def test_focal_loss_equals_ce_loss_multi_class(self) -> None:
"""
Focal loss with predictions for multiple classes matches ce loss.
"""
inputs = logit(
torch.tensor(
[[
[0.95, 0.55, 0.12, 0.05],
[0.09, 0.95, 0.36, 0.11],
[0.06, 0.12, 0.56, 0.07],
[0.09, 0.15, 0.25, 0.45],
]],
dtype=torch.float32,
))
targets = torch.tensor(
[[[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]],
dtype=torch.float32,
)
focal_loss = sigmoid_focal_loss(inputs,
targets,
gamma=0,
alpha=-1,
reduction="mean")
ce_loss = F.binary_cross_entropy_with_logits(inputs,
targets,
reduction="mean")
self.assertEqual(ce_loss, focal_loss)
def score_box_point_loss(target_scores, target_box_deltas, target_point_deltas,
pred_logits, pred_box_deltas, pred_point_deltas,
point_mask):
# target_scores: [B, k]
# target_box_deltas: [B, k, 4]
# target_point_deltas: [B, k, p, 2]
# point_mask: [B, k]
score_loss = sigmoid_focal_loss(pred_logits,
target_scores,
reduction='sum')
maxscore, _ = target_scores.max(-1)
pos_mask = maxscore > 0.5
target_box_deltas_pos = target_box_deltas[pos_mask].reshape(-1, 2, 2)
pred_box_deltas_pos = pred_box_deltas[pos_mask].reshape(-1, 2, 2)
box_loss = euclidean_loss(pred_box_deltas_pos, target_box_deltas_pos,
'sum')
point_mask = pos_mask & point_mask
target_point = target_point_deltas[point_mask] # -1, p, 2
pred_point = pred_point_deltas[point_mask] # -1, p, 2
point_loss = euclidean_loss(pred_point, target_point, 'sum')
npos = target_box_deltas_pos.shape[0]
npoint = target_point.shape[0]
return score_loss / npos, box_loss / npos, point_loss / npoint
def test_focal_loss_equals_ce_loss(self) -> None:
"""
No weighting of easy/hard (gamma = 0) or positive/negative (alpha = 0).
"""
inputs = logit(
torch.tensor([[[0.95], [0.90], [0.98], [0.99]]],
dtype=torch.float32))
targets = torch.tensor([[[1], [1], [1], [1]]], dtype=torch.float32)
inputs_fl = inputs.clone().requires_grad_()
targets_fl = targets.clone()
inputs_ce = inputs.clone().requires_grad_()
targets_ce = targets.clone()
focal_loss = sigmoid_focal_loss(inputs_fl,
targets_fl,
gamma=0,
alpha=-1,
reduction="mean")
ce_loss = F.binary_cross_entropy_with_logits(inputs_ce,
targets_ce,
reduction="mean")
self.assertEqual(ce_loss, focal_loss.data)
focal_loss.backward()
ce_loss.backward()
self.assertTrue(
# pyre-ignore
torch.allclose(inputs_fl.grad.data, inputs_ce.grad.data))
def test_positives_ignored_focal_loss(self) -> None:
"""
With alpha = 0 postive examples have focal loss of 0.
"""
inputs = logit(
torch.tensor(
[[[0.05], [0.12], [0.89], [0.79]]], dtype=torch.float32
)
)
targets = torch.tensor([[[1], [1], [0], [0]]], dtype=torch.float32)
focal_loss = (
sigmoid_focal_loss(inputs, targets, gamma=2, alpha=0)
.squeeze()
.numpy()
)
ce_loss = (
F.binary_cross_entropy_with_logits(
inputs, targets, reduction="none"
)
.squeeze()
.numpy()
)
targets = targets.squeeze().numpy()
self.assertTrue(np.all(ce_loss[targets == 1] > 0))
self.assertTrue(np.all(focal_loss[targets == 1] == 0))
def run_focal_loss() -> None:
fl = sigmoid_focal_loss(inputs,
targets,
gamma=0,
alpha=alpha,
reduction="mean")
fl.backward()
torch.cuda.synchronize()
def forward(self, input: Tensor, target: Tensor):
"""
Shape:
- Input: : $(N, *)$ where $*$ means, any number of additional dimensions.
- Target: : $(N, *)$, same shape as the input.
- Output: scalar. If `reduction` is 'none', then $(N, *)$ , same shape as input.
"""
loss = sigmoid_focal_loss(input, target, self.gamma, self.alpha, self.reduction)
return loss
def test_hard_ex_focal_loss_similar_to_ce_loss(self) -> None:
"""
With gamma = 2 loss of hard examples is unchanged.
"""
inputs = logit(
torch.tensor([0.05, 0.12, 0.09, 0.17], dtype=torch.float32))
targets = torch.tensor([1, 1, 1, 1], dtype=torch.float32)
focal_loss = sigmoid_focal_loss(inputs, targets, gamma=2, alpha=-1)
ce_loss = F.binary_cross_entropy_with_logits(inputs,
targets,
reduction="none")
loss_ratio = (ce_loss / focal_loss).squeeze()
correct_ratio = 1.0 / ((1.0 - inputs.sigmoid())**2)
self.assertTrue(np.allclose(loss_ratio, correct_ratio))
def test_easy_ex_focal_loss_weighted_less_than_ce_loss(self) -> None:
"""
With gamma = 2, alpha = 0.5 loss of easy examples is downweighted.
"""
inputs = logit(
torch.tensor([[[0.95], [0.90], [0.6], [0.3]]],
dtype=torch.float64))
targets = torch.tensor([[[1], [1], [1], [1]]], dtype=torch.float64)
focal_loss = sigmoid_focal_loss(inputs, targets, gamma=2, alpha=0.5)
ce_loss = F.binary_cross_entropy_with_logits(inputs,
targets,
reduction="none")
loss_ratio = (ce_loss / focal_loss).squeeze()
correct_ratio = 2.0 / ((1.0 - inputs.squeeze().sigmoid())**2)
self.assertTrue(np.allclose(loss_ratio, correct_ratio))
def test_easy_ex_focal_loss_less_than_ce_loss(self) -> None:
"""
With gamma = 2 loss of easy examples is downweighted.
"""
N = 5
inputs = logit(torch.rand(N))
targets = torch.randint(0, 2, (N, )).float()
focal_loss = sigmoid_focal_loss(inputs, targets, gamma=2, alpha=-1)
ce_loss = F.binary_cross_entropy_with_logits(inputs,
targets,
reduction="none")
loss_ratio = (ce_loss / focal_loss).squeeze()
prob = torch.sigmoid(inputs)
p_t = prob * targets + (1 - prob) * (1 - targets)
correct_ratio = 1.0 / ((1.0 - p_t)**2)
self.assertTrue(np.allclose(loss_ratio, correct_ratio))
def test_sum_focal_loss_equals_ce_loss(self) -> None:
"""
Sum of focal loss across all examples matches ce loss.
"""
inputs = logit(
torch.tensor([[[0.05], [0.12], [0.89], [0.79]]],
dtype=torch.float32))
targets = torch.tensor([[[1], [1], [0], [0]]], dtype=torch.float32)
focal_loss = sigmoid_focal_loss(inputs,
targets,
gamma=0,
alpha=-1,
reduction="sum")
ce_loss = F.binary_cross_entropy_with_logits(inputs,
targets,
reduction="sum")
self.assertEqual(ce_loss, focal_loss)
def score_box_loss(target_scores, target_deltas, pred_logits, pred_deltas):
# target_scores: [B, k]
# target_deltas: [B, k, 4]
# pred_scores: [B, k]
# pred_deltas: [B, k, 4]
score_loss = sigmoid_focal_loss(pred_logits,
target_scores,
reduction='sum')
maxscore, _ = target_scores.max(-1)
pos_mask = maxscore > 0.5
target_box_deltas_pos = target_deltas[pos_mask].reshape(-1, 2, 2)
pred_box_deltas_pos = pred_deltas[pos_mask].reshape(-1, 2, 2)
box_loss = euclidean_loss(pred_box_deltas_pos, target_box_deltas_pos,
'sum')
npos = target_box_deltas_pos.shape[0]
return score_loss / npos, box_loss / npos
def test_mean_focal_loss_equals_ce_loss(self) -> None:
"""
Mean value of focal loss across all examples matches ce loss.
"""
inputs = logit(
torch.tensor(
[[0.05, 0.9], [0.52, 0.45], [0.89, 0.8], [0.39, 0.5]],
dtype=torch.float32,
))
targets = torch.tensor([[1, 0], [1, 0], [1, 1], [0, 1]],
dtype=torch.float32)
focal_loss = sigmoid_focal_loss(inputs,
targets,
gamma=0,
alpha=-1,
reduction="mean")
ce_loss = F.binary_cross_entropy_with_logits(inputs,
targets,
reduction="mean")
self.assertEqual(ce_loss, focal_loss)
def focal_loss(self):
from fvcore.nn import sigmoid_focal_loss
self._log_accuracy()
bg_class_ind = self.pred_class_logits.shape[1] - 1
fg_inds = (self.gt_classes >= 0) & (self.gt_classes < bg_class_ind)
num_foreground = fg_inds.nonzero().numel()
gt_classes_target = torch.zeros_like(self.pred_class_logits)
gt_classes_target[fg_inds, self.gt_classes[fg_inds]] = 1
loss_cls = sigmoid_focal_loss(
self.pred_class_logits[:, 0:bg_class_ind],
gt_classes_target[:, 0:bg_class_ind],
alpha=self.focal_loss_alpha,
gamma=self.focal_loss_gamma,
reduction="sum",
) / max(1, num_foreground)
# logger.info("___________________________softmax cross entropy loss would have been: {}_________________________"
# "__".format(F.cross_entropy(self.pred_class_logits, self.gt_classes, reduction="mean")))
# logger.info("classification loss divided by: {}".format(max(1, num_foreground)))
return loss_cls
def losses(self, gt_clses, gt_belongs, gt_masks, pred_logits, pred_points,
ins_feature):
"""
Args:
For `gt_clses`, `gt_belongs`, `gt_masks` parameters, see
:meth:`TensorMask.get_ground_truth`.
For `pred_logits` and `pred_points`, see
:meth:`TensorMaskHead.forward`.
Returns:
losses (dict[str: Tensor]): mapping from a named loss to a scalar tensor
storing the loss. Used during training only. The potential dict keys are:
"loss_cls", and "loss_mask".
"""
num_fg = gt_belongs.size(0)
w, h = pred_points.size(2), pred_points.size(3)
max_side = w * 0.5 if w > h else h * 0.5
# print(num_fg,torch.sum(gt_clses))
loss_normalizer = torch.tensor(max(1, num_fg),
dtype=torch.float32,
device=self.device)
# classification
loss_cls = (sigmoid_focal_loss(
pred_logits,
gt_clses,
alpha=self.focal_loss_alpha,
gamma=self.focal_loss_gamma,
reduction="sum",
) / loss_normalizer)
losses = {
"loss_cls": loss_cls,
}
# mask prediction
loss_mask = 0
if num_fg < 1:
losses = {
"loss_cls": loss_cls,
"loss_mask": torch.sum(pred_points * 0)
}
return losses
pred_points_valids = pred_points[gt_belongs[:, 0], :, gt_belongs[:, 1],
gt_belongs[:, 2]]
N, P = pred_points_valids.size()
pred_location = pred_points_valids.detach().view(N, P // 2, 2)
location = (
pred_location + gt_belongs[:, 1:].view(N, 1, 2)
) * self.classify_feature_strides[-1] / self.ins_feature_strides[0]
location = location.float()
pred_points_valids = pred_points_valids.view(N, P // 2, 2, 1)
Q = gt_masks.size(1)
gt_masks = gt_masks.view(N, 1, Q, 2)
gt_masks = gt_masks.transpose(2, 3)
pred_points_valids_contour = pred_points_valids
batch_index = gt_belongs[:, 0]
# print(pred_points_valids_contour.size(),gt_masks.size())
#chamfer distance:
# loss_mask=chamfer_loss(pred_points_valids_contour,gt_masks)*self.mask_loss_weight
#normlized chamfer distance
loss_mask1 = normlize_chamfer_loss(
pred_points_valids_contour, gt_masks,
max_side=max_side) * self.mask_loss_weight
dist = emd_l1_loss2(pred_points_valids_contour, gt_masks)
loss_mask2 = torch.mean(dist)
points_delta = self.ins_head(ins_feature, location, batch_index)
contour_weight = torch.zeros_like(points_delta) + 0.1
contour_weight[:, :81 * 3, :] = 1
loss_delta = torch.mean(
torch.abs(points_delta - dist) * contour_weight)
losses["loss_mask"] = loss_mask1 + loss_mask2
losses["loss_delta"] = loss_delta
return losses
for img_data, labels_data in tqdm(dataloader_train):
labels_data[labels_data > 1] = 1
img_label = labels_data.numpy().astype(np.float)
optimizer.zero_grad()
if opt.gpu_id >= 0:
img_data = img_data.cuda()
labels_data = labels_data.cuda().float()
classes = model(img_data)
loss_dis = sigmoid_focal_loss(classes,
labels_data.view(-1, 1),
alpha=0.25,
gamma=2,
reduction="mean")
loss_dis_data = loss_dis.item()
loss_dis.backward()
optimizer.step()
classes = classes.view(1, -1).squeeze(0)
classes = torch.sigmoid(classes)
output_dis = classes.data.cpu().numpy()
tol_label = np.concatenate((tol_label, img_label))
tol_pred = np.concatenate((tol_pred, output_dis))
loss_train += loss_dis_data
def losses(
self,
gt_clses,
gt_belongs,
gt_masks,
gt_ins,
pred_logits,
pred_points,
ins_feature
):
"""
Args:
For `gt_clses`, `gt_belongs`, `gt_masks` parameters, see
:meth:`TensorMask.get_ground_truth`.
For `pred_logits` and `pred_points`, see
:meth:`TensorMaskHead.forward`.
Returns:
losses (dict[str: Tensor]): mapping from a named loss to a scalar tensor
storing the loss. Used during training only. The potential dict keys are:
"loss_cls", and "loss_mask".
"""
num_fg=gt_belongs.size(0)
w,h=pred_points.size(2),pred_points.size(3)
max_side=w*0.5 if w>h else h*0.5
# print(num_fg,torch.sum(gt_clses))
loss_normalizer = torch.tensor(max(1, num_fg), dtype=torch.float32, device=self.device)
# classification
loss_cls = (
sigmoid_focal_loss(
pred_logits,
gt_clses,
alpha=self.focal_loss_alpha,
gamma=self.focal_loss_gamma,
reduction="sum",
)
/ loss_normalizer
)
losses = {"loss_cls": loss_cls, }
# mask prediction
loss_mask = 0
if num_fg<1:
losses= {"loss_cls": loss_cls, "loss_mask":torch.sum(pred_points*0)}
return losses
pred_points_valids=pred_points[gt_belongs[:,0],:,gt_belongs[:,1],gt_belongs[:,2]]
N,P=pred_points_valids.size()
# pred_location=(pred_points_valids.view(N,P//2,2)+gt_belongs[:,1:].view(N,1,2))*4
pred_points_valids=pred_points_valids.view(N,P//2,2,1)
Q=gt_masks.size(1)
location=(gt_masks+gt_belongs[:,1:].view(N,1,2))*self.classify_feature_strides[-1]/self.ins_feature_strides[0]
location=location.float()
gt_masks=gt_masks.view(N,1,Q,2)
gt_masks=gt_masks.transpose(2,3)
pred_points_valids_contour=pred_points_valids
batch_index=gt_belongs[:,0]
pred_ins=self.ins_head(ins_feature,location,batch_index)
# _,_,h,w=gt_ins.size()
pred_ins=F.interpolate(pred_ins,scale_factor=self.ins_feature_strides[0],mode='bilinear').squeeze(1)
loss_ins=F.binary_cross_entropy_with_logits(pred_ins,gt_ins)*self.ins_loss_weight
losses['loss_ins'] = loss_ins
# print(pred_points_valids_contour.size(),gt_masks.size())
#chamfer distance:
# loss_mask=chamfer_loss(pred_points_valids_contour,gt_masks)*self.mask_loss_weight
#normlized chamfer distance
# loss_mask=normlize_chamfer_loss(pred_points_valids_contour,gt_masks,max_side=max_side)*self.mask_loss_weight
loss_mask=emd_l1_loss(pred_points_valids_contour,gt_masks)*self.mask_loss_weight
losses["loss_mask"] = loss_mask
return losses
def losses(
self,
gt_clses,
gt_belongs,
gt_masks,
pred_logits,
pred_points,
):
"""
Args:
For `gt_clses`, `gt_belongs`, `gt_masks` parameters, see
:meth:`TensorMask.get_ground_truth`.
For `pred_logits` and `pred_points`, see
:meth:`TensorMaskHead.forward`.
Returns:
losses (dict[str: Tensor]): mapping from a named loss to a scalar tensor
storing the loss. Used during training only. The potential dict keys are:
"loss_cls", and "loss_mask".
"""
num_fg = gt_belongs.size(0)
w, h = pred_points.size(2), pred_points.size(3)
max_side = w * 0.5 if w > h else h * 0.5
# print(num_fg,torch.sum(gt_clses))
loss_normalizer = torch.tensor(max(1, num_fg),
dtype=torch.float32,
device=self.device)
# classification
loss_cls = (sigmoid_focal_loss(
pred_logits,
gt_clses,
alpha=self.focal_loss_alpha,
gamma=self.focal_loss_gamma,
reduction="sum",
) / loss_normalizer)
losses = {
"loss_cls": loss_cls,
}
# mask prediction
loss_mask = 0
if num_fg < 1:
losses = {
"loss_cls": loss_cls,
"loss_mask": torch.sum(pred_points * 0)
}
return losses
pred_points_valids = pred_points[gt_belongs[:, 0], :, gt_belongs[:, 1],
gt_belongs[:, 2]]
N, P = pred_points_valids.size()
pred_points_valids = pred_points_valids.view(N, P // 2, 2, 1)
Q = gt_masks.size(1)
gt_masks = gt_masks.view(N, 1, Q, 2)
gt_masks = gt_masks.transpose(2, 3)
pred_points_valids_contour = pred_points_valids
if self.circumscribed:
num_points = P // 2
edge = np.sqrt(num_points)
# top=[0+i for i in range(edge)]
# left=[i*edge+0 for i in range(1,edge-1)]
# right=[i*edge+edge-1 for i in range(1,edge-1)]
# bottom=[(edge-1)*edge+i for i in range(edge)]
# Index=top+left+right+bottom
All = [i * edge + j for i in range(edge) for j in range(edge)]
Inner = [
i * edge + j for i in range(1, edge - 1)
for j in range(1, edge - 1)
]
Out = list(set(All) - set(Inner))
Out = torch.tensor(Out, dtype=torch.int64, device=self.device)
pred_points_valids_contour = pred_points_valids[:, Out, :, :]
# Inner=torch.tensor(Inner,dtype=torch.int64,device=self.device)
# pred_points_valids_inner=pred_points_valids[:,Inner,:,:]
# print(pred_points_valids_contour.size(),gt_masks.size())
#chamfer distance:
# loss_mask1=chamfer_loss(pred_points_valids_contour,gt_masks)*self.mask_loss_weight
#normlized chamfer distance
loss_mask1 = normlize_chamfer_loss(
pred_points_valids_contour, gt_masks,
max_side=max_side) * self.mask_loss_weight
loss_mask2 = emd_l1_loss(pred_points_valids_contour,
gt_masks) * self.mask_loss_weight
losses["loss_mask"] = loss_mask1 + loss_mask2
return losses
