Spaces:
Runtime error
Runtime error
File size: 7,453 Bytes
04fbff5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class Loss(nn.Module):
def __init__(self, loss_weight, keys, mapping=None) -> None:
'''
mapping: map the kwargs keys into desired ones.
'''
super().__init__()
self.loss_weight = loss_weight
self.keys = keys
self.mapping = mapping
if isinstance(mapping, dict):
self.mapping = {k: v for k, v in mapping if v in keys}
def forward(self, **kwargs):
params = {k: v for k, v in kwargs.items() if k in self.keys}
if self.mapping is not None:
for k, v in kwargs.items():
if self.mapping.get(k) is not None:
params[self.mapping[k]] = v
return self._forward(**params) * self.loss_weight
def _forward(self, **kwargs):
pass
class CharbonnierLoss(Loss):
def __init__(self, loss_weight, keys) -> None:
super().__init__(loss_weight, keys)
def _forward(self, imgt_pred, imgt):
diff = imgt_pred - imgt
loss = ((diff ** 2 + 1e-6) ** 0.5).mean()
return loss
class AdaCharbonnierLoss(Loss):
def __init__(self, loss_weight, keys) -> None:
super().__init__(loss_weight, keys)
def _forward(self, imgt_pred, imgt, weight):
alpha = weight / 2
epsilon = 10 ** (-(10 * weight - 1) / 3)
diff = imgt_pred - imgt
loss = ((diff ** 2 + epsilon ** 2) ** alpha).mean()
return loss
class TernaryLoss(Loss):
def __init__(self, loss_weight, keys, patch_size=7):
super().__init__(loss_weight, keys)
self.patch_size = patch_size
out_channels = patch_size * patch_size
self.w = np.eye(out_channels).reshape((patch_size, patch_size, 1, out_channels))
self.w = np.transpose(self.w, (3, 2, 0, 1))
self.w = torch.tensor(self.w, dtype=torch.float32)
def transform(self, tensor):
self.w = self.w.to(tensor.device)
tensor_ = tensor.mean(dim=1, keepdim=True)
patches = F.conv2d(tensor_, self.w, padding=self.patch_size//2, bias=None)
loc_diff = patches - tensor_
loc_diff_norm = loc_diff / torch.sqrt(0.81 + loc_diff ** 2)
return loc_diff_norm
def valid_mask(self, tensor):
padding = self.patch_size//2
b, c, h, w = tensor.size()
inner = torch.ones(b, 1, h - 2 * padding, w - 2 * padding).type_as(tensor)
mask = F.pad(inner, [padding] * 4)
return mask
def _forward(self, imgt_pred, imgt):
loc_diff_x = self.transform(imgt_pred)
loc_diff_y = self.transform(imgt)
diff = loc_diff_x - loc_diff_y.detach()
dist = (diff ** 2 / (0.1 + diff ** 2)).mean(dim=1, keepdim=True)
mask = self.valid_mask(imgt_pred)
loss = (dist * mask).mean()
return loss
class GeometryLoss(Loss):
def __init__(self, loss_weight, keys, patch_size=3):
super().__init__(loss_weight, keys)
self.patch_size = patch_size
out_channels = patch_size * patch_size
self.w = np.eye(out_channels).reshape((patch_size, patch_size, 1, out_channels))
self.w = np.transpose(self.w, (3, 2, 0, 1))
self.w = torch.tensor(self.w).float()
def transform(self, tensor):
b, c, h, w = tensor.size()
self.w = self.w.to(tensor.device)
tensor_ = tensor.reshape(b*c, 1, h, w)
patches = F.conv2d(tensor_, self.w, padding=self.patch_size // 2, bias=None)
loc_diff = patches - tensor_
loc_diff_ = loc_diff.reshape(b, c*(self.patch_size ** 2), h, w)
loc_diff_norm = loc_diff_ / torch.sqrt(0.81 + loc_diff_ ** 2)
return loc_diff_norm
def valid_mask(self, tensor):
padding = self.patch_size // 2
b, c, h, w = tensor.size()
inner = torch.ones(b, 1, h - 2 * padding, w - 2 * padding).type_as(tensor)
mask = F.pad(inner, [padding] * 4)
return mask
def _forward(self, ft_pred, ft_gt):
loss = 0.
for pred, gt in zip(ft_pred, ft_gt):
loc_diff_x = self.transform(pred)
loc_diff_y = self.transform(gt)
diff = loc_diff_x - loc_diff_y
dist = (diff ** 2 / (0.1 + diff ** 2)).mean(dim=1, keepdim=True)
mask = self.valid_mask(pred)
loss = loss + (dist * mask).mean()
return loss
class IFRFlowLoss(Loss):
def __init__(self, loss_weight, keys, beta=0.3) -> None:
super().__init__(loss_weight, keys)
self.beta = beta
self.ada_cb_loss = AdaCharbonnierLoss(1.0, ['imgt_pred', 'imgt', 'weight'])
def _forward(self, flow0_pred, flow1_pred, flow):
robust_weight0 = self.get_robust_weight(flow0_pred[0], flow[:, 0:2])
robust_weight1 = self.get_robust_weight(flow1_pred[0], flow[:, 2:4])
loss = 0
for lvl in range(1, len(flow0_pred)):
scale_factor = 2**lvl
loss = loss + self.ada_cb_loss(**{
'imgt_pred': self.resize(flow0_pred[lvl], scale_factor),
'imgt': flow[:, 0:2],
'weight': robust_weight0
})
loss = loss + self.ada_cb_loss(**{
'imgt_pred': self.resize(flow1_pred[lvl], scale_factor),
'imgt': flow[:, 2:4],
'weight': robust_weight1
})
return loss
def resize(self, x, scale_factor):
return scale_factor * F.interpolate(x, scale_factor=scale_factor, mode="bilinear", align_corners=False)
def get_robust_weight(self, flow_pred, flow_gt):
epe = ((flow_pred.detach() - flow_gt) ** 2).sum(dim=1, keepdim=True) ** 0.5
robust_weight = torch.exp(-self.beta * epe)
return robust_weight
class MultipleFlowLoss(Loss):
def __init__(self, loss_weight, keys, beta=0.3) -> None:
super().__init__(loss_weight, keys)
self.beta = beta
self.ada_cb_loss = AdaCharbonnierLoss(1.0, ['imgt_pred', 'imgt', 'weight'])
def _forward(self, flow0_pred, flow1_pred, flow):
robust_weight0 = self.get_mutli_flow_robust_weight(flow0_pred[0], flow[:, 0:2])
robust_weight1 = self.get_mutli_flow_robust_weight(flow1_pred[0], flow[:, 2:4])
loss = 0
for lvl in range(1, len(flow0_pred)):
scale_factor = 2**lvl
loss = loss + self.ada_cb_loss(**{
'imgt_pred': self.resize(flow0_pred[lvl], scale_factor),
'imgt': flow[:, 0:2],
'weight': robust_weight0
})
loss = loss + self.ada_cb_loss(**{
'imgt_pred': self.resize(flow1_pred[lvl], scale_factor),
'imgt': flow[:, 2:4],
'weight': robust_weight1
})
return loss
def resize(self, x, scale_factor):
return scale_factor * F.interpolate(x, scale_factor=scale_factor, mode="bilinear", align_corners=False)
def get_mutli_flow_robust_weight(self, flow_pred, flow_gt):
b, num_flows, c, h, w = flow_pred.shape
flow_pred = flow_pred.view(b, num_flows, c, h, w)
flow_gt = flow_gt.repeat(1, num_flows, 1, 1).view(b, num_flows, c, h, w)
epe = ((flow_pred.detach() - flow_gt) ** 2).sum(dim=2, keepdim=True).max(1)[0] ** 0.5
robust_weight = torch.exp(-self.beta * epe)
return robust_weight |