File size: 10,123 Bytes
3334467 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 |
import torch
import torch.nn.functional as F
from scipy.optimize import linear_sum_assignment
from torch import nn
from torch.cuda.amp import autocast
def batch_dice_loss(inputs: torch.Tensor, targets: torch.Tensor):
"""
Compute the DICE loss, similar to generalized IOU for masks
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
"""
inputs = inputs.sigmoid()
inputs = inputs.flatten(1)
numerator = 2 * torch.einsum("nc,mc->nm", inputs, targets)
denominator = inputs.sum(-1)[:, None] + targets.sum(-1)[None, :]
loss = 1 - (numerator + 1) / (denominator + 1)
return loss
batch_dice_loss_jit = torch.jit.script(
batch_dice_loss
)
def batch_sigmoid_ce_loss(inputs: torch.Tensor, targets: torch.Tensor):
"""
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
Returns:
Loss tensor
"""
hw = inputs.shape[1]
pos = F.binary_cross_entropy_with_logits(
inputs, torch.ones_like(inputs), reduction="none"
)
neg = F.binary_cross_entropy_with_logits(
inputs, torch.zeros_like(inputs), reduction="none"
)
loss = torch.einsum("nc,mc->nm", pos, targets) + torch.einsum(
"nc,mc->nm", neg, (1 - targets)
)
return loss / hw
def point_sample(input, point_coords, **kwargs):
"""
A wrapper around :function:`torch.nn.functional.grid_sample` to support 3D point_coords tensors.
Unlike :function:`torch.nn.functional.grid_sample` it assumes `point_coords` to lie inside
[0, 1] x [0, 1] square.
Args:
input (Tensor): A tensor of shape (N, C, H, W) that contains features map on a H x W grid.
point_coords (Tensor): A tensor of shape (N, P, 2) or (N, Hgrid, Wgrid, 2) that contains
[0, 1] x [0, 1] normalized point coordinates.
Returns:
output (Tensor): A tensor of shape (N, C, P) or (N, C, Hgrid, Wgrid) that contains
features for points in `point_coords`. The features are obtained via bilinear
interplation from `input` the same way as :function:`torch.nn.functional.grid_sample`.
"""
add_dim = False
if point_coords.dim() == 3:
add_dim = True
point_coords = point_coords.unsqueeze(2)
output = F.grid_sample(input.float(), 2.0 * point_coords.float() - 1.0, **kwargs)
if add_dim:
output = output.squeeze(3)
return output
def match_pred(out_mask, tgt_mask):
cost_mask = 5
cost_dice = 5
num_points = 12544
out_mask = out_mask[:, None]
tgt_mask = tgt_mask[:, None]
num_pred = out_mask.shape[0]
num_tgt = tgt_mask.shape[0]
point_coords = torch.rand(1, num_points, 2, device=out_mask.device)
tgt_mask = point_sample(
tgt_mask,
point_coords.repeat(tgt_mask.shape[0], 1, 1),
align_corners=False,
).squeeze(1)
out_mask = point_sample(
out_mask,
point_coords.repeat(out_mask.shape[0], 1, 1),
align_corners=False,
).squeeze(1)
with autocast(enabled=False):
out_mask = out_mask.float()
tgt_mask = tgt_mask.float()
cost_mask = batch_sigmoid_ce_loss(out_mask, tgt_mask)
cost_dice = batch_dice_loss(out_mask, tgt_mask)
C = (
cost_mask
+ cost_dice
)
C = C.reshape(num_pred, num_tgt).cpu()
indices = linear_sum_assignment(C)
return indices
class HungarianMatcher(nn.Module):
"""This class computes an assignment between the targets and the predictions of the network
For efficiency reasons, the targets don't include the no_object. Because of this, in general,
there are more predictions than targets. In this case, we do a 1-to-1 matching of the best predictions,
while the others are un-matched (and thus treated as non-objects).
"""
def __init__(self, cost_class: float = 2, cost_mask: float = 5, cost_dice: float = 5, num_points: int = 0):
"""Creates the matcher
Params:
cost_class: This is the relative weight of the classification error in the matching cost
cost_mask: This is the relative weight of the focal loss of the binary mask in the matching cost
cost_dice: This is the relative weight of the dice loss of the binary mask in the matching cost
"""
super().__init__()
self.cost_class = cost_class
self.cost_mask = cost_mask
self.cost_dice = cost_dice
assert cost_class != 0 or cost_mask != 0 or cost_dice != 0, "all costs cant be 0"
self.num_points = num_points
@torch.no_grad()
def memory_efficient_forward(self, outputs, targets):
"""More memory-friendly matching"""
bs, num_queries = outputs["pred_logits"].shape[:2]
indices = []
for b in range(bs):
out_prob = outputs["pred_logits"][b].softmax(-1)
tgt_ids = targets[b]["labels"]
try:
cost_class = -out_prob[:, tgt_ids.tolist()]
except:
print("out_prob_shape:", out_prob.shape, tgt_ids.tolist())
out_mask = outputs["pred_masks"][b]
tgt_mask = targets[b]["masks"].to(out_mask)
out_mask = out_mask[:, None]
tgt_mask = tgt_mask[:, None]
point_coords = torch.rand(1, self.num_points, 2, device=out_mask.device)
tgt_mask = point_sample(
tgt_mask,
point_coords.repeat(tgt_mask.shape[0], 1, 1),
align_corners=False,
).squeeze(1)
out_mask = point_sample(
out_mask,
point_coords.repeat(out_mask.shape[0], 1, 1),
align_corners=False,
).squeeze(1)
with autocast(enabled=False):
out_mask = out_mask.float()
tgt_mask = tgt_mask.float()
cost_mask = batch_sigmoid_ce_loss(out_mask, tgt_mask)
cost_dice = batch_dice_loss(out_mask, tgt_mask)
C = (
self.cost_mask * cost_mask
+ self.cost_class * cost_class
+ self.cost_dice * cost_dice
)
C = C.reshape(num_queries, -1).cpu()
try:
indices.append(linear_sum_assignment(C))
except:
print("out_prob is nan:", torch.isnan(out_prob).any())
return [
(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64))
for i, j in indices
]
@torch.no_grad()
def forward(self, outputs, targets):
"""Performs the matching
Params:
outputs: This is a dict that contains at least these entries:
"pred_logits": Tensor of dim [batch_size, num_queries, num_classes] with the classification logits
"pred_masks": Tensor of dim [batch_size, num_queries, H_pred, W_pred] with the predicted masks
targets: This is a list of targets (len(targets) = batch_size), where each target is a dict containing:
"labels": Tensor of dim [num_target_boxes] (where num_target_boxes is the number of ground-truth
objects in the target) containing the class labels
"masks": Tensor of dim [num_target_boxes, H_gt, W_gt] containing the target masks
Returns:
A list of size batch_size, containing tuples of (index_i, index_j) where:
- index_i is the indices of the selected predictions (in order)
- index_j is the indices of the corresponding selected targets (in order)
For each batch element, it holds:
len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
"""
return self.memory_efficient_forward(outputs, targets)
def __repr__(self, _repr_indent=4):
head = "Matcher " + self.__class__.__name__
body = [
"cost_class: {}".format(self.cost_class),
"cost_mask: {}".format(self.cost_mask),
"cost_dice: {}".format(self.cost_dice),
]
lines = [head] + [" " * _repr_indent + line for line in body]
return "\n".join(lines)
|