import torch import torch.nn as nn from einops.einops import rearrange from .backbone import ResNet_8_2 from .utils.position_encoding import PositionEncodingSine from .xoftr_module import LocalFeatureTransformer, FineProcess class XoFTR_Pretrain(nn.Module): def __init__(self, config): super().__init__() # Misc self.config = config self.patch_size = config["pretrain_patch_size"] # Modules self.backbone = ResNet_8_2(config['resnet']) self.pos_encoding = PositionEncodingSine(config['coarse']['d_model']) self.loftr_coarse = LocalFeatureTransformer(config['coarse']) self.fine_process = FineProcess(config) self.mask_token_f = nn.Parameter(torch.zeros(1, config['resnet']["block_dims"][0], 1, 1)) self.mask_token_m = nn.Parameter(torch.zeros(1, config['resnet']["block_dims"][1], 1, 1)) self.mask_token_c = nn.Parameter(torch.zeros(1, config['resnet']["block_dims"][2], 1, 1)) self.out_proj = nn.Linear(config['resnet']["block_dims"][0], 4) torch.nn.init.normal_(self.mask_token_f, std=.02) torch.nn.init.normal_(self.mask_token_m, std=.02) torch.nn.init.normal_(self.mask_token_c, std=.02) def upsample_mae_mask(self, mae_mask, scale): assert len(mae_mask.shape) == 2 p = int(mae_mask.shape[1] ** .5) return mae_mask.reshape(-1, p, p).repeat_interleave(scale, axis=1).repeat_interleave(scale, axis=2) def upsample_mask(self, mask, scale): return mask.repeat_interleave(scale, axis=1).repeat_interleave(scale, axis=2) def mask_layer(self, feat, mae_mask, mae_mask_scale, mask=None, mask_scale=None, mask_token=None): """ Mask the feature map and replace with trainable inpu tokens if available Args: feat (torch.Tensor): [N, C, H, W] mae_mask (torch.Tensor): (N, L) mask for masked image modeling mae_mask_scale (int): the scale of layer to mae mask mask (torch.Generator): mask for padded input image mask_scale (int): the scale of layer to mask (mask is created on course scale) mask_token (torch.Tensor): [1, C, 1, 1] learnable mae mask token Returns: feat (torch.Tensor): [N, C, H, W] """ mae_mask = self.upsample_mae_mask(mae_mask, mae_mask_scale) mae_mask = mae_mask.unsqueeze(1).type_as(feat) if mask is not None: mask = self.upsample_mask(mask, mask_scale) mask = mask.unsqueeze(1).type_as(feat) mae_mask = mask * mae_mask feat = feat * (1. - mae_mask) if mask_token is not None: mask_token = mask_token.repeat(feat.shape[0], 1, feat.shape[2], feat.shape[3]) feat += mask_token * mae_mask return feat def forward(self, data): """ Update: data (dict): { 'image0': (torch.Tensor): (N, 1, H, W) 'image1': (torch.Tensor): (N, 1, H, W) 'mask0'(optional) : (torch.Tensor): (N, H, W) '0' indicates a padded position 'mask1'(optional) : (torch.Tensor): (N, H, W) } """ # 1. Local Feature CNN data.update({ 'bs': data['image0'].size(0), 'hw0_i': data['image0'].shape[2:], 'hw1_i': data['image1'].shape[2:] }) image0 = data["image0_norm"] if "image0_norm" in data else data["image0"] image1 = data["image1_norm"] if "image1_norm" in data else data["image1"] mask0 = mask1 = None # mask fro madded images if 'mask0' in data: mask0, mask1 = data['mask0'], data['mask1'] # mask input images image0 = self.mask_layer(image0, data["mae_mask0"], mae_mask_scale=self.patch_size, mask=mask0, mask_scale=8) image1 = self.mask_layer(image1, data["mae_mask1"], mae_mask_scale=self.patch_size, mask=mask1, mask_scale=8) data.update({"masked_image0":image0.clone().detach().cpu(), "masked_image1":image1.clone().detach().cpu()}) if data['hw0_i'] == data['hw1_i']: # faster & better BN convergence feats_c, feats_m, feats_f = self.backbone(torch.cat([image0, image1], dim=0)) (feat_c0, feat_c1) = feats_c.split(data['bs']) (feat_m0, feat_m1) = feats_m.split(data['bs']) (feat_f0, feat_f1) = feats_f.split(data['bs']) else: # handle different input shapes feat_c0, feat_m0, feat_f0 = self.backbone(image0) feat_c1, feat_m1, feat_f1 = self.backbone(image1) # mask output layers of backbone and replace with trainable token feat_c0 = self.mask_layer(feat_c0, data["mae_mask0"], mae_mask_scale=self.patch_size // 8, mask=mask0, mask_scale=1, mask_token=self.mask_token_c) feat_c1 = self.mask_layer(feat_c1, data["mae_mask1"], mae_mask_scale=self.patch_size // 8, mask=mask1, mask_scale=1, mask_token=self.mask_token_c) feat_m0 = self.mask_layer(feat_m0, data["mae_mask0"], mae_mask_scale=self.patch_size // 4, mask=mask0, mask_scale=2, mask_token=self.mask_token_m) feat_m1 = self.mask_layer(feat_m1, data["mae_mask1"], mae_mask_scale=self.patch_size // 4, mask=mask1, mask_scale=2, mask_token=self.mask_token_m) feat_f0 = self.mask_layer(feat_f0, data["mae_mask0"], mae_mask_scale=self.patch_size // 2, mask=mask0, mask_scale=4, mask_token=self.mask_token_f) feat_f1 = self.mask_layer(feat_f1, data["mae_mask1"], mae_mask_scale=self.patch_size // 2, mask=mask1, mask_scale=4, mask_token=self.mask_token_f) data.update({ 'hw0_c': feat_c0.shape[2:], 'hw1_c': feat_c1.shape[2:], 'hw0_m': feat_m0.shape[2:], 'hw1_m': feat_m1.shape[2:], 'hw0_f': feat_f0.shape[2:], 'hw1_f': feat_f1.shape[2:] }) # save coarse features for fine matching module feat_c0_pre, feat_c1_pre = feat_c0.clone(), feat_c1.clone() # 2. Coarse-level loftr module # add featmap with positional encoding, then flatten it to sequence [N, HW, C] feat_c0 = rearrange(self.pos_encoding(feat_c0), 'n c h w -> n (h w) c') feat_c1 = rearrange(self.pos_encoding(feat_c1), 'n c h w -> n (h w) c') mask_c0 = mask_c1 = None # mask is useful in training if 'mask0' in data: mask_c0, mask_c1 = data['mask0'].flatten(-2), data['mask1'].flatten(-2) feat_c0, feat_c1 = self.loftr_coarse(feat_c0, feat_c1, mask_c0, mask_c1) # 3. Fine-level maching module as decoder # generate window locations from mae mask to reconstruct mae_mask_c0 = self.upsample_mae_mask( data["mae_mask0"], self.patch_size // 8) if mask0 is not None: mae_mask_c0 = mae_mask_c0 * mask0.type_as(mae_mask_c0) mae_mask_c1 = self.upsample_mae_mask( data["mae_mask1"], self.patch_size // 8) if mask1 is not None: mae_mask_c1 = mae_mask_c1 * mask1.type_as(mae_mask_c1) mae_mask_c = torch.logical_or(mae_mask_c0, mae_mask_c1) b_ids, i_ids = mae_mask_c.flatten(1).nonzero(as_tuple=True) j_ids = i_ids # b_ids, i_ids and j_ids are masked location for both images # ids_image0 and ids_image1 determines which indeces belogs to which image ids_image0 = mae_mask_c0.flatten(1)[b_ids, i_ids] ids_image1 = mae_mask_c1.flatten(1)[b_ids, j_ids] data.update({'b_ids': b_ids, 'i_ids': i_ids, 'j_ids': j_ids, 'ids_image0': ids_image0==1, 'ids_image1': ids_image1==1}) # fine level matching module feat_f0_unfold, feat_f1_unfold = self.fine_process( feat_f0, feat_f1, feat_m0, feat_m1, feat_c0, feat_c1, feat_c0_pre, feat_c1_pre, data) # output projection 5x5 window to 10x10 window pred0 = self.out_proj(feat_f0_unfold) pred1 = self.out_proj(feat_f1_unfold) data.update({"pred0":pred0, "pred1": pred1}) def load_state_dict(self, state_dict, *args, **kwargs): for k in list(state_dict.keys()): if k.startswith('matcher.'): state_dict[k.replace('matcher.', '', 1)] = state_dict.pop(k) return super().load_state_dict(state_dict, *args, **kwargs)