, self.jj)
+        ii1 = ii % (self.M * self.mem)
+        jj1 = jj % (self.mem)
+        corr1 = altcorr.corr(self.gmap, self.pyramid[0], coords / 1, ii1, jj1, 3)
+        corr2 = altcorr.corr(self.gmap, self.pyramid[1], coords / 4, ii1, jj1, 3)
+        return torch.stack([corr1, corr2], -1).view(1, len(ii), -1)
+
+    def reproject(self, indicies=None):
+        """ reproject patch k from i -> j """
+        (ii, jj, kk) = indicies if indicies is not None else (self.ii, self.jj, self.kk)
+        coords = pops.transform(SE3(self.poses), self.patches, self.intrinsics, ii, jj, kk)
+        return coords.permute(0, 1, 4, 2, 3).contiguous()
+
+    def append_factors(self, ii, jj):
+        self.jj = torch.cat([self.jj, jj])
+        self.kk = torch.cat([self.kk, ii])
+        self.ii = torch.cat([self.ii, self.ix[ii]])
+
+        net = torch.zeros(1, len(ii), self.DIM, **self.kwargs)
+        self.net = torch.cat([self.net, net], dim=1)
+
+    def remove_factors(self, m):
+        self.ii = self.ii[~m]
+        self.jj = self.jj[~m]
+        self.kk = self.kk[~m]
+        self.net = self.net[:,~m]
+
+    def motion_probe(self):
+        """ kinda hacky way to ensure enough motion for initialization """
+        kk = torch.arange(self.m-self.M, self.m, device="cuda")
+        jj = self.n * torch.ones_like(kk)
+        ii = self.ix[kk]
+
+        net = torch.zeros(1, len(ii), self.DIM, **self.kwargs)
+        coords = self.reproject(indicies=(ii, jj, kk))
+
+        with autocast(enabled=self.cfg.MIXED_PRECISION):
+            corr = self.corr(coords, indicies=(kk, jj))
+            ctx = self.imap[:,kk % (self.M * self.mem)]
+            net, (delta, weight, _) = \
+                self.network.update(net, ctx, corr, None, ii, jj, kk)
+
+        return torch.quantile(delta.norm(dim=-1).float(), 0.5)
+
+    def motionmag(self, i, j):
+        k = (self.ii == i) & (self.jj == j)
+        ii = self.ii[k]
+        jj = self.jj[k]
+        kk = self.kk[k]
+
+        flow = pops.flow_mag(SE3(self.poses), self.patches, self.intrinsics, ii, jj, kk, beta=0.5)
+        return flow.mean().item()
+
+    def keyframe(self):
+
+        i = self.n - self.cfg.KEYFRAME_INDEX - 1
+        j = self.n - self.cfg.KEYFRAME_INDEX + 1
+        m = self.motionmag(i, j) + self.motionmag(j, i)
+ 
+        if m / 2 < self.cfg.KEYFRAME_THRESH:
+            k = self.n - self.cfg.KEYFRAME_INDEX
+            t0 = self.tstamps_[k-1].item()
+            t1 = self.tstamps_[k].item()
+
+            dP = SE3(self.poses_[k]) * SE3(self.poses_[k-1]).inv()
+            self.delta[t1] = (t0, dP)
+
+            to_remove = (self.ii == k) | (self.jj == k)
+            self.remove_factors(to_remove)
+
+            self.kk[self.ii > k] -= self.M
+            self.ii[self.ii > k] -= 1
+            self.jj[self.jj > k] -= 1
+
+            for i in range(k, self.n-1):
+                self.tstamps_[i] = self.tstamps_[i+1]
+                self.colors_[i] = self.colors_[i+1]
+                self.poses_[i] = self.poses_[i+1]
+                self.patches_[i] = self.patches_[i+1]
+                self.intrinsics_[i] = self.intrinsics_[i+1]
+
+                self.imap_[i%self.mem] = self.imap_[(i+1) % self.mem]
+                self.gmap_[i%self.mem] = self.gmap_[(i+1) % self.mem]
+                self.fmap1_[0,i%self.mem] = self.fmap1_[0,(i+1)%self.mem]
+                self.fmap2_[0,i%self.mem] = self.fmap2_[0,(i+1)%self.mem]
+
+            self.n -= 1
+            self.m-= self.M
+
+        to_remove = self.ix[self.kk] < self.n - self.cfg.REMOVAL_WINDOW
+        self.remove_factors(to_remove)
+
+    def update(self):
+        with Timer("other", enabled=self.enable_timing):
+            coords = self.reproject()
+
+            with autocast(enabled=True):
+                corr = self.corr(coords)
+                ctx = self.imap[:,self.kk % (self.M * self.mem)]
+                self.net, (delta, weight, _) = \
+                    self.network.update(self.net, ctx, corr, None, self.ii, self.jj, self.kk)
+
+            lmbda = torch.as_tensor([1e-4], device="cuda")
+            weight = weight.float()
+            target = coords[...,self.P//2,self.P//2] + delta.float()
+
+        with Timer("BA", enabled=self.enable_timing):
+            t0 = self.n - self.cfg.OPTIMIZATION_WINDOW if self.is_initialized else 1
+            t0 = max(t0, 1)
+
+            try:
+                fastba.BA(self.poses, self.patches, self.intrinsics, 
+                    target, weight, lmbda, self.ii, self.jj, self.kk, t0, self.n, 2)
+            except:
+                print("Warning BA failed...")
+            
+            points = pops.point_cloud(SE3(self.poses), self.patches[:, :self.m], self.intrinsics, self.ix[:self.m])
+            points = (points[...,1,1,:3] / points[...,1,1,3:]).reshape(-1, 3)
+            self.points_[:len(points)] = points[:]
+                
+    def __edges_all(self):
+        return flatmeshgrid(
+            torch.arange(0, self.m, device="cuda"),
+            torch.arange(0, self.n, device="cuda"), indexing='ij')
+
+    def __edges_forw(self):
+        r=self.cfg.PATCH_LIFETIME
+        t0 = self.M * max((self.n - r), 0)
+        t1 = self.M * max((self.n - 1), 0)
+        return flatmeshgrid(
+            torch.arange(t0, t1, device="cuda"),
+            torch.arange(self.n-1, self.n, device="cuda"), indexing='ij')
+
+    def __edges_back(self):
+        r=self.cfg.PATCH_LIFETIME
+        t0 = self.M * max((self.n - 1), 0)
+        t1 = self.M * max((self.n - 0), 0)
+        return flatmeshgrid(torch.arange(t0, t1, device="cuda"),
+            torch.arange(max(self.n-r, 0), self.n, device="cuda"), indexing='ij')
+
+    def __call__(self, tstamp, image, intrinsics):
+        """ track new frame """
+
+        if (self.n+1) >= self.N:
+            raise Exception(f'The buffer size is too small. You can increase it using "--buffer {self.N*2}"')
+
+        if self.viewer is not None:
+            self.viewer.update_image(image)
+
+        image = 2 * (image[None,None] / 255.0) - 0.5
+        
+        with autocast(enabled=self.cfg.MIXED_PRECISION):
+            fmap, gmap, imap, patches, _, clr = \
+                self.network.patchify(image,
+                    patches_per_image=self.cfg.PATCHES_PER_FRAME, 
+                    gradient_bias=self.cfg.GRADIENT_BIAS, 
+                    return_color=True)
+
+        ### update state attributes ###
+        self.tlist.append(tstamp)
+        self.tstamps_[self.n] = self.counter
+        self.intrinsics_[self.n] = intrinsics / self.RES
+
+        # color info for visualization
+        clr = (clr[0,:,[2,1,0]] + 0.5) * (255.0 / 2)
+        self.colors_[self.n] = clr.to(torch.uint8)
+
+        self.index_[self.n + 1] = self.n + 1
+        self.index_map_[self.n + 1] = self.m + self.M
+
+        if self.n > 1:
+            if self.cfg.MOTION_MODEL == 'DAMPED_LINEAR':
+                P1 = SE3(self.poses_[self.n-1])
+                P2 = SE3(self.poses_[self.n-2])
+                
+                xi = self.cfg.MOTION_DAMPING * (P1 * P2.inv()).log()
+                tvec_qvec = (SE3.exp(xi) * P1).data
+                self.poses_[self.n] = tvec_qvec
+            else:
+                tvec_qvec = self.poses[self.n-1]
+                self.poses_[self.n] = tvec_qvec
+
+        # TODO better depth initialization
+        patches[:,:,2] = torch.rand_like(patches[:,:,2,0,0,None,None])
+        if self.is_initialized:
+            s = torch.median(self.patches_[self.n-3:self.n,:,2])
+            patches[:,:,2] = s
+
+        self.patches_[self.n] = patches
+
+        ### update network attributes ###
+        self.imap_[self.n % self.mem] = imap.squeeze()
+        self.gmap_[self.n % self.mem] = gmap.squeeze()
+        self.fmap1_[:, self.n % self.mem] = F.avg_pool2d(fmap[0], 1, 1)
+        self.fmap2_[:, self.n % self.mem] = F.avg_pool2d(fmap[0], 4, 4)
+
+        self.counter += 1        
+        if self.n > 0 and not self.is_initialized:
+            if self.motion_probe() < 2.0:
+                self.delta[self.counter - 1] = (self.counter - 2, Id[0])
+                return
+
+        self.n += 1
+        self.m += self.M
+
+        # relative pose
+        self.append_factors(*self.__edges_forw())
+        self.append_factors(*self.__edges_back())
+
+        if self.n == 8 and not self.is_initialized:
+            self.is_initialized = True            
+
+            for itr in range(12):
+                self.update()
+        
+        elif self.is_initialized:
+            self.update()
+            self.keyframe()
+
+            
+
+
+
+
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/extractor.py

@@ -0,0 +1,264 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class ResidualBlock(nn.Module):
+    def __init__(self, in_planes, planes, norm_fn='group', stride=1):
+        super(ResidualBlock, self).__init__()
+  
+        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, stride=stride)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1)
+        self.relu = nn.ReLU(inplace=True)
+
+        num_groups = planes // 8
+
+        if norm_fn == 'group':
+            self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            if not stride == 1:
+                self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+        
+        elif norm_fn == 'batch':
+            self.norm1 = nn.BatchNorm2d(planes)
+            self.norm2 = nn.BatchNorm2d(planes)
+            if not stride == 1:
+                self.norm3 = nn.BatchNorm2d(planes)
+        
+        elif norm_fn == 'instance':
+            self.norm1 = nn.InstanceNorm2d(planes)
+            self.norm2 = nn.InstanceNorm2d(planes)
+            if not stride == 1:
+                self.norm3 = nn.InstanceNorm2d(planes)
+
+        elif norm_fn == 'none':
+            self.norm1 = nn.Sequential()
+            self.norm2 = nn.Sequential()
+            if not stride == 1:
+                self.norm3 = nn.Sequential()
+
+        if stride == 1:
+            self.downsample = None
+        
+        else:    
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm3)
+
+    def forward(self, x):
+        y = x
+        y = self.relu(self.norm1(self.conv1(y)))
+        y = self.relu(self.norm2(self.conv2(y)))
+
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        return self.relu(x+y)
+
+
+class BottleneckBlock(nn.Module):
+    def __init__(self, in_planes, planes, norm_fn='group', stride=1):
+        super(BottleneckBlock, self).__init__()
+  
+        self.conv1 = nn.Conv2d(in_planes, planes//4, kernel_size=1, padding=0)
+        self.conv2 = nn.Conv2d(planes//4, planes//4, kernel_size=3, padding=1, stride=stride)
+        self.conv3 = nn.Conv2d(planes//4, planes, kernel_size=1, padding=0)
+        self.relu = nn.ReLU(inplace=True)
+
+        num_groups = planes // 8
+
+        if norm_fn == 'group':
+            self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)
+            self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)
+            self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+            if not stride == 1:
+                self.norm4 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
+        
+        elif norm_fn == 'batch':
+            self.norm1 = nn.BatchNorm2d(planes//4)
+            self.norm2 = nn.BatchNorm2d(planes//4)
+            self.norm3 = nn.BatchNorm2d(planes)
+            if not stride == 1:
+                self.norm4 = nn.BatchNorm2d(planes)
+        
+        elif norm_fn == 'instance':
+            self.norm1 = nn.InstanceNorm2d(planes//4)
+            self.norm2 = nn.InstanceNorm2d(planes//4)
+            self.norm3 = nn.InstanceNorm2d(planes)
+            if not stride == 1:
+                self.norm4 = nn.InstanceNorm2d(planes)
+
+        elif norm_fn == 'none':
+            self.norm1 = nn.Sequential()
+            self.norm2 = nn.Sequential()
+            self.norm3 = nn.Sequential()
+            if not stride == 1:
+                self.norm4 = nn.Sequential()
+
+        if stride == 1:
+            self.downsample = None
+        
+        else:    
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm4)
+
+    def forward(self, x):
+        y = x
+        y = self.relu(self.norm1(self.conv1(y)))
+        y = self.relu(self.norm2(self.conv2(y)))
+        y = self.relu(self.norm3(self.conv3(y)))
+
+        if self.downsample is not None:
+            x = self.downsample(x)
+
+        return self.relu(x+y)
+
+DIM=32
+
+class BasicEncoder(nn.Module):
+    def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0, multidim=False):
+        super(BasicEncoder, self).__init__()
+        self.norm_fn = norm_fn
+        self.multidim = multidim
+
+        if self.norm_fn == 'group':
+            self.norm1 = nn.GroupNorm(num_groups=8, num_channels=DIM)
+            
+        elif self.norm_fn == 'batch':
+            self.norm1 = nn.BatchNorm2d(DIM)
+
+        elif self.norm_fn == 'instance':
+            self.norm1 = nn.InstanceNorm2d(DIM)
+
+        elif self.norm_fn == 'none':
+            self.norm1 = nn.Sequential()
+
+        self.conv1 = nn.Conv2d(3, DIM, kernel_size=7, stride=2, padding=3)
+        self.relu1 = nn.ReLU(inplace=True)
+
+        self.in_planes = DIM
+        self.layer1 = self._make_layer(DIM,  stride=1)
+        self.layer2 = self._make_layer(2*DIM, stride=2)
+        self.layer3 = self._make_layer(4*DIM, stride=2)
+
+        # output convolution
+        self.conv2 = nn.Conv2d(4*DIM, output_dim, kernel_size=1)
+
+        if self.multidim:
+            self.layer4 = self._make_layer(256, stride=2)
+            self.layer5 = self._make_layer(512, stride=2)
+
+            self.in_planes = 256
+            self.layer6 = self._make_layer(256, stride=1)
+
+            self.in_planes = 128
+            self.layer7 = self._make_layer(128, stride=1)
+
+            self.up1 = nn.Conv2d(512, 256, 1)
+            self.up2 = nn.Conv2d(256, 128, 1)
+            self.conv3 = nn.Conv2d(128, output_dim, kernel_size=1)
+
+        if dropout > 0:
+            self.dropout = nn.Dropout2d(p=dropout)
+        else:
+            self.dropout = None
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+            elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)):
+                if m.weight is not None:
+                    nn.init.constant_(m.weight, 1)
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, dim, stride=1):
+        layer1 = ResidualBlock(self.in_planes, dim, self.norm_fn, stride=stride)
+        layer2 = ResidualBlock(dim, dim, self.norm_fn, stride=1)
+        layers = (layer1, layer2)
+        
+        self.in_planes = dim
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        b, n, c1, h1, w1 = x.shape
+        x = x.view(b*n, c1, h1, w1)
+
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.relu1(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+
+        x = self.conv2(x)
+
+        _, c2, h2, w2 = x.shape
+        return x.view(b, n, c2, h2, w2)
+
+
+class BasicEncoder4(nn.Module):
+    def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0, multidim=False):
+        super(BasicEncoder4, self).__init__()
+        self.norm_fn = norm_fn
+        self.multidim = multidim
+
+        if self.norm_fn == 'group':
+            self.norm1 = nn.GroupNorm(num_groups=8, num_channels=DIM)
+            
+        elif self.norm_fn == 'batch':
+            self.norm1 = nn.BatchNorm2d(DIM)
+
+        elif self.norm_fn == 'instance':
+            self.norm1 = nn.InstanceNorm2d(DIM)
+
+        elif self.norm_fn == 'none':
+            self.norm1 = nn.Sequential()
+
+        self.conv1 = nn.Conv2d(3, DIM, kernel_size=7, stride=2, padding=3)
+        self.relu1 = nn.ReLU(inplace=True)
+
+        self.in_planes = DIM
+        self.layer1 = self._make_layer(DIM,  stride=1)
+        self.layer2 = self._make_layer(2*DIM, stride=2)
+
+        # output convolution
+        self.conv2 = nn.Conv2d(2*DIM, output_dim, kernel_size=1)
+
+        if dropout > 0:
+            self.dropout = nn.Dropout2d(p=dropout)
+        else:
+            self.dropout = None
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+            elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)):
+                if m.weight is not None:
+                    nn.init.constant_(m.weight, 1)
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, dim, stride=1):
+        layer1 = ResidualBlock(self.in_planes, dim, self.norm_fn, stride=stride)
+        layer2 = ResidualBlock(dim, dim, self.norm_fn, stride=1)
+        layers = (layer1, layer2)
+        
+        self.in_planes = dim
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        b, n, c1, h1, w1 = x.shape
+        x = x.view(b*n, c1, h1, w1)
+
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.relu1(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+
+        x = self.conv2(x)
+
+        _, c2, h2, w2 = x.shape
+        return x.view(b, n, c2, h2, w2)

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/fastba/__init__.py

@@ -0,0 +1 @@
+from .ba import BA, neighbors, reproject

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/fastba/ba.py

@@ -0,0 +1,8 @@
+import torch
+import cuda_ba
+
+neighbors = cuda_ba.neighbors
+reproject = cuda_ba.reproject
+
+def BA(poses, patches, intrinsics, target, weight, lmbda, ii, jj, kk, t0, t1, iterations=2):
+    return cuda_ba.forward(poses.data, patches, intrinsics, target, weight, lmbda, ii, jj, kk, t0, t1, iterations)

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/lietorch/__init__.py

@@ -0,0 +1,2 @@
+__all__ = ['groups']
+from .groups import LieGroupParameter, SO3, RxSO3, SE3, Sim3, cat, stack

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/lietorch/broadcasting.py

@@ -0,0 +1,31 @@
+import torch
+import numpy as np
+
+def check_broadcastable(x, y):
+    assert len(x.shape) == len(y.shape)
+    for (n, m) in zip(x.shape[:-1], y.shape[:-1]):
+        assert n==m or n==1 or m==1
+
+def broadcast_inputs(x, y):
+    """ Automatic broadcasting of missing dimensions """
+    if y is None:
+        xs, xd = x.shape[:-1], x.shape[-1] 
+        return (x.view(-1, xd).contiguous(), ), x.shape[:-1]
+
+    check_broadcastable(x, y)
+
+    xs, xd = x.shape[:-1], x.shape[-1] 
+    ys, yd = y.shape[:-1], y.shape[-1]
+    out_shape = [max(n,m) for (n,m) in zip(xs,ys)]
+
+    if x.shape[:-1] == y.shape[-1]:
+        x1 = x.view(-1, xd)
+        y1 = y.view(-1, yd)
+
+    else:
+        x_expand = [m if n==1 else 1 for (n,m) in zip(xs, ys)]
+        y_expand = [n if m==1 else 1 for (n,m) in zip(xs, ys)]
+        x1 = x.repeat(x_expand + [1]).reshape(-1, xd).contiguous()
+        y1 = y.repeat(y_expand + [1]).reshape(-1, yd).contiguous()
+
+    return (x1, y1), tuple(out_shape)

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/lietorch/gradcheck.py

@@ -0,0 +1,592 @@
+import torch
+
+TORCH_MAJOR = int(torch.__version__.split('.')[0])
+TORCH_MINOR = int(torch.__version__.split('.')[1])
+
+from torch.types import _TensorOrTensors
+if TORCH_MAJOR == 1 and TORCH_MINOR < 8:
+    from torch._six import container_abcs, istuple
+else:
+    import collections.abc as container_abcs
+
+import torch.testing
+from torch.overrides import is_tensor_like
+from itertools import product
+import warnings
+from typing import Callable, Union, Optional, Iterable, List
+
+def zero_gradients(x):
+    if isinstance(x, torch.Tensor):
+        if x.grad is not None:
+            x.grad.detach_()
+            x.grad.zero_()
+    elif isinstance(x, container_abcs.Iterable):
+        for elem in x:
+            zero_gradients(elem)
+
+
+def make_jacobian(input, num_out):
+    if is_tensor_like(input):
+        if not input.is_floating_point() and not input.is_complex():
+            return None
+        if not input.requires_grad:
+            return None
+        return input.new_zeros((input.nelement(), num_out), dtype=input.dtype, layout=torch.strided)
+    elif isinstance(input, container_abcs.Iterable) and not isinstance(input, str):
+        jacobians = list(filter(
+            lambda x: x is not None, (make_jacobian(elem, num_out) for elem in input)))
+        if not jacobians:
+            return None
+        return type(input)(jacobians)  # type: ignore
+    else:
+        return None
+
+
+def iter_tensors(x: Union[torch.Tensor, Iterable[torch.Tensor]], only_requiring_grad: bool = False) -> Iterable[torch.Tensor]:
+    if is_tensor_like(x):
+        # mypy doesn't narrow type of `x` to torch.Tensor
+        if x.requires_grad or not only_requiring_grad:  # type: ignore
+            yield x  # type: ignore
+    elif isinstance(x, container_abcs.Iterable) and not isinstance(x, str):
+        for elem in x:
+            for result in iter_tensors(elem, only_requiring_grad):
+                yield result
+
+def get_numerical_jacobian(fn, input, target=None, eps=1e-3, grad_out=1.0):
+    """
+    input: input to `fn`
+    target: the Tensors wrt whom Jacobians are calculated (default=`input`)
+    grad_out: grad output value used to calculate gradients.
+
+    Note that `target` may not even be part of `input` to `fn`, so please be
+    **very careful** in this to not clone `target`.
+    """
+    if target is None:
+        target = input
+    output_size = fn(input).numel()
+    jacobian = make_jacobian(target, output_size)
+
+    # It's much easier to iterate over flattened lists of tensors.
+    # These are reference to the same objects in jacobian, so any changes
+    # will be reflected in it as well.
+    x_tensors = iter_tensors(target, True)
+    j_tensors = iter_tensors(jacobian)
+
+    def update_jacobians(x, idx, d, d_idx, is_mkldnn=False):
+
+        # compute_jacobian only works for pure real
+        # or pure imaginary delta
+        def compute_gradient(delta):
+            # we currently assume that the norm of delta equals eps
+            assert(delta == eps or delta == (eps * 1j))
+
+            def fn_out():
+                if not is_mkldnn:
+                    # x is a view into input and so this works
+                    return fn(input).clone()
+                else:
+                    # convert the dense tensor back to have mkldnn layout
+                    return fn([x.to_mkldnn()])
+
+            orig = x[idx].item()
+            x[idx] = orig - delta
+            outa = fn_out()
+            x[idx] = orig + delta
+            outb = fn_out()
+            x[idx] = orig
+            r = (outb - outa) / (2 * eps)
+            return r.detach().reshape(-1)
+
+        # for details on the algorithm used here, refer:
+        # Section 3.5.3 https://arxiv.org/pdf/1701.00392.pdf
+        # s = fn(z) where z = x for real valued input
+        # and z = x + yj for complex valued input
+        ds_dx = compute_gradient(eps)
+        if x.is_complex():  # C -> C, C -> R
+            ds_dy = compute_gradient(eps * 1j)
+            # conjugate wirtinger derivative
+            conj_w_d = 0.5 * (ds_dx + ds_dy * 1j)
+            # wirtinger derivative
+            w_d = 0.5 * (ds_dx - ds_dy * 1j)
+            d[d_idx] = grad_out.conjugate() * conj_w_d + grad_out * w_d.conj()
+        elif ds_dx.is_complex():  # R -> C
+            # w_d = conj_w_d = 0.5 * ds_dx
+            # dL_dz_conj = 0.5 * [grad_out.conj() * ds_dx + grad_out * ds_dx.conj()]
+            #            = 0.5 * [grad_out.conj() * ds_dx + (grad_out.conj() * ds_dx).conj()]
+            #            = 0.5 * 2 * real(grad_out.conj() * ds_dx)
+            #            = real(grad_out.conj() * ds_dx)
+            d[d_idx] = torch.real(grad_out.conjugate() * ds_dx)
+        else:   # R -> R
+            d[d_idx] = ds_dx * grad_out
+
+    # TODO: compare structure
+    for x_tensor, d_tensor in zip(x_tensors, j_tensors):
+        if x_tensor.is_sparse:
+            def get_stride(size):
+                dim = len(size)
+                tmp = 1
+                stride = [0] * dim
+                for i in reversed(range(dim)):
+                    stride[i] = tmp
+                    tmp *= size[i]
+                return stride
+
+            x_nnz = x_tensor._nnz()
+            x_size = list(x_tensor.size())
+            x_indices = x_tensor._indices().t()
+            x_values = x_tensor._values()
+            x_stride = get_stride(x_size)
+
+            # Use .data here to get around the version check
+            x_values = x_values.data
+
+            for i in range(x_nnz):
+                x_value = x_values[i]
+                for x_idx in product(*[range(m) for m in x_values.size()[1:]]):
+                    indices = x_indices[i].tolist() + list(x_idx)
+                    d_idx = sum(indices[k] * x_stride[k] for k in range(len(x_size)))
+                    update_jacobians(x_value, x_idx, d_tensor, d_idx)
+        elif x_tensor.layout == torch._mkldnn:  # type: ignore
+            # Use .data here to get around the version check
+            x_tensor = x_tensor.data
+            if len(input) != 1:
+                raise ValueError('gradcheck currently only supports functions with 1 input, but got: ',
+                                 len(input))
+            for d_idx, x_idx in enumerate(product(*[range(m) for m in x_tensor.size()])):
+                # this is really inefficient, but without indexing implemented, there's
+                # not really a better way than converting back and forth
+                x_tensor_dense = x_tensor.to_dense()
+                update_jacobians(x_tensor_dense, x_idx, d_tensor, d_idx, is_mkldnn=True)
+        else:
+            # Use .data here to get around the version check
+            x_tensor = x_tensor.data
+            for d_idx, x_idx in enumerate(product(*[range(m) for m in x_tensor.size()])):
+                update_jacobians(x_tensor, x_idx, d_tensor, d_idx)
+
+    return jacobian
+
+
+def get_analytical_jacobian(input, output, nondet_tol=0.0, grad_out=1.0):
+    # it is easier to call to_dense() on the sparse output than
+    # to modify analytical jacobian
+    if output.is_sparse:
+        raise ValueError('Sparse output is not supported at gradcheck yet. '
+                         'Please call to_dense() on the output of fn for gradcheck.')
+    if output.layout == torch._mkldnn:  # type: ignore
+        raise ValueError('MKLDNN output is not supported at gradcheck yet. '
+                         'Please call to_dense() on the output of fn for gradcheck.')
+    diff_input_list = list(iter_tensors(input, True))
+    jacobian = make_jacobian(input, output.numel())
+    jacobian_reentrant = make_jacobian(input, output.numel())
+    grad_output = torch.zeros_like(output, memory_format=torch.legacy_contiguous_format)
+    flat_grad_output = grad_output.view(-1)
+    reentrant = True
+    correct_grad_sizes = True
+    correct_grad_types = True
+
+    for i in range(flat_grad_output.numel()):
+        flat_grad_output.zero_()
+        flat_grad_output[i] = grad_out
+        for jacobian_c in (jacobian, jacobian_reentrant):
+            grads_input = torch.autograd.grad(output, diff_input_list, grad_output,
+                                              retain_graph=True, allow_unused=True)
+            for jacobian_x, d_x, x in zip(jacobian_c, grads_input, diff_input_list):
+                if d_x is not None and d_x.size() != x.size():
+                    correct_grad_sizes = False
+                elif d_x is not None and d_x.dtype != x.dtype:
+                    correct_grad_types = False
+                elif jacobian_x.numel() != 0:
+                    if d_x is None:
+                        jacobian_x[:, i].zero_()
+                    else:
+                        d_x_dense = d_x.to_dense() if not d_x.layout == torch.strided else d_x
+                        assert jacobian_x[:, i].numel() == d_x_dense.numel()
+                        jacobian_x[:, i] = d_x_dense.contiguous().view(-1)
+
+    for jacobian_x, jacobian_reentrant_x in zip(jacobian, jacobian_reentrant):
+        if jacobian_x.numel() != 0 and (jacobian_x - jacobian_reentrant_x).abs().max() > nondet_tol:
+            reentrant = False
+
+    return jacobian, reentrant, correct_grad_sizes, correct_grad_types
+
+
+def _as_tuple(x):
+    if TORCH_MAJOR == 1 and TORCH_MINOR < 8:
+        b_tuple = istuple(x)  
+    else:
+        b_tuple = isinstance(x, tuple)
+    
+    if b_tuple:
+        return x
+    elif isinstance(x, list):
+        return tuple(x)
+    else:
+        return x,
+    
+
+
+def _differentiable_outputs(x):
+    return tuple(o for o in _as_tuple(x) if o.requires_grad)
+
+
+# Note [VarArg of Tensors]
+# ~~~~~~~~~~~~~~~~~~~~~~~~
+# 'func' accepts a vararg of tensors, which isn't expressable in the type system at the moment.
+# If https://mypy.readthedocs.io/en/latest/additional_features.html?highlight=callable#extended-callable-types is accepted,
+# the '...' first argument of Callable can be replaced with VarArg(Tensor).
+# For now, we permit any input.
+# the '...' first argument of Callable can be replaced with VarArg(Tensor).
+# For now, we permit any input.
+
+def gradcheck(
+    func: Callable[..., Union[_TensorOrTensors]],  # See Note [VarArg of Tensors]
+    inputs: _TensorOrTensors,
+    eps: float = 1e-6,
+    atol: float = 1e-5,
+    rtol: float = 1e-3,
+    raise_exception: bool = True,
+    check_sparse_nnz: bool = False,
+    nondet_tol: float = 0.0,
+    check_undefined_grad: bool = True,
+    check_grad_dtypes: bool = False
+) -> bool:
+    r"""Check gradients computed via small finite differences against analytical
+    gradients w.r.t. tensors in :attr:`inputs` that are of floating point or complex type
+    and with ``requires_grad=True``.
+
+    The check between numerical and analytical gradients uses :func:`~torch.allclose`.
+
+    For complex functions, no notion of Jacobian exists. Gradcheck verifies if the numerical and
+    analytical values of Wirtinger and Conjugate Wirtinger derivative are consistent. The gradient
+    computation is done under the assumption that the overall function has a real valued output.
+    For functions with complex output, gradcheck compares the numerical and analytical gradients
+    for two values of :attr:`grad_output`: 1 and 1j. For more details, check out
+    :ref:`complex_autograd-doc`.
+
+    .. note::
+        The default values are designed for :attr:`input` of double precision.
+        This check will likely fail if :attr:`input` is of less precision, e.g.,
+        ``FloatTensor``.
+
+    .. warning::
+       If any checked tensor in :attr:`input` has overlapping memory, i.e.,
+       different indices pointing to the same memory address (e.g., from
+       :func:`torch.expand`), this check will likely fail because the numerical
+       gradients computed by point perturbation at such indices will change
+       values at all other indices that share the same memory address.
+
+    Args:
+        func (function): a Python function that takes Tensor inputs and returns
+            a Tensor or a tuple of Tensors
+        inputs (tuple of Tensor or Tensor): inputs to the function
+        eps (float, optional): perturbation for finite differences
+        atol (float, optional): absolute tolerance
+        rtol (float, optional): relative tolerance
+        raise_exception (bool, optional): indicating whether to raise an exception if
+            the check fails. The exception gives more information about the
+            exact nature of the failure. This is helpful when debugging gradchecks.
+        check_sparse_nnz (bool, optional): if True, gradcheck allows for SparseTensor input,
+            and for any SparseTensor at input, gradcheck will perform check at nnz positions only.
+        nondet_tol (float, optional): tolerance for non-determinism. When running
+            identical inputs through the differentiation, the results must either match
+            exactly (default, 0.0) or be within this tolerance.
+        check_undefined_grad (bool, options): if True, check if undefined output grads
+            are supported and treated as zeros, for ``Tensor`` outputs.
+
+    Returns:
+        True if all differences satisfy allclose condition
+    """
+    def fail_test(msg):
+        if raise_exception:
+            raise RuntimeError(msg)
+        return False
+
+    tupled_inputs = _as_tuple(inputs)
+    if not check_sparse_nnz and any(t.is_sparse for t in tupled_inputs if isinstance(t, torch.Tensor)):
+        return fail_test('gradcheck expects all tensor inputs are dense when check_sparse_nnz is set to False.')
+
+    # Make sure that gradients are saved for at least one input
+    any_input_requiring_grad = False
+    for idx, inp in enumerate(tupled_inputs):
+        if is_tensor_like(inp) and inp.requires_grad:
+            if not (inp.dtype == torch.float64 or inp.dtype == torch.complex128):
+                warnings.warn(
+                    f'Input #{idx} requires gradient and '
+                    'is not a double precision floating point or complex. '
+                    'This check will likely fail if all the inputs are '
+                    'not of double precision floating point or complex. ')
+            content = inp._values() if inp.is_sparse else inp
+            # TODO: To cover more problematic cases, replace stride = 0 check with
+            # "any overlap in memory" once we have a proper function to check it.
+            if content.layout is not torch._mkldnn:  # type: ignore
+                if not all(st > 0 or sz <= 1 for st, sz in zip(content.stride(), content.size())):
+                    raise RuntimeError(
+                        'The {}th input has a dimension with stride 0. gradcheck only '
+                        'supports inputs that are non-overlapping to be able to '
+                        'compute the numerical gradients correctly. You should call '
+                        '.contiguous on the input before passing it to gradcheck.')
+            any_input_requiring_grad = True
+            inp.retain_grad()
+    if not any_input_requiring_grad:
+        raise ValueError(
+            'gradcheck expects at least one input tensor to require gradient, '
+            'but none of the them have requires_grad=True.')
+
+    func_out = func(*tupled_inputs)
+    output = _differentiable_outputs(func_out)
+
+    if not output:
+        for i, o in enumerate(func_out):
+            def fn(input):
+                return _as_tuple(func(*input))[i]
+            numerical = get_numerical_jacobian(fn, tupled_inputs, eps=eps)
+            for n in numerical:
+                if torch.ne(n, 0).sum() > 0:
+                    return fail_test('Numerical gradient for function expected to be zero')
+        return True
+
+    for i, o in enumerate(output):
+        if not o.requires_grad:
+            continue
+
+        def fn(input):
+            return _as_tuple(func(*input))[i]
+
+        analytical, reentrant, correct_grad_sizes, correct_grad_types = get_analytical_jacobian(tupled_inputs,
+                                                                                                o,
+                                                                                                nondet_tol=nondet_tol)
+        numerical = get_numerical_jacobian(fn, tupled_inputs, eps=eps)
+
+        return analytical, numerical
+
+        out_is_complex = o.is_complex()
+
+        if out_is_complex:
+            # analytical vjp with grad_out = 1.0j
+            analytical_with_imag_grad_out, reentrant_with_imag_grad_out, \
+                correct_grad_sizes_with_imag_grad_out, correct_grad_types_with_imag_grad_out \
+                = get_analytical_jacobian(tupled_inputs, o, nondet_tol=nondet_tol, grad_out=1j)
+            numerical_with_imag_grad_out = get_numerical_jacobian(fn, tupled_inputs, eps=eps, grad_out=1j)
+
+        if not correct_grad_types and check_grad_dtypes:
+            return fail_test('Gradient has dtype mismatch')
+
+        if out_is_complex and not correct_grad_types_with_imag_grad_out and check_grad_dtypes:
+            return fail_test('Gradient (calculated using complex valued grad output) has dtype mismatch')
+
+        if not correct_grad_sizes:
+            return fail_test('Analytical gradient has incorrect size')
+
+        if out_is_complex and not correct_grad_sizes_with_imag_grad_out:
+            return fail_test('Analytical gradient (calculated using complex valued grad output) has incorrect size')
+
+        def checkIfNumericalAnalyticAreClose(a, n, j, error_str=''):
+            if not torch.allclose(a, n, rtol, atol):
+                return fail_test(error_str + 'Jacobian mismatch for output %d with respect to input %d,\n'
+                                 'numerical:%s\nanalytical:%s\n' % (i, j, n, a))
+
+        inp_tensors = iter_tensors(tupled_inputs, True)
+
+        for j, (a, n, inp) in enumerate(zip(analytical, numerical, inp_tensors)):
+            if a.numel() != 0 or n.numel() != 0:
+                if o.is_complex():
+                    # C -> C, R -> C
+                    a_with_imag_grad_out = analytical_with_imag_grad_out[j]
+                    n_with_imag_grad_out = numerical_with_imag_grad_out[j]
+                    checkIfNumericalAnalyticAreClose(a_with_imag_grad_out, n_with_imag_grad_out, j,
+                                                     "Gradients failed to compare equal for grad output = 1j. ")
+                if inp.is_complex():
+                    # C -> R, C -> C
+                    checkIfNumericalAnalyticAreClose(a, n, j,
+                                                     "Gradients failed to compare equal for grad output = 1. ")
+                else:
+                    # R -> R, R -> C
+                    checkIfNumericalAnalyticAreClose(a, n, j)
+
+
+        def not_reentrant_error(error_str=''):
+            error_msg = "Backward" + error_str + " is not reentrant, i.e., running backward with same \
+                        input and grad_output multiple times gives different values, \
+                        although analytical gradient matches numerical gradient. \
+                        The tolerance for nondeterminism was {}.".format(nondet_tol)
+            return fail_test(error_msg)
+
+        if not reentrant:
+            return not_reentrant_error()
+
+        if out_is_complex and not reentrant_with_imag_grad_out:
+            return not_reentrant_error(' (calculated using complex valued grad output)')
+
+    # check if the backward multiplies by grad_output
+    output = _differentiable_outputs(func(*tupled_inputs))
+    if any([o.requires_grad for o in output]):
+        diff_input_list: List[torch.Tensor] = list(iter_tensors(tupled_inputs, True))
+        if not diff_input_list:
+            raise RuntimeError("no Tensors requiring grad found in input")
+        grads_input = torch.autograd.grad(output, diff_input_list,
+                                          [torch.zeros_like(o, memory_format=torch.legacy_contiguous_format) for o in output],
+                                          allow_unused=True)
+        for gi, di in zip(grads_input, diff_input_list):
+            if gi is None:
+                continue
+            if isinstance(gi, torch.Tensor) and gi.layout != torch.strided:
+                if gi.layout != di.layout:
+                    return fail_test('grad is incorrect layout (' + str(gi.layout) + ' is not ' + str(di.layout) + ')')
+                if gi.layout == torch.sparse_coo:
+                    if gi.sparse_dim() != di.sparse_dim():
+                        return fail_test('grad is sparse tensor, but has incorrect sparse_dim')
+                    if gi.dense_dim() != di.dense_dim():
+                        return fail_test('grad is sparse tensor, but has incorrect dense_dim')
+                gi = gi.to_dense()
+                di = di.to_dense()
+            if not gi.eq(0).all():
+                return fail_test('backward not multiplied by grad_output')
+            if gi.dtype != di.dtype or gi.device != di.device or gi.is_sparse != di.is_sparse:
+                return fail_test("grad is incorrect type")
+            if gi.size() != di.size():
+                return fail_test('grad is incorrect size')
+
+        if check_undefined_grad:
+            def warn_bc_breaking():
+                warnings.warn((
+                    'Backwards compatibility: New undefined gradient support checking '
+                    'feature is enabled by default, but it may break existing callers '
+                    'of this function. If this is true for you, you can call this '
+                    'function with "check_undefined_grad=False" to disable the feature'))
+
+            def check_undefined_grad_support(output_to_check):
+                grads_output = [torch.zeros_like(o, memory_format=torch.legacy_contiguous_format) for o in output_to_check]
+                try:
+                    grads_input = torch.autograd.grad(output_to_check,
+                                                      diff_input_list,
+                                                      grads_output,
+                                                      allow_unused=True)
+                except RuntimeError:
+                    warn_bc_breaking()
+                    return fail_test((
+                        'Expected backward function to handle undefined output grads. '
+                        'Please look at "Notes about undefined output gradients" in '
+                        '"tools/autograd/derivatives.yaml"'))
+
+                for gi, i in zip(grads_input, diff_input_list):
+                    if (gi is not None) and (not gi.eq(0).all()):
+                        warn_bc_breaking()
+                        return fail_test((
+                            'Expected all input grads to be undefined or zero when all output grads are undefined '
+                            'or zero. Please look at "Notes about undefined output gradients" in '
+                            '"tools/autograd/derivatives.yaml"'))
+                return True
+
+            # All backward functions must work properly if all output grads are undefined
+            outputs_to_check = [[
+                torch._C._functions.UndefinedGrad()(o) for o in _differentiable_outputs(func(*tupled_inputs))
+                # This check filters out Tensor-likes that aren't instances of Tensor.
+                if isinstance(o, torch.Tensor)
+            ]]
+
+            # If there are multiple output grads, we should be able to undef one at a time without error
+            if len(outputs_to_check[0]) > 1:
+                for undef_grad_idx in range(len(output)):
+                    output_to_check = _differentiable_outputs(func(*tupled_inputs))
+                    outputs_to_check.append([
+                        torch._C._functions.UndefinedGrad()(o) if idx == undef_grad_idx else o
+                        for idx, o in enumerate(output_to_check)])
+
+            for output_to_check in outputs_to_check:
+                if not check_undefined_grad_support(output_to_check):
+                    return False
+
+    return True
+
+
+def gradgradcheck(
+    func: Callable[..., _TensorOrTensors],  # See Note [VarArg of Tensors]
+    inputs: _TensorOrTensors,
+    grad_outputs: Optional[_TensorOrTensors] = None,
+    eps: float = 1e-6,
+    atol: float = 1e-5,
+    rtol: float = 1e-3,
+    gen_non_contig_grad_outputs: bool = False,
+    raise_exception: bool = True,
+    nondet_tol: float = 0.0,
+    check_undefined_grad: bool = True,
+    check_grad_dtypes: bool = False
+) -> bool:
+    r"""Check gradients of gradients computed via small finite differences
+    against analytical gradients w.r.t. tensors in :attr:`inputs` and
+    :attr:`grad_outputs` that are of floating point or complex type and with
+    ``requires_grad=True``.
+
+    This function checks that backpropagating through the gradients computed
+    to the given :attr:`grad_outputs` are correct.
+
+    The check between numerical and analytical gradients uses :func:`~torch.allclose`.
+
+    .. note::
+        The default values are designed for :attr:`input` and
+        :attr:`grad_outputs` of double precision. This check will likely fail if
+        they are of less precision, e.g., ``FloatTensor``.
+
+    .. warning::
+       If any checked tensor in :attr:`input` and :attr:`grad_outputs` has
+       overlapping memory, i.e., different indices pointing to the same memory
+       address (e.g., from :func:`torch.expand`), this check will likely fail
+       because the numerical gradients computed by point perturbation at such
+       indices will change values at all other indices that share the same
+       memory address.
+
+    Args:
+        func (function): a Python function that takes Tensor inputs and returns
+            a Tensor or a tuple of Tensors
+        inputs (tuple of Tensor or Tensor): inputs to the function
+        grad_outputs (tuple of Tensor or Tensor, optional): The gradients with
+            respect to the function's outputs.
+        eps (float, optional): perturbation for finite differences
+        atol (float, optional): absolute tolerance
+        rtol (float, optional): relative tolerance
+        gen_non_contig_grad_outputs (bool, optional): if :attr:`grad_outputs` is
+            ``None`` and :attr:`gen_non_contig_grad_outputs` is ``True``, the
+            randomly generated gradient outputs are made to be noncontiguous
+        raise_exception (bool, optional): indicating whether to raise an exception if
+            the check fails. The exception gives more information about the
+            exact nature of the failure. This is helpful when debugging gradchecks.
+        nondet_tol (float, optional): tolerance for non-determinism. When running
+            identical inputs through the differentiation, the results must either match
+            exactly (default, 0.0) or be within this tolerance. Note that a small amount
+            of nondeterminism in the gradient will lead to larger inaccuracies in
+            the second derivative.
+        check_undefined_grad (bool, options): if True, check if undefined output grads
+            are supported and treated as zeros
+
+    Returns:
+        True if all differences satisfy allclose condition
+    """
+    tupled_inputs = _as_tuple(inputs)
+
+    if grad_outputs is None:
+        # If grad_outputs is not specified, create random Tensors of the same
+        # shape, type, and device as the outputs
+        def randn_like(x):
+            y = torch.testing.randn_like(
+                x if (x.is_floating_point() or x.is_complex()) else x.double(), memory_format=torch.legacy_contiguous_format)
+            if gen_non_contig_grad_outputs:
+                y = torch.testing.make_non_contiguous(y)
+            return y.requires_grad_()
+        outputs = _as_tuple(func(*tupled_inputs))
+        tupled_grad_outputs = tuple(randn_like(x) for x in outputs)
+    else:
+        tupled_grad_outputs = _as_tuple(grad_outputs)
+
+    num_outputs = len(tupled_grad_outputs)
+
+    def new_func(*args):
+        input_args = args[:-num_outputs]
+        grad_outputs = args[-num_outputs:]
+        outputs = _differentiable_outputs(func(*input_args))
+        input_args = tuple(x for x in input_args if isinstance(x, torch.Tensor) and x.requires_grad)
+        grad_inputs = torch.autograd.grad(outputs, input_args, grad_outputs, create_graph=True)
+        return grad_inputs
+
+    return gradcheck(new_func, tupled_inputs + tupled_grad_outputs, eps, atol, rtol, raise_exception,
+                     nondet_tol=nondet_tol, check_undefined_grad=check_undefined_grad,
+                     check_grad_dtypes=check_grad_dtypes)

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/lietorch/group_ops.py

@@ -0,0 +1,102 @@
+import lietorch_backends
+import torch
+import torch.nn.functional as F
+
+
+
+class GroupOp(torch.autograd.Function):
+    """ group operation base class """
+
+    @classmethod
+    def forward(cls, ctx, group_id, *inputs):
+        ctx.group_id = group_id
+        ctx.save_for_backward(*inputs)
+        out = cls.forward_op(ctx.group_id, *inputs)
+        return out
+
+    @classmethod
+    def backward(cls, ctx, grad):
+        error_str = "Backward operation not implemented for {}".format(cls)
+        assert cls.backward_op is not None, error_str
+
+        inputs = ctx.saved_tensors
+        grad = grad.contiguous()
+        grad_inputs = cls.backward_op(ctx.group_id, grad, *inputs)
+        return (None, ) + tuple(grad_inputs)
+        
+
+class Exp(GroupOp):
+    """ exponential map """
+    forward_op, backward_op = lietorch_backends.expm, lietorch_backends.expm_backward
+
+class Log(GroupOp):
+    """ logarithm map """
+    forward_op, backward_op = lietorch_backends.logm, lietorch_backends.logm_backward
+
+class Inv(GroupOp):
+    """ group inverse """
+    forward_op, backward_op = lietorch_backends.inv, lietorch_backends.inv_backward
+
+class Mul(GroupOp):
+    """ group multiplication """
+    forward_op, backward_op = lietorch_backends.mul, lietorch_backends.mul_backward
+
+class Adj(GroupOp):
+    """ adjoint operator """
+    forward_op, backward_op = lietorch_backends.adj, lietorch_backends.adj_backward
+
+class AdjT(GroupOp):
+    """ adjoint operator """
+    forward_op, backward_op = lietorch_backends.adjT, lietorch_backends.adjT_backward
+
+class Act3(GroupOp):
+    """ action on point """
+    forward_op, backward_op = lietorch_backends.act, lietorch_backends.act_backward
+
+class Act4(GroupOp):
+    """ action on point """
+    forward_op, backward_op = lietorch_backends.act4, lietorch_backends.act4_backward
+
+class Jinv(GroupOp):
+    """ adjoint operator """
+    forward_op, backward_op = lietorch_backends.Jinv, None
+
+class ToMatrix(GroupOp):
+    """ convert to matrix representation """
+    forward_op, backward_op = lietorch_backends.as_matrix, None
+
+
+
+
+### conversion operations to/from Euclidean embeddings ###
+
+class FromVec(torch.autograd.Function):
+    """ convert vector into group object """
+
+    @classmethod
+    def forward(cls, ctx, group_id, *inputs):
+        ctx.group_id = group_id
+        ctx.save_for_backward(*inputs)
+        return inputs[0]
+
+    @classmethod
+    def backward(cls, ctx, grad):
+        inputs = ctx.saved_tensors
+        J = lietorch_backends.projector(ctx.group_id, *inputs)
+        return None, torch.matmul(grad.unsqueeze(-2), torch.linalg.pinv(J)).squeeze(-2)
+
+class ToVec(torch.autograd.Function):
+    """ convert group object to vector """
+
+    @classmethod
+    def forward(cls, ctx, group_id, *inputs):
+        ctx.group_id = group_id
+        ctx.save_for_backward(*inputs)
+        return inputs[0]
+
+    @classmethod
+    def backward(cls, ctx, grad):
+        inputs = ctx.saved_tensors
+        J = lietorch_backends.projector(ctx.group_id, *inputs)
+        return None, torch.matmul(grad.unsqueeze(-2), J).squeeze(-2)
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/lietorch/groups.py

@@ -0,0 +1,322 @@
+import torch
+import numpy as np
+
+# group operations implemented in cuda
+from .group_ops import Exp, Log, Inv, Mul, Adj, AdjT, Jinv, Act3, Act4, ToMatrix, ToVec, FromVec
+from .broadcasting import broadcast_inputs
+
+
+class LieGroupParameter(torch.Tensor):
+    """ Wrapper class for LieGroup """
+
+    from torch._C import _disabled_torch_function_impl
+    __torch_function__ = _disabled_torch_function_impl
+
+    def __new__(cls, group, requires_grad=True):
+        data = torch.zeros(group.tangent_shape, 
+                           device=group.data.device, 
+                           dtype=group.data.dtype, 
+                           requires_grad=True)
+
+        return torch.Tensor._make_subclass(cls, data, requires_grad)
+
+    def __init__(self, group):
+        self.group = group
+
+    def retr(self):
+        return self.group.retr(self)
+
+    def log(self):
+        return self.retr().log()
+
+    def inv(self):
+        return self.retr().inv()
+
+    def adj(self, a):
+        return self.retr().adj(a)
+
+    def __mul__(self, other):
+        if isinstance(other, LieGroupParameter):
+            return self.retr() * other.retr()
+        else:
+            return self.retr() * other
+
+    def add_(self, update, alpha):
+        self.group = self.group.exp(alpha*update) * self.group
+
+    def __getitem__(self, index):
+        return self.retr().__getitem__(index)
+
+
+class LieGroup:
+    """ Base class for Lie Group """
+
+    def __init__(self, data):
+        self.data = data
+
+    def __repr__(self):
+        return "{}: size={}, device={}, dtype={}".format(
+            self.group_name, self.shape, self.device, self.dtype)
+
+    @property
+    def shape(self):
+        return self.data.shape[:-1]
+
+    @property
+    def device(self):
+        return self.data.device
+
+    @property
+    def dtype(self):
+        return self.data.dtype
+
+    def vec(self):
+        return self.apply_op(ToVec, self.data)
+
+    @property
+    def tangent_shape(self):
+        return self.data.shape[:-1] + (self.manifold_dim,)
+
+    @classmethod
+    def Identity(cls, *batch_shape, **kwargs):
+        """ Construct identity element with batch shape """
+        
+        if isinstance(batch_shape[0], tuple):
+            batch_shape = batch_shape[0]
+        
+        elif isinstance(batch_shape[0], list):
+            batch_shape = tuple(batch_shape[0])
+
+        numel = np.prod(batch_shape)
+        data = cls.id_elem.reshape(1,-1)
+
+        if 'device' in kwargs:
+            data = data.to(kwargs['device'])
+
+        if 'dtype' in kwargs:
+            data = data.type(kwargs['dtype'])
+
+        data = data.repeat(numel, 1)
+        return cls(data).view(batch_shape)
+
+    @classmethod
+    def IdentityLike(cls, G):
+        return cls.Identity(G.shape, device=G.data.device, dtype=G.data.dtype)
+
+    @classmethod
+    def InitFromVec(cls, data):
+        return cls(cls.apply_op(FromVec, data))
+
+    @classmethod
+    def Random(cls, *batch_shape, sigma=1.0, **kwargs):
+        """ Construct random element with batch_shape by random sampling in tangent space"""
+
+        if isinstance(batch_shape[0], tuple):
+            batch_shape = batch_shape[0]
+        
+        elif isinstance(batch_shape[0], list):
+            batch_shape = tuple(batch_shape[0])
+        
+        tangent_shape = batch_shape + (cls.manifold_dim,)
+        xi = torch.randn(tangent_shape, **kwargs)
+        return cls.exp(sigma * xi)
+
+    @classmethod
+    def apply_op(cls, op, x, y=None):
+        """ Apply group operator """
+        inputs, out_shape = broadcast_inputs(x, y)
+
+        data = op.apply(cls.group_id, *inputs)
+        return data.view(out_shape + (-1,))
+
+    @classmethod
+    def exp(cls, x):
+        """ exponential map: x -> X """
+        return cls(cls.apply_op(Exp, x))
+
+    def quaternion(self):
+        """ extract quaternion """
+        return self.apply_op(Quat, self.data)
+
+    def log(self):
+        """ logarithm map """
+        return self.apply_op(Log, self.data)
+
+    def inv(self):
+        """ group inverse """
+        return self.__class__(self.apply_op(Inv, self.data))
+
+    def mul(self, other):
+        """ group multiplication """
+        return self.__class__(self.apply_op(Mul, self.data, other.data))
+
+    def retr(self, a):
+        """ retraction: Exp(a) * X """
+        dX = self.__class__.apply_op(Exp, a)
+        return self.__class__(self.apply_op(Mul, dX, self.data))
+
+    def adj(self, a):
+        """ adjoint operator: b = A(X) * a """
+        return self.apply_op(Adj, self.data, a)
+
+    def adjT(self, a):
+        """ transposed adjoint operator: b = a * A(X) """
+        return self.apply_op(AdjT, self.data, a)
+
+    def Jinv(self, a):
+        return self.apply_op(Jinv, self.data, a)
+
+    def act(self, p):
+        """ action on a point cloud """
+        
+        # action on point
+        if p.shape[-1] == 3:
+            return self.apply_op(Act3, self.data, p)
+        
+        # action on homogeneous point
+        elif p.shape[-1] == 4:
+            return self.apply_op(Act4, self.data, p)
+
+    def matrix(self):
+        """ convert element to 4x4 matrix """
+        I = torch.eye(4, dtype=self.dtype, device=self.device)
+        I = I.view([1] * (len(self.data.shape) - 1) + [4, 4])
+        return self.__class__(self.data[...,None,:]).act(I).transpose(-1,-2)
+
+    def translation(self):
+        """ extract translation component """
+        p = torch.as_tensor([0.0, 0.0, 0.0, 1.0], dtype=self.dtype, device=self.device)
+        p = p.view([1] * (len(self.data.shape) - 1) + [4,])
+        return self.apply_op(Act4, self.data, p)
+
+    def detach(self):
+        return self.__class__(self.data.detach())
+
+    def view(self, dims):
+        data_reshaped = self.data.view(dims + (self.embedded_dim,))
+        return self.__class__(data_reshaped)
+
+    def __mul__(self, other):
+        # group multiplication
+
+        if isinstance(other, LieGroup):
+            return self.mul(other)
+
+        # action on point
+        elif isinstance(other, torch.Tensor):
+            return self.act(other)
+
+    def __getitem__(self, index):
+        return self.__class__(self.data[index])
+
+    def __setitem__(self, index, item):
+        self.data[index] = item.data
+
+    def to(self, *args, **kwargs):
+        return self.__class__(self.data.to(*args, **kwargs))
+
+    def cpu(self):
+        return self.__class__(self.data.cpu())
+
+    def cuda(self):
+        return self.__class__(self.data.cuda())
+
+    def float(self, device):
+        return self.__class__(self.data.float())
+
+    def double(self, device):
+        return self.__class__(self.data.double())
+
+    def unbind(self, dim=0):
+        return [self.__class__(x) for x in self.data.unbind(dim=dim)]
+        
+
+class SO3(LieGroup):
+    group_name = 'SO3'
+    group_id = 1
+    manifold_dim = 3
+    embedded_dim = 4
+    
+    # unit quaternion
+    id_elem = torch.as_tensor([0.0, 0.0, 0.0, 1.0])
+
+    def __init__(self, data):
+        if isinstance(data, SE3):
+            data = data.data[..., 3:7]
+
+        super(SO3, self).__init__(data)
+
+
+class RxSO3(LieGroup):
+    group_name = 'RxSO3'
+    group_id = 2
+    manifold_dim = 4
+    embedded_dim = 5
+    
+    # unit quaternion
+    id_elem = torch.as_tensor([0.0, 0.0, 0.0, 1.0, 1.0])
+
+    def __init__(self, data):
+        if isinstance(data, Sim3):
+            data = data.data[..., 3:8]
+
+        super(RxSO3, self).__init__(data)
+
+
+class SE3(LieGroup):
+    group_name = 'SE3'
+    group_id = 3
+    manifold_dim = 6
+    embedded_dim = 7
+
+    # translation, unit quaternion
+    id_elem = torch.as_tensor([0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0])
+
+    def __init__(self, data):
+        if isinstance(data, SO3):
+            translation = torch.zeros_like(data.data[...,:3])
+            data = torch.cat([translation, data.data], -1)
+
+        super(SE3, self).__init__(data)
+
+    def scale(self, s):
+        t, q = self.data.split([3,4], -1)
+        t = t * s.unsqueeze(-1)
+        return SE3(torch.cat([t, q], dim=-1))
+
+
+class Sim3(LieGroup):
+    group_name = 'Sim3'
+    group_id = 4
+    manifold_dim = 7
+    embedded_dim = 8
+
+    # translation, unit quaternion, scale
+    id_elem = torch.as_tensor([0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 1.0])
+
+    def __init__(self, data):
+
+        if isinstance(data, SO3):
+            scale = torch.ones_like(SO3.data[...,:1])
+            translation = torch.zeros_like(SO3.data[...,:3])
+            data = torch.cat([translation, SO3.data, scale], -1)
+
+        elif isinstance(data, SE3):
+            scale = torch.ones_like(data.data[...,:1])
+            data = torch.cat([data.data, scale], -1)
+
+        elif isinstance(data, Sim3):
+            data = data.data
+
+        super(Sim3, self).__init__(data)
+
+
+def cat(group_list, dim):
+    """ Concatenate groups along dimension """
+    data = torch.cat([X.data for X in group_list], dim=dim)
+    return group_list[0].__class__(data)
+
+def stack(group_list, dim):
+    """ Concatenate groups along dimension """
+    data = torch.stack([X.data for X in group_list], dim=dim)
+    return group_list[0].__class__(data)

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/lietorch/run_tests.py

@@ -0,0 +1,302 @@
+import torch
+import lietorch
+
+from lietorch import SO3, RxSO3, SE3, Sim3
+from gradcheck import gradcheck, get_analytical_jacobian
+
+
+### forward tests ###
+
+def make_homogeneous(p):
+    return torch.cat([p, torch.ones_like(p[...,:1])], dim=-1)
+
+def matv(A, b):
+    return torch.matmul(A, b[...,None])[..., 0]
+
+def test_exp_log(Group, device='cuda'):
+    """ check Log(Exp(x)) == x """
+    a = .2*torch.randn(2,3,4,5,6,7,Group.manifold_dim, device=device).double()
+    b = Group.exp(a).log()
+    assert torch.allclose(a,b,atol=1e-8), "should be identity"
+    print("\t-", Group, "Passed exp-log test")
+    
+def test_inv(Group, device='cuda'):
+    """ check X * X^{-1} == 0 """
+    X = Group.exp(.1*torch.randn(2,3,4,5,Group.manifold_dim, device=device).double())
+    a = (X * X.inv()).log()
+    assert torch.allclose(a, torch.zeros_like(a), atol=1e-8), "should be 0"
+    print("\t-", Group, "Passed inv test")
+
+def test_adj(Group, device='cuda'):
+    """ check X * Exp(a) == Exp(Adj(X,a)) * X 0 """
+    X = Group.exp(torch.randn(2,3,4,5, Group.manifold_dim, device=device).double())
+    a = torch.randn(2,3,4,5, Group.manifold_dim, device=device).double()
+
+    b = X.adj(a)
+    Y1 = X * Group.exp(a)
+    Y2 = Group.exp(b) * X
+
+    c = (Y1 * Y2.inv()).log()
+    assert torch.allclose(c, torch.zeros_like(c), atol=1e-8), "should be 0"
+    print("\t-", Group, "Passed adj test")
+    
+
+def test_act(Group, device='cuda'):
+    X = Group.exp(torch.randn(1, Group.manifold_dim, device=device).double())
+    p = torch.randn(1,3,device=device).double()
+
+    p1 = X.act(p)
+    p2 = matv(X.matrix(), make_homogeneous(p))
+
+    assert torch.allclose(p1, p2[...,:3], atol=1e-8), "should be 0"
+    print("\t-", Group, "Passed act test")
+
+
+### backward tests ###
+def test_exp_log_grad(Group, device='cuda', tol=1e-8):
+    
+    D = Group.manifold_dim
+
+    def fn(a):
+        return Group.exp(a).log()
+
+    a = torch.zeros(1, Group.manifold_dim, requires_grad=True, device=device).double()
+    analytical, reentrant, correct_grad_sizes, correct_grad_types = \
+        get_analytical_jacobian((a,), fn(a))
+
+    assert torch.allclose(analytical[0], torch.eye(D, device=device).double(), atol=tol)
+
+    a = .2 * torch.randn(1, Group.manifold_dim, requires_grad=True, device=device).double()
+    analytical, reentrant, correct_grad_sizes, correct_grad_types = \
+        get_analytical_jacobian((a,), fn(a))
+
+    assert torch.allclose(analytical[0], torch.eye(D, device=device).double(), atol=tol)
+
+    print("\t-", Group, "Passed eye-grad test")
+
+
+def test_inv_log_grad(Group, device='cuda', tol=1e-8):
+
+    D = Group.manifold_dim
+    X = Group.exp(.2*torch.randn(1,D,device=device).double())
+
+    def fn(a):
+        return (Group.exp(a) * X).inv().log()
+
+    a = torch.zeros(1, D, requires_grad=True, device=device).double()
+    analytical, numerical = gradcheck(fn, [a], eps=1e-4)
+
+    # assert torch.allclose(analytical[0], numerical[0], atol=tol)
+    if not torch.allclose(analytical[0], numerical[0], atol=tol):
+        print(analytical[0])
+        print(numerical[0])
+
+    print("\t-", Group, "Passed inv-grad test")
+
+
+def test_adj_grad(Group, device='cuda'):
+    D = Group.manifold_dim
+    X = Group.exp(.5*torch.randn(1,Group.manifold_dim, device=device).double())
+    
+    def fn(a, b):
+        return (Group.exp(a) * X).adj(b)
+
+    a = torch.zeros(1, D, requires_grad=True, device=device).double()
+    b = torch.randn(1, D, requires_grad=True, device=device).double()
+
+    analytical, numerical = gradcheck(fn, [a, b], eps=1e-4)
+    assert torch.allclose(analytical[0], numerical[0], atol=1e-8)
+    assert torch.allclose(analytical[1], numerical[1], atol=1e-8)
+
+    print("\t-", Group, "Passed adj-grad test")
+
+
+def test_adjT_grad(Group, device='cuda'):
+    D = Group.manifold_dim
+    X = Group.exp(.5*torch.randn(1,Group.manifold_dim, device=device).double())
+    
+    def fn(a, b):
+        return (Group.exp(a) * X).adjT(b)
+
+    a = torch.zeros(1, D, requires_grad=True, device=device).double()
+    b = torch.randn(1, D, requires_grad=True, device=device).double()
+
+    analytical, numerical = gradcheck(fn, [a, b], eps=1e-4)
+
+    assert torch.allclose(analytical[0], numerical[0], atol=1e-8)
+    assert torch.allclose(analytical[1], numerical[1], atol=1e-8)
+
+    print("\t-", Group, "Passed adjT-grad test")
+
+
+def test_act_grad(Group, device='cuda'):
+    D = Group.manifold_dim
+    X = Group.exp(5*torch.randn(1,D, device=device).double())
+    
+    def fn(a, b):
+        return (X*Group.exp(a)).act(b)
+
+    a = torch.zeros(1, D, requires_grad=True, device=device).double()
+    b = torch.randn(1, 3, requires_grad=True, device=device).double()
+
+    analytical, numerical = gradcheck(fn, [a, b], eps=1e-4)
+
+    assert torch.allclose(analytical[0], numerical[0], atol=1e-8)
+    assert torch.allclose(analytical[1], numerical[1], atol=1e-8)
+
+    print("\t-", Group, "Passed act-grad test")
+
+
+def test_matrix_grad(Group, device='cuda'):
+    D = Group.manifold_dim
+    X = Group.exp(torch.randn(1, D, device=device).double())
+    
+    def fn(a):
+        return (Group.exp(a) * X).matrix()
+
+    a = torch.zeros(1, D, requires_grad=True, device=device).double()
+    analytical, numerical = gradcheck(fn, [a], eps=1e-4)
+    assert torch.allclose(analytical[0], numerical[0], atol=1e-6)
+
+    print("\t-", Group, "Passed matrix-grad test")
+
+
+def extract_translation_grad(Group, device='cuda'):
+    """ prototype function """
+
+    D = Group.manifold_dim
+    X = Group.exp(5*torch.randn(1,D, device=device).double())
+    
+    def fn(a):
+        return (Group.exp(a)*X).translation()
+
+    a = torch.zeros(1, D, requires_grad=True, device=device).double()
+
+    analytical, numerical = gradcheck(fn, [a], eps=1e-4)
+
+    assert torch.allclose(analytical[0], numerical[0], atol=1e-8)
+    print("\t-", Group, "Passed translation grad test")
+
+
+def test_vec_grad(Group, device='cuda', tol=1e-6):
+
+    D = Group.manifold_dim
+    X = Group.exp(5*torch.randn(1,D, device=device).double())
+    
+    def fn(a):
+        return (Group.exp(a)*X).vec()
+
+    a = torch.zeros(1, D, requires_grad=True, device=device).double()
+
+    analytical, numerical = gradcheck(fn, [a], eps=1e-4)
+
+    assert torch.allclose(analytical[0], numerical[0], atol=tol)
+    print("\t-", Group, "Passed tovec grad test")
+
+
+def test_fromvec_grad(Group, device='cuda', tol=1e-6):
+
+    def fn(a):
+        if Group == SO3:
+            a = a / a.norm(dim=-1, keepdim=True)
+
+        elif Group == RxSO3:
+            q, s = a.split([4, 1], dim=-1)
+            q = q / q.norm(dim=-1, keepdim=True)
+            a = torch.cat([q, s.exp()], dim=-1)
+
+        elif Group == SE3:
+            t, q = a.split([3, 4], dim=-1)
+            q = q / q.norm(dim=-1, keepdim=True)
+            a = torch.cat([t, q], dim=-1)
+
+        elif Group == Sim3:
+            t, q, s = a.split([3, 4, 1], dim=-1)
+            q = q / q.norm(dim=-1, keepdim=True)
+            a = torch.cat([t, q, s.exp()], dim=-1)
+
+        return Group.InitFromVec(a).vec()
+
+    D = Group.embedded_dim
+    a = torch.randn(1, 2, D, requires_grad=True, device=device).double()
+
+    analytical, numerical = gradcheck(fn, [a], eps=1e-4)
+
+    assert torch.allclose(analytical[0], numerical[0], atol=tol)
+    print("\t-", Group, "Passed fromvec grad test")
+
+
+
+def scale(device='cuda'):
+    
+    def fn(a, s):
+        X = SE3.exp(a)
+        X.scale(s)
+        return X.log()
+
+    s = torch.rand(1, requires_grad=True, device=device).double()
+    a = torch.randn(1, 6, requires_grad=True, device=device).double()
+    
+    analytical, numerical = gradcheck(fn, [a, s], eps=1e-3)
+    print(analytical[1])
+    print(numerical[1])
+
+
+    assert torch.allclose(analytical[0], numerical[0], atol=1e-8)
+    assert torch.allclose(analytical[1], numerical[1], atol=1e-8)
+
+    print("\t-", "Passed se3-to-sim3 test")
+
+    
+if __name__ == '__main__':
+
+
+    print("Testing lietorch forward pass (CPU) ...")
+    for Group in [SO3, RxSO3, SE3, Sim3]:
+        test_exp_log(Group, device='cpu')
+        test_inv(Group, device='cpu')
+        test_adj(Group, device='cpu')
+        test_act(Group, device='cpu')
+
+    print("Testing lietorch backward pass (CPU)...")
+    for Group in [SO3, RxSO3, SE3, Sim3]:
+        if Group == Sim3:
+            tol = 1e-3
+        else:
+            tol = 1e-8
+
+        test_exp_log_grad(Group, device='cpu', tol=tol)
+        test_inv_log_grad(Group, device='cpu', tol=tol)
+        test_adj_grad(Group, device='cpu')
+        test_adjT_grad(Group, device='cpu')
+        test_act_grad(Group, device='cpu')
+        test_matrix_grad(Group, device='cpu')
+        extract_translation_grad(Group, device='cpu')
+        test_vec_grad(Group, device='cpu')
+        test_fromvec_grad(Group, device='cpu')
+
+    print("Testing lietorch forward pass (GPU) ...")
+    for Group in [SO3, RxSO3, SE3, Sim3]:
+        test_exp_log(Group, device='cuda')
+        test_inv(Group, device='cuda')
+        test_adj(Group, device='cuda')
+        test_act(Group, device='cuda')
+
+    print("Testing lietorch backward pass (GPU)...")
+    for Group in [SO3, RxSO3, SE3, Sim3]:
+        if Group == Sim3:
+            tol = 1e-3
+        else:
+            tol = 1e-8
+
+        test_exp_log_grad(Group, device='cuda', tol=tol)
+        test_inv_log_grad(Group, device='cuda', tol=tol)
+        test_adj_grad(Group, device='cuda')
+        test_adjT_grad(Group, device='cuda')
+        test_act_grad(Group, device='cuda')
+        test_matrix_grad(Group, device='cuda')
+        extract_translation_grad(Group, device='cuda')
+        test_vec_grad(Group, device='cuda')
+        test_fromvec_grad(Group, device='cuda')
+
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/logger.py

@@ -0,0 +1,58 @@
+
+import torch
+from torch.utils.tensorboard import SummaryWriter
+
+
+SUM_FREQ = 100
+
+class Logger:
+    def __init__(self, name, scheduler):
+        self.total_steps = 0
+        self.running_loss = {}
+        self.writer = None
+        self.name = name
+        self.scheduler = scheduler
+
+    def _print_training_status(self):
+        if self.writer is None:
+            self.writer = SummaryWriter("runs/{}".format(self.name))
+            print([k for k in self.running_loss])
+
+        lr = self.scheduler.get_lr().pop()
+        metrics_data = [self.running_loss[k]/SUM_FREQ for k in self.running_loss.keys()]
+        training_str = "[{:6d}, {:10.7f}] ".format(self.total_steps+1, lr)
+        metrics_str = ("{:10.4f}, "*len(metrics_data)).format(*metrics_data)
+        
+        # print the training status
+        print(training_str + metrics_str)
+
+        for key in self.running_loss:
+            val = self.running_loss[key] / SUM_FREQ
+            self.writer.add_scalar(key, val, self.total_steps)
+            self.running_loss[key] = 0.0
+
+    def push(self, metrics):
+
+        for key in metrics:
+            if key not in self.running_loss:
+                self.running_loss[key] = 0.0
+
+            self.running_loss[key] += metrics[key]
+
+        if self.total_steps % SUM_FREQ == SUM_FREQ-1:
+            self._print_training_status()
+            self.running_loss = {}
+
+        self.total_steps += 1
+
+    def write_dict(self, results):
+        if self.writer is None:
+            self.writer = SummaryWriter("runs/{}".format(self.name))
+            print([k for k in self.running_loss])
+            
+        for key in results:
+            self.writer.add_scalar(key, results[key], self.total_steps)
+
+    def close(self):
+        self.writer.close()
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/net.py

@@ -0,0 +1,270 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from collections import OrderedDict
+
+import torch_scatter
+from torch_scatter import scatter_sum
+
+from . import fastba
+from . import altcorr
+from . import lietorch
+from .lietorch import SE3
+
+from .extractor import BasicEncoder, BasicEncoder4
+from .blocks import GradientClip, GatedResidual, SoftAgg
+
+from .utils import *
+from .ba import BA
+from . import projective_ops as pops
+
+autocast = torch.cuda.amp.autocast
+import matplotlib.pyplot as plt
+
+DIM = 384
+
+class Update(nn.Module):
+    def __init__(self, p):
+        super(Update, self).__init__()
+
+        self.c1 = nn.Sequential(
+            nn.Linear(DIM, DIM),
+            nn.ReLU(inplace=True),
+            nn.Linear(DIM, DIM))
+
+        self.c2 = nn.Sequential(
+            nn.Linear(DIM, DIM),
+            nn.ReLU(inplace=True),
+            nn.Linear(DIM, DIM))
+        
+        self.norm = nn.LayerNorm(DIM, eps=1e-3)
+
+        self.agg_kk = SoftAgg(DIM)
+        self.agg_ij = SoftAgg(DIM)
+
+        self.gru = nn.Sequential(
+            nn.LayerNorm(DIM, eps=1e-3),
+            GatedResidual(DIM),
+            nn.LayerNorm(DIM, eps=1e-3),
+            GatedResidual(DIM),
+        )
+
+        self.corr = nn.Sequential(
+            nn.Linear(2*49*p*p, DIM),
+            nn.ReLU(inplace=True),
+            nn.Linear(DIM, DIM),
+            nn.LayerNorm(DIM, eps=1e-3),
+            nn.ReLU(inplace=True),
+            nn.Linear(DIM, DIM),
+        )
+
+        self.d = nn.Sequential(
+            nn.ReLU(inplace=False),
+            nn.Linear(DIM, 2),
+            GradientClip())
+
+        self.w = nn.Sequential(
+            nn.ReLU(inplace=False),
+            nn.Linear(DIM, 2),
+            GradientClip(),
+            nn.Sigmoid())
+
+
+    def forward(self, net, inp, corr, flow, ii, jj, kk):
+        """ update operator """
+
+        net = net + inp + self.corr(corr)
+        net = self.norm(net)
+
+        ix, jx = fastba.neighbors(kk, jj)
+        mask_ix = (ix >= 0).float().reshape(1, -1, 1)
+        mask_jx = (jx >= 0).float().reshape(1, -1, 1)
+
+        net = net + self.c1(mask_ix * net[:,ix])
+        net = net + self.c2(mask_jx * net[:,jx])
+
+        net = net + self.agg_kk(net, kk)
+        net = net + self.agg_ij(net, ii*12345 + jj)
+
+        net = self.gru(net)
+
+        return net, (self.d(net), self.w(net), None)
+
+
+class Patchifier(nn.Module):
+    def __init__(self, patch_size=3):
+        super(Patchifier, self).__init__()
+        self.patch_size = patch_size
+        self.fnet = BasicEncoder4(output_dim=128, norm_fn='instance')
+        self.inet = BasicEncoder4(output_dim=DIM, norm_fn='none')
+
+    def __image_gradient(self, images):
+        gray = ((images + 0.5) * (255.0 / 2)).sum(dim=2)
+        dx = gray[...,:-1,1:] - gray[...,:-1,:-1]
+        dy = gray[...,1:,:-1] - gray[...,:-1,:-1]
+        g = torch.sqrt(dx**2 + dy**2)
+        g = F.avg_pool2d(g, 4, 4)
+        return g
+
+    def forward(self, images, patches_per_image=80, disps=None, gradient_bias=False, return_color=False):
+        """ extract patches from input images """
+        fmap = self.fnet(images) / 4.0
+        imap = self.inet(images) / 4.0
+
+        b, n, c, h, w = fmap.shape
+        P = self.patch_size
+
+        # bias patch selection towards regions with high gradient
+        if gradient_bias:
+            g = self.__image_gradient(images)
+            x = torch.randint(1, w-1, size=[n, 3*patches_per_image], device="cuda")
+            y = torch.randint(1, h-1, size=[n, 3*patches_per_image], device="cuda")
+
+            coords = torch.stack([x, y], dim=-1).float()
+            g = altcorr.patchify(g[0,:,None], coords, 0).view(n, 3 * patches_per_image)
+            
+            ix = torch.argsort(g, dim=1)
+            x = torch.gather(x, 1, ix[:, -patches_per_image:])
+            y = torch.gather(y, 1, ix[:, -patches_per_image:])
+
+        else:
+            x = torch.randint(1, w-1, size=[n, patches_per_image], device="cuda")
+            y = torch.randint(1, h-1, size=[n, patches_per_image], device="cuda")
+        
+        coords = torch.stack([x, y], dim=-1).float()
+        imap = altcorr.patchify(imap[0], coords, 0).view(b, -1, DIM, 1, 1)
+        gmap = altcorr.patchify(fmap[0], coords, P//2).view(b, -1, 128, P, P)
+
+        if return_color:
+            clr = altcorr.patchify(images[0], 4*(coords + 0.5), 0).view(b, -1, 3)
+
+        if disps is None:
+            disps = torch.ones(b, n, h, w, device="cuda")
+
+        grid, _ = coords_grid_with_index(disps, device=fmap.device)
+        patches = altcorr.patchify(grid[0], coords, P//2).view(b, -1, 3, P, P)
+
+        index = torch.arange(n, device="cuda").view(n, 1)
+        index = index.repeat(1, patches_per_image).reshape(-1)
+
+        if return_color:
+            return fmap, gmap, imap, patches, index, clr
+
+        return fmap, gmap, imap, patches, index
+
+
+class CorrBlock:
+    def __init__(self, fmap, gmap, radius=3, dropout=0.2, levels=[1,4]):
+        self.dropout = dropout
+        self.radius = radius
+        self.levels = levels
+
+        self.gmap = gmap
+        self.pyramid = pyramidify(fmap, lvls=levels)
+
+    def __call__(self, ii, jj, coords):
+        corrs = []
+        for i in range(len(self.levels)):
+            corrs += [ altcorr.corr(self.gmap, self.pyramid[i], coords / self.levels[i], ii, jj, self.radius, self.dropout) ]
+        return torch.stack(corrs, -1).view(1, len(ii), -1)
+
+
+class VONet(nn.Module):
+    def __init__(self, use_viewer=False):
+        super(VONet, self).__init__()
+        self.P = 3
+        self.patchify = Patchifier(self.P)
+        self.update = Update(self.P)
+
+        self.DIM = DIM
+        self.RES = 4
+
+
+    @autocast(enabled=False)
+    def forward(self, images, poses, disps, intrinsics, M=1024, STEPS=12, P=1, structure_only=False, rescale=False):
+        """ Estimates SE3 or Sim3 between pair of frames """
+
+        images = 2 * (images / 255.0) - 0.5
+        intrinsics = intrinsics / 4.0
+        disps = disps[:, :, 1::4, 1::4].float()
+
+        fmap, gmap, imap, patches, ix = self.patchify(images, disps=disps)
+
+        corr_fn = CorrBlock(fmap, gmap)
+
+        b, N, c, h, w = fmap.shape
+        p = self.P
+
+        patches_gt = patches.clone()
+        Ps = poses
+
+        d = patches[..., 2, p//2, p//2]
+        patches = set_depth(patches, torch.rand_like(d))
+
+        kk, jj = flatmeshgrid(torch.where(ix < 8)[0], torch.arange(0,8, device="cuda"))
+        ii = ix[kk]
+
+        imap = imap.view(b, -1, DIM)
+        net = torch.zeros(b, len(kk), DIM, device="cuda", dtype=torch.float)
+        
+        Gs = SE3.IdentityLike(poses)
+
+        if structure_only:
+            Gs.data[:] = poses.data[:]
+
+        traj = []
+        bounds = [-64, -64, w + 64, h + 64]
+        
+        while len(traj) < STEPS:
+            Gs = Gs.detach()
+            patches = patches.detach()
+
+            n = ii.max() + 1
+            if len(traj) >= 8 and n < images.shape[1]:
+                if not structure_only: Gs.data[:,n] = Gs.data[:,n-1]
+                kk1, jj1 = flatmeshgrid(torch.where(ix  < n)[0], torch.arange(n, n+1, device="cuda"))
+                kk2, jj2 = flatmeshgrid(torch.where(ix == n)[0], torch.arange(0, n+1, device="cuda"))
+
+                ii = torch.cat([ix[kk1], ix[kk2], ii])
+                jj = torch.cat([jj1, jj2, jj])
+                kk = torch.cat([kk1, kk2, kk])
+
+                net1 = torch.zeros(b, len(kk1) + len(kk2), DIM, device="cuda")
+                net = torch.cat([net1, net], dim=1)
+
+                if np.random.rand() < 0.1:
+                    k = (ii != (n - 4)) & (jj != (n - 4))
+                    ii = ii[k]
+                    jj = jj[k]
+                    kk = kk[k]
+                    net = net[:,k]
+
+                patches[:,ix==n,2] = torch.median(patches[:,(ix == n-1) | (ix == n-2),2])
+                n = ii.max() + 1
+
+            coords = pops.transform(Gs, patches, intrinsics, ii, jj, kk)
+            coords1 = coords.permute(0, 1, 4, 2, 3).contiguous()
+
+            corr = corr_fn(kk, jj, coords1)
+            net, (delta, weight, _) = self.update(net, imap[:,kk], corr, None, ii, jj, kk)
+
+            lmbda = 1e-4
+            target = coords[...,p//2,p//2,:] + delta
+
+            ep = 10
+            for itr in range(2):
+                Gs, patches = BA(Gs, patches, intrinsics, target, weight, lmbda, ii, jj, kk, 
+                    bounds, ep=ep, fixedp=1, structure_only=structure_only)
+
+            kl = torch.as_tensor(0)
+            dij = (ii - jj).abs()
+            k = (dij > 0) & (dij <= 2)
+
+            coords = pops.transform(Gs, patches, intrinsics, ii[k], jj[k], kk[k])
+            coords_gt, valid, _ = pops.transform(Ps, patches_gt, intrinsics, ii[k], jj[k], kk[k], jacobian=True)
+
+            traj.append((valid, coords, coords_gt, Gs[:,:n], Ps[:,:n], kl))
+
+        return traj
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/plot_utils.py

@@ -0,0 +1,52 @@
+from copy import deepcopy
+
+import matplotlib.pyplot as plt
+import numpy as np
+from evo.core import sync
+from evo.core.trajectory import PoseTrajectory3D
+from evo.tools import plot
+from pathlib import Path
+
+
+def make_traj(args) -> PoseTrajectory3D:
+    if isinstance(args, tuple):
+        traj, tstamps = args
+        return PoseTrajectory3D(positions_xyz=traj[:,:3], orientations_quat_wxyz=traj[:,3:], timestamps=tstamps)
+    assert isinstance(args, PoseTrajectory3D), type(args)
+    return deepcopy(args)
+
+def best_plotmode(traj):
+    _, i1, i2 = np.argsort(np.var(traj.positions_xyz, axis=0))
+    plot_axes = "xyz"[i2] + "xyz"[i1]
+    return getattr(plot.PlotMode, plot_axes)
+
+def plot_trajectory(pred_traj, gt_traj=None, title="", filename="", align=True, correct_scale=True):
+    pred_traj = make_traj(pred_traj)
+
+    if gt_traj is not None:
+        gt_traj = make_traj(gt_traj)
+        gt_traj, pred_traj = sync.associate_trajectories(gt_traj, pred_traj)
+
+        if align:
+            pred_traj.align(gt_traj, correct_scale=correct_scale)
+
+    plot_collection = plot.PlotCollection("PlotCol")
+    fig = plt.figure(figsize=(8, 8))
+    plot_mode = best_plotmode(gt_traj if (gt_traj is not None) else pred_traj)
+    ax = plot.prepare_axis(fig, plot_mode)
+    ax.set_title(title)
+    if gt_traj is not None:
+        plot.traj(ax, plot_mode, gt_traj, '--', 'gray', "Ground Truth")
+    plot.traj(ax, plot_mode, pred_traj, '-', 'blue', "Predicted")
+    plot_collection.add_figure("traj (error)", fig)
+    plot_collection.export(filename, confirm_overwrite=False)
+    plt.close(fig=fig)
+    print(f"Saved {filename}")
+
+def save_trajectory_tum_format(traj, filename):
+    traj = make_traj(traj)
+    tostr = lambda a: ' '.join(map(str, a))
+    with Path(filename).open('w') as f:
+        for i in range(traj.num_poses):
+            f.write(f"{traj.timestamps[i]} {tostr(traj.positions_xyz[i])} {tostr(traj.orientations_quat_wxyz[i][[1,2,3,0]])}\n")
+    print(f"Saved {filename}")

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/projective_ops.py

@@ -0,0 +1,121 @@
+import torch
+import torch.nn.functional as F
+
+from .lietorch import SE3, Sim3
+
+MIN_DEPTH = 0.2
+
+def extract_intrinsics(intrinsics):
+    return intrinsics[...,None,None,:].unbind(dim=-1)
+
+def coords_grid(ht, wd, **kwargs):
+    y, x = torch.meshgrid(
+        torch.arange(ht).to(**kwargs).float(),
+        torch.arange(wd).to(**kwargs).float())
+
+    return torch.stack([x, y], dim=-1)
+
+
+def iproj(patches, intrinsics):
+    """ inverse projection """
+    x, y, d = patches.unbind(dim=2)
+    fx, fy, cx, cy = intrinsics[...,None,None].unbind(dim=2)
+
+    i = torch.ones_like(d)
+    xn = (x - cx) / fx
+    yn = (y - cy) / fy
+
+    X = torch.stack([xn, yn, i, d], dim=-1)
+    return X
+
+
+def proj(X, intrinsics, depth=False):
+    """ projection """
+
+    X, Y, Z, W = X.unbind(dim=-1)
+    fx, fy, cx, cy = intrinsics[...,None,None].unbind(dim=2)
+
+    # d = 0.01 * torch.ones_like(Z)
+    # d[Z > 0.01] = 1.0 / Z[Z > 0.01]
+    # d = torch.ones_like(Z)
+    # d[Z.abs() > 0.1] = 1.0 / Z[Z.abs() > 0.1]
+
+    d = 1.0 / Z.clamp(min=0.1)
+    x = fx * (d * X) + cx
+    y = fy * (d * Y) + cy
+
+    if depth:
+        return torch.stack([x, y, d], dim=-1)
+
+    return torch.stack([x, y], dim=-1)
+
+
+def transform(poses, patches, intrinsics, ii, jj, kk, depth=False, valid=False, jacobian=False, tonly=False):
+    """ projective transform """
+
+    # backproject
+    X0 = iproj(patches[:,kk], intrinsics[:,ii])
+
+    # transform
+    Gij = poses[:, jj] * poses[:, ii].inv()
+
+    if tonly:
+        Gij[...,3:] = torch.as_tensor([0,0,0,1], device=Gij.device)
+
+    X1 = Gij[:,:,None,None] * X0
+
+    # project
+    x1 = proj(X1, intrinsics[:,jj], depth)
+
+
+    if jacobian:
+        p = X1.shape[2]
+        X, Y, Z, H = X1[...,p//2,p//2,:].unbind(dim=-1)
+        o = torch.zeros_like(H)
+        i = torch.zeros_like(H)
+
+        fx, fy, cx, cy = intrinsics[:,jj].unbind(dim=-1)
+
+        d = torch.zeros_like(Z)
+        d[Z.abs() > 0.2] = 1.0 / Z[Z.abs() > 0.2]
+
+        Ja = torch.stack([
+            H,  o,  o,  o,  Z, -Y,
+            o,  H,  o, -Z,  o,  X, 
+            o,  o,  H,  Y, -X,  o,
+            o,  o,  o,  o,  o,  o,
+        ], dim=-1).view(1, len(ii), 4, 6)
+        
+        Jp = torch.stack([
+             fx*d,     o, -fx*X*d*d,  o,
+                o,  fy*d, -fy*Y*d*d,  o,
+        ], dim=-1).view(1, len(ii), 2, 4)
+
+        Jj = torch.matmul(Jp, Ja)
+        Ji = -Gij[:,:,None].adjT(Jj)
+        
+        Jz = torch.matmul(Jp, Gij.matrix()[...,:,3:])
+
+        return x1, (Z > 0.2).float(), (Ji, Jj, Jz)
+
+    if valid:
+        return x1, (X1[...,2] > 0.2).float()
+        
+    return x1
+
+def point_cloud(poses, patches, intrinsics, ix):
+    """ generate point cloud from patches """
+    return poses[:,ix,None,None].inv() * iproj(patches, intrinsics[:,ix])
+
+
+def flow_mag(poses, patches, intrinsics, ii, jj, kk, beta=0.3):
+    """ projective transform """
+
+    coords0 = transform(poses, patches, intrinsics, ii, ii, kk)
+    coords1 = transform(poses, patches, intrinsics, ii, jj, kk, tonly=False)
+    coords2 = transform(poses, patches, intrinsics, ii, jj, kk, tonly=True)
+
+    flow1 = (coords1 - coords0).norm(dim=-1)
+    flow2 = (coords2 - coords0).norm(dim=-1)
+
+    return beta * flow1 + (1-beta) * flow2

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/stream.py

@@ -0,0 +1,87 @@
+import os
+import cv2
+import numpy as np
+from multiprocessing import Process, Queue
+from pathlib import Path
+from itertools import chain
+
+def image_stream(queue, imagedir, calib, stride, skip=0):
+    """ image generator """
+
+    calib = np.loadtxt(calib, delimiter=" ")
+    fx, fy, cx, cy = calib[:4]
+
+    K = np.eye(3)
+    K[0,0] = fx
+    K[0,2] = cx
+    K[1,1] = fy
+    K[1,2] = cy
+
+    img_exts = ["*.png", "*.jpeg", "*.jpg"]
+    image_list = sorted(chain.from_iterable(Path(imagedir).glob(e) for e in img_exts))[skip::stride]
+
+    for t, imfile in enumerate(image_list):
+        image = cv2.imread(str(imfile))
+        if len(calib) > 4:
+            image = cv2.undistort(image, K, calib[4:])
+
+        if 0:
+            image = cv2.resize(image, None, fx=0.5, fy=0.5)
+            intrinsics = np.array([fx / 2, fy / 2, cx / 2, cy / 2])
+
+        else:
+            intrinsics = np.array([fx, fy, cx, cy])
+            
+        h, w, _ = image.shape
+        image = image[:h-h%16, :w-w%16]
+
+        queue.put((t, image, intrinsics))
+
+    queue.put((-1, image, intrinsics))
+
+
+def video_stream(queue, imagedir, calib, stride, skip=0):
+    """ video generator """
+
+    calib = np.loadtxt(calib, delimiter=" ")
+    fx, fy, cx, cy = calib[:4]
+
+    K = np.eye(3)
+    K[0,0] = fx
+    K[0,2] = cx
+    K[1,1] = fy
+    K[1,2] = cy
+
+    cap = cv2.VideoCapture(imagedir)
+
+    t = 0
+
+    for _ in range(skip):
+        ret, image = cap.read()
+
+    while True:
+        # Capture frame-by-frame
+        for _ in range(stride):
+            ret, image = cap.read()
+            # if frame is read correctly ret is True
+            if not ret:
+                break
+
+        if not ret:
+            break
+
+        if len(calib) > 4:
+            image = cv2.undistort(image, K, calib[4:])
+
+        image = cv2.resize(image, None, fx=0.5, fy=0.5, interpolation=cv2.INTER_AREA)
+        h, w, _ = image.shape
+        image = image[:h-h%16, :w-w%16]
+
+        intrinsics = np.array([fx*.5, fy*.5, cx*.5, cy*.5])
+        queue.put((t, image, intrinsics))
+
+        t += 1
+
+    queue.put((-1, image, intrinsics))
+    cap.release()
+

third-party/DPVO/build/lib.win-amd64-3.9/dpvo/utils.py

@@ -0,0 +1,87 @@
+import torch
+import torch.nn.functional as F
+
+
+all_times = []
+
+class Timer:
+    def __init__(self, name, enabled=True):
+        self.name = name
+        self.enabled = enabled
+
+        if self.enabled:
+            self.start = torch.cuda.Event(enable_timing=True)
+            self.end = torch.cuda.Event(enable_timing=True)
+
+    def __enter__(self):
+        if self.enabled:
+            self.start.record()
+        
+    def __exit__(self, type, value, traceback):
+        global all_times
+        if self.enabled:
+            self.end.record()
+            torch.cuda.synchronize()
+
+            elapsed = self.start.elapsed_time(self.end)
+            all_times.append(elapsed)
+            print(self.name, elapsed)
+
+
+def coords_grid(b, n, h, w, **kwargs):
+    """ coordinate grid """
+    x = torch.arange(0, w, dtype=torch.float, **kwargs)
+    y = torch.arange(0, h, dtype=torch.float, **kwargs)
+    coords = torch.stack(torch.meshgrid(y, x, indexing="ij"))
+    return coords[[1,0]].view(1, 1, 2, h, w).repeat(b, n, 1, 1, 1)
+
+def coords_grid_with_index(d, **kwargs):
+    """ coordinate grid with frame index"""
+    b, n, h, w = d.shape
+    i = torch.ones_like(d)
+    x = torch.arange(0, w, dtype=torch.float, **kwargs)
+    y = torch.arange(0, h, dtype=torch.float, **kwargs)
+
+    y, x = torch.stack(torch.meshgrid(y, x, indexing="ij"))
+    y = y.view(1, 1, h, w).repeat(b, n, 1, 1)
+    x = x.view(1, 1, h, w).repeat(b, n, 1, 1)
+
+    coords = torch.stack([x, y, d], dim=2)
+    index = torch.arange(0, n, dtype=torch.float, **kwargs)
+    index = index.view(1, n, 1, 1, 1).repeat(b, 1, 1, h, w)
+
+    return coords, index
+
+def patchify(x, patch_size=3):
+    """ extract patches from video """
+    b, n, c, h, w = x.shape
+    x = x.view(b*n, c, h, w)
+    y = F.unfold(x, patch_size)
+    y = y.transpose(1,2)
+    return y.reshape(b, -1, c, patch_size, patch_size)
+
+
+def pyramidify(fmap, lvls=[1]):
+    """ turn fmap into a pyramid """
+    b, n, c, h, w = fmap.shape
+
+    pyramid = []
+    for lvl in lvls:
+        gmap =  F.avg_pool2d(fmap.view(b*n, c, h, w), lvl, stride=lvl)
+        pyramid += [ gmap.view(b, n, c, h//lvl, w//lvl) ]
+        
+    return pyramid
+
+def all_pairs_exclusive(n, **kwargs):
+    ii, jj = torch.meshgrid(torch.arange(n, **kwargs), torch.arange(n, **kwargs))
+    k = ii != jj
+    return ii[k].reshape(-1), jj[k].reshape(-1)
+
+def set_depth(patches, depth):
+    patches[...,2,:,:] = depth[...,None,None]
+    return patches
+
+def flatmeshgrid(*args, **kwargs):
+    grid = torch.meshgrid(*args, **kwargs)
+    return (x.reshape(-1) for x in grid)
+

third-party/DPVO/build/lib.win-amd64-3.9/lietorch_backends.cp39-win_amd64.pyd

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2c6a2dd2d29f5ee56eeb0912681982366f36a0bc7420a1ea05b05d64d761a2f
+size 2983936

third-party/DPVO/build/temp.win-amd64-3.9/Release/.ninja_log

@@ -0,0 +1,6 @@
+# ninja log v5
+0	10746	7510839870546033	C:/Users/thpap/PycharmProjects/motionbert-meta-sapiens/WHAM/third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/altcorr/correlation.obj	8a6ea87c38b5029e
+5	34954	7510840112325833	C:/Users/thpap/PycharmProjects/motionbert-meta-sapiens/WHAM/third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/altcorr/correlation_kernel.obj	f7d1343ce924f3b5
+7	10971	7510840239521303	C:/Users/thpap/PycharmProjects/motionbert-meta-sapiens/WHAM/third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/fastba/ba.obj	1448633d654dbba0
+1	33893	7510840468461448	C:/Users/thpap/PycharmProjects/motionbert-meta-sapiens/WHAM/third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/fastba/ba_cuda.obj	d1400b5b4a6daa70
+0	10676	7510840592828284	C:/Users/thpap/PycharmProjects/motionbert-meta-sapiens/WHAM/third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/lietorch/src/lietorch.obj	ba69a789264375ff

third-party/DPVO/build/temp.win-amd64-3.9/Release/build.ninja

@@ -0,0 +1,29 @@
+ninja_required_version = 1.3
+cxx = cl
+nvcc = C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.3\bin\nvcc
+
+cflags = /nologo /Ox /W3 /GL /DNDEBUG /MD /MD /wd4819 /wd4251 /wd4244 /wd4267 /wd4275 /wd4018 /wd4190 /EHsc -IC:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\dpvo/lietorch/include -IC:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\thirdparty/eigen-3.4.0 -IC:\Users\thpap\miniconda3\envs\wham\lib\site-packages\torch\include -IC:\Users\thpap\miniconda3\envs\wham\lib\site-packages\torch\include\torch\csrc\api\include -IC:\Users\thpap\miniconda3\envs\wham\lib\site-packages\torch\include\TH -IC:\Users\thpap\miniconda3\envs\wham\lib\site-packages\torch\include\THC "-IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.3\include" -IC:\Users\thpap\miniconda3\envs\wham\include -IC:\Users\thpap\miniconda3\envs\wham\include "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\include" "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\include" "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\winrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\cppwinrt" "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\include" "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\include" "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\winrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\cppwinrt"
+post_cflags = -O3 -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=lietorch_backends -D_GLIBCXX_USE_CXX11_ABI=0 /std:c++14
+cuda_cflags = --use-local-env -Xcompiler /MD -Xcompiler /wd4819 -Xcompiler /wd4251 -Xcompiler /wd4244 -Xcompiler /wd4267 -Xcompiler /wd4275 -Xcompiler /wd4018 -Xcompiler /wd4190 -Xcompiler /EHsc -Xcudafe --diag_suppress=base_class_has_different_dll_interface -Xcudafe --diag_suppress=field_without_dll_interface -Xcudafe --diag_suppress=dll_interface_conflict_none_assumed -Xcudafe --diag_suppress=dll_interface_conflict_dllexport_assumed -IC:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\dpvo/lietorch/include -IC:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\thirdparty/eigen-3.4.0 -IC:\Users\thpap\miniconda3\envs\wham\lib\site-packages\torch\include -IC:\Users\thpap\miniconda3\envs\wham\lib\site-packages\torch\include\torch\csrc\api\include -IC:\Users\thpap\miniconda3\envs\wham\lib\site-packages\torch\include\TH -IC:\Users\thpap\miniconda3\envs\wham\lib\site-packages\torch\include\THC "-IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.3\include" -IC:\Users\thpap\miniconda3\envs\wham\include -IC:\Users\thpap\miniconda3\envs\wham\include "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\include" "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\include" "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\winrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\cppwinrt" "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\ATLMFC\include" "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\include" "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\shared" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\winrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.22621.0\cppwinrt"
+cuda_post_cflags = -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -O3 --allow-unsupported-compiler -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=lietorch_backends -D_GLIBCXX_USE_CXX11_ABI=0 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_86,code=sm_86
+ldflags = 
+
+rule compile
+  command = cl /showIncludes $cflags -c $in /Fo$out $post_cflags
+  deps = msvc
+
+rule cuda_compile
+  depfile = $out.d
+  deps = gcc
+  command = $nvcc --generate-dependencies-with-compile --dependency-output $out.d $cuda_cflags -c $in -o $out $cuda_post_cflags
+
+
+
+build C$:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\build\temp.win-amd64-3.9\Release\dpvo/lietorch/src/lietorch.obj: compile C$:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\dpvo\lietorch\src\lietorch.cpp
+build C$:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\build\temp.win-amd64-3.9\Release\dpvo/lietorch/src/lietorch_cpu.obj: compile C$:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\dpvo\lietorch\src\lietorch_cpu.cpp
+build C$:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\build\temp.win-amd64-3.9\Release\dpvo/lietorch/src/lietorch_gpu.obj: cuda_compile C$:\Users\thpap\PycharmProjects\motionbert-meta-sapiens\WHAM\third-party\DPVO\dpvo\lietorch\src\lietorch_gpu.cu
+
+
+
+
+

third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/altcorr/correlation.obj

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f1ef51859e20505bb755ea5e5ce48fba33b480d99b464405edb944ee5c40191
+size 37419806

third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/altcorr/correlation_kernel.obj

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c41ced021c0c59fdfba4692ce3769cbc85e0a7b72ad5a1c332c86ff5ca5015b3
+size 1461202

third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/fastba/ba.obj

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c345d67f220f9216fd2ce229b4d4a25c91ba82568459eb137ed18ae96d77aad
+size 37657608

third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/fastba/ba_cuda.obj

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:766b7b9f519617f73e61fefca21afde2a56852bfc92bec64cb88fcf60307657e
+size 1195236

third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/lietorch/src/lietorch.obj

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc39b10aa93073560e20f0b855a64775989f4368fc8dc067d086a10546a18c63
+size 44606503

third-party/DPVO/build/temp.win-amd64-3.9/Release/dpvo/lietorch/src/lietorch_cpu.obj

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1957cc4bf3d2edd2496ff09abebd3da682200fc79241e85be4ba902323d90d47
+size 124596555