Machine-Learning-Collection
Machine-Learning-Collection copied to clipboard
Published
20 hours ago •
aladdinpersson
aladdinpersson
Output becomes zero after optimizer.step() yolo-v1 model
I encountered this error while i was trying train the model on my local gpu
Here : Machine-Learning-Collection/ML/Pytorch/object_detection/YOLO/
This is the test script that i have used to test the yolo-v1 model
if __name__ == '__main__':
csv_file_path = 'PascalVOC_YOLO/100examples.csv'
img_dir = 'PascalVOC_YOLO/images'
label_path = 'PascalVOC_YOLO/labels'
learning_rate = 1e-10
num_workers = 2
batch_size = 2
weight_decay = 1e-4
sample_dataset = VOCDataset( csv_file_path , img_dir , label_path, transform=transform)
sample_loader = DataLoader(
dataset=sample_dataset,
batch_size=2,
num_workers=2,
pin_memory=True,
shuffle=True,
drop_last=True,
)
device = 'cuda'if torch.cuda.is_available() else 'cpu'
model = Yolov1(split_size=7, num_boxes=2, num_classes=20).to(device).half()
optimizer = optim.Adam( model.parameters() , lr=learning_rate , weight_decay=weight_decay)
loss_func = YoloLoss().to(device)
for _ in range(2):
print('iter : ',_,'\n')
x , y = next( iter(sample_loader) )
x , y = Variable(x).to(device).half() , Variable(y).to(device).half()
# print( 'infinite : ' , torch.isfinite(x))
# print('x : ',x)
out = model(x)
print('out : ',out , '\n')
loss = loss_func(out,y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print( 'loss : ',loss , '\n')
print( 'loss : data ', loss.data , '\n')
print(' loss : grad ',loss.grad , '\n')
for name, param in model.named_parameters():
print(name, torch.isfinite(param.grad).all() , torch.max(abs(param.grad)) )
print('\n')
Note : i am using half() because of the cuda error => RuntimeError: cuDNN error: CUDNN_STATUS_EXECUTION_FAILED
while running the script i was getting this output below
iter : 0
out : tensor([[-0.1432, 0.0819, 0.0342, ..., -0.0377, -0.0745, 0.1312],
[ 0.1110, -0.0650, 0.2410, ..., -0.0765, 0.3328, 0.1908]],
device='cuda:0', dtype=torch.float16, grad_fn=<AddmmBackward>)
loss : tensor(1., device='cuda:0', dtype=torch.float16, grad_fn=<ClampBackward>)
loss : data tensor(1., device='cuda:0', dtype=torch.float16)
test.py:94: UserWarning: The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the gradient for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations.
print(' loss : grad ',loss.grad , '\n')
loss : grad None
darknet.0.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.0.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.0.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.2.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.2.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.2.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.4.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.4.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.4.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.5.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.5.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.5.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.6.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.6.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.6.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.7.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.7.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.7.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.9.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.9.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.9.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.10.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.10.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.10.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.11.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.11.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.11.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.12.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.12.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.12.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.13.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.13.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.13.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.14.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.14.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.14.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.15.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.15.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.15.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.16.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.16.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.16.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.17.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.17.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.17.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.18.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.18.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.18.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.20.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.20.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.20.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.21.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.21.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.21.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.22.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.22.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.22.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.23.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.23.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.23.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.24.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.24.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.24.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.25.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.25.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.25.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.26.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.26.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.26.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.27.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.27.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.27.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
fcs.1.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
fcs.1.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
fcs.4.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
fcs.4.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
iter : 1
out : tensor([[nan, nan, nan, ..., nan, nan, nan],
[nan, nan, nan, ..., nan, nan, nan]], device='cuda:0',
dtype=torch.float16, grad_fn=<AddmmBackward>)
[W python_anomaly_mode.cpp:104] Warning: Error detected in MseLossBackward. Traceback of forward call that caused the error:
File "test.py", line 86, in <module>
loss = loss_func(out,y)
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/mnt/e/workspace/@training/@datasets/cnns/yolo/yolo-v1-pytorch/loss.py", line 120, in forward
torch.flatten(exists_box * target[..., :20], end_dim=-2,),
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/nn/modules/loss.py", line 528, in forward
return F.mse_loss(input, target, reduction=self.reduction)
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/nn/functional.py", line 2929, in mse_loss
return torch._C._nn.mse_loss(expanded_input, expanded_target, _Reduction.get_enum(reduction))
(function _print_stack)
Traceback (most recent call last):
File "test.py", line 89, in <module>
loss.backward()
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/tensor.py", line 245, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/autograd/__init__.py", line 147, in backward
allow_unreachable=True, accumulate_grad=True) # allow_unreachable flag
RuntimeError: Function 'MseLossBackward' returned nan values in its 0th output.
(dev) buckaroo@hansolo:/mnt/e/workspace/@training/@datasets/cnns/yolo/yolo-v1-pytorch$ python3 test.py
iter : 0
out : tensor([[-0.1044, -0.3135, -0.4897, ..., -0.1079, -0.0055, -0.0380],
[ 0.1190, -0.3154, -0.0910, ..., -0.0995, -0.1595, -0.0576]],
device='cuda:0', dtype=torch.float16, grad_fn=<AddmmBackward>)
loss : tensor(1.0010, device='cuda:0', dtype=torch.float16, grad_fn=<AddBackward0>)
loss : data tensor(1.0010, device='cuda:0', dtype=torch.float16)
test.py:94: UserWarning: The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the gradient for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations.
print(' loss : grad ',loss.grad , '\n')
loss : grad None
darknet.0.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.0.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.0.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.2.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.2.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.2.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.4.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.4.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.4.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.5.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.5.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.5.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.6.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.6.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.6.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.7.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.7.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.7.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.9.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.9.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.9.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.10.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.10.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.10.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.11.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.11.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.11.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.12.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.12.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.12.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.13.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.13.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.13.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.14.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.14.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.14.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.15.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.15.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.15.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.16.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.16.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.16.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.17.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.17.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.17.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.18.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.18.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.18.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.20.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.20.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.20.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.21.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.21.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.21.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.22.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.22.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.22.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.23.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.23.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.23.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.24.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.24.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.24.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.25.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.25.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.25.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.26.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.26.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.26.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.27.conv.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.27.batchnorm.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
darknet.27.batchnorm.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
fcs.1.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
fcs.1.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
fcs.4.weight tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
fcs.4.bias tensor(True, device='cuda:0') tensor(0., device='cuda:0', dtype=torch.float16)
iter : 1
out : tensor([[nan, nan, nan, ..., nan, nan, nan],
[nan, nan, nan, ..., nan, nan, nan]], device='cuda:0',
dtype=torch.float16, grad_fn=<AddmmBackward>)
Observations
- As you can see that for the first time , the model out is a valid tensor with values ( i.e before
optimizer.step()) - when the iteration 1 begins ( i.e after
optimizer.step()) output becomesnan
Debug method : 0
- after setting this
torch.autograd.set_detect_anomaly(True)globally - i found this result below
[W python_anomaly_mode.cpp:104] Warning: Error detected in MseLossBackward. Traceback of forward call that caused the error:
File "test.py", line 86, in <module>
loss = loss_func(out,y)
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/mnt/e/workspace/@training/@datasets/cnns/yolo/yolo-v1-pytorch/loss.py", line 120, in forward
torch.flatten(exists_box * target[..., :20], end_dim=-2,),
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/nn/modules/loss.py", line 528, in forward
return F.mse_loss(input, target, reduction=self.reduction)
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/nn/functional.py", line 2929, in mse_loss
return torch._C._nn.mse_loss(expanded_input, expanded_target, _Reduction.get_enum(reduction))
(function _print_stack)
Traceback (most recent call last):
File "test.py", line 89, in <module>
loss.backward()
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/tensor.py", line 245, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File "/home/buckaroo/miniconda3/envs/dev/lib/python3.7/site-packages/torch/autograd/__init__.py", line 147, in backward
allow_unreachable=True, accumulate_grad=True) # allow_unreachable flag
RuntimeError: Function 'MseLossBackward' returned nan values in its 0th output.
so i have tried
- clamping the loss tensors
torch.clamp(value, min=0.0 , max=1.0)inloss.py - adding epsilon (1e-6) after torch.sqrt() like
torch.sqrt(val+epsilon)inloss.py
But this didnot fix my issue.
Reference
Getting NaN values in backward pass
Output of Model is nan every time
Nan Loss coming after some time
Getting Nan after first iteration with custom loss
Weights become NaN values after first batch step
NaN values popping up during loss.backward()
So kindly help me debug this issue , thanks in advance
