Error while training efficientdet
Hi,
I am new to the problem of object detection and currently learning more about it. I am trying to train EfficientDet for this problem. However, I am getting this error even after resizing the images to the same size along with the bboxes using albumentations. I am using Ross Wightman's pytorch implementation of EfficientDet. Could someone please help me out here?
Thanks in advance!
Batchsize = 4
images resized to 224
0%| | 0/4349 [00:00<?, ?it/s]
torch.Size([4, 3, 224, 224])
0%| | 0/4349 [00:02<?, ?it/s]
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
/tmp/ipykernel_331/3869611587.py in <module>
1 for i in range(CFG.EPOCHS):
----> 2 avg_train_loss = train_model(train_loader, model, optimizer)
3 avg_val_loss = val_model(val_loader, model)
/tmp/ipykernel_331/3403888848.py in train_model(dataloader, model, optimizer)
18
19 with torch.autograd.set_detect_anomaly(True):
---> 20 output = model(images, input_targets)
21 print(output)
22
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1188 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1189 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1190 return forward_call(*input, **kwargs)
1191 # Do not call functions when jit is used
1192 full_backward_hooks, non_full_backward_hooks = [], []
/opt/conda/lib/python3.7/site-packages/effdet/bench.py in forward(self, x, target)
131
132 def forward(self, x, target: Dict[str, torch.Tensor]):
--> 133 class_out, box_out = self.model(x)
134 if self.anchor_labeler is None:
135 # target should contain pre-computed anchor labels if labeler not present in bench
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1188 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1189 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1190 return forward_call(*input, **kwargs)
1191 # Do not call functions when jit is used
1192 full_backward_hooks, non_full_backward_hooks = [], []
/opt/conda/lib/python3.7/site-packages/effdet/efficientdet.py in forward(self, x)
601 def forward(self, x):
602 x = self.backbone(x)
--> 603 x = self.fpn(x)
604 x_class = self.class_net(x)
605 x_box = self.box_net(x)
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1188 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1189 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1190 return forward_call(*input, **kwargs)
1191 # Do not call functions when jit is used
1192 full_backward_hooks, non_full_backward_hooks = [], []
/opt/conda/lib/python3.7/site-packages/effdet/efficientdet.py in forward(self, x)
344 for resample in self.resample.values():
345 x.append(resample(x[-1]))
--> 346 x = self.cell(x)
347 return x
348
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1188 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1189 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1190 return forward_call(*input, **kwargs)
1191 # Do not call functions when jit is used
1192 full_backward_hooks, non_full_backward_hooks = [], []
/opt/conda/lib/python3.7/site-packages/effdet/efficientdet.py in forward(self, x)
34 def forward(self, x: List[torch.Tensor]) -> List[torch.Tensor]:
35 for module in self:
---> 36 x = module(x)
37 return x
38
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1188 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1189 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1190 return forward_call(*input, **kwargs)
1191 # Do not call functions when jit is used
1192 full_backward_hooks, non_full_backward_hooks = [], []
/opt/conda/lib/python3.7/site-packages/effdet/efficientdet.py in forward(self, x)
277 def forward(self, x: List[torch.Tensor]):
278 for fn in self.fnode:
--> 279 x.append(fn(x))
280 return x[-self.num_levels::]
281
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1188 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1189 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1190 return forward_call(*input, **kwargs)
1191 # Do not call functions when jit is used
1192 full_backward_hooks, non_full_backward_hooks = [], []
/opt/conda/lib/python3.7/site-packages/effdet/efficientdet.py in forward(self, x)
238
239 def forward(self, x: List[torch.Tensor]) -> torch.Tensor:
--> 240 return self.after_combine(self.combine(x))
241
242
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1188 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1189 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1190 return forward_call(*input, **kwargs)
1191 # Do not call functions when jit is used
1192 full_backward_hooks, non_full_backward_hooks = [], []
/opt/conda/lib/python3.7/site-packages/effdet/efficientdet.py in forward(self, x)
219 weights_sum = torch.sum(edge_weights)
220 out = torch.stack(
--> 221 [(nodes[i] * edge_weights[i]) / (weights_sum + 0.0001) for i in range(len(nodes))], dim=-1)
222 elif self.weight_method == 'sum':
223 out = torch.stack(nodes, dim=-1)
RuntimeError: stack expects each tensor to be equal size, but got [4, 64, 4, 4] at entry 0 and [4, 64, 8, 8] at entry 1
It's a shape mismatch error. If you are using Data Transforms then try to add a manual transforms.Resize() in transforms.Compose()
You may refer to this link:- https://discuss.pytorch.org/t/runtimeerror-stack-expects-each-tensor-to-be-equal-size-but-got-3-224-224-at-entry-0-and-3-224-336-at-entry-3/87211
As I can get only this information from the stack trace, so if possible try to share the notebook.
Hi @Raghvender , thank you for the reply.
I have resized the images using albumentations. I tried again by setting the batch size=1 to check if the images are not getting resized properly, but im still getting the same error.
Notebook link - https://www.kaggle.com/code/sagarikajadon/ship-effdet/notebook
Let me check and get back to you
As you are using efficientdet_d0, you need to use the image size defined in model_config (512x512)
Because, even if you resize your image to 224, the model is build for input of size 512.
You have two option, either you change your CFG to (512, 512) or you can git clone efficientdet-pytorch and edit directly the model_config.py file.
Resolved it, thank you @Yohann84L