Chapter 6: find_lr(), train() - RuntimeError Output size is too small

[MarcusFra]

Hi,

after having inserted the missing train_loader argument in the find_lr() function (we should open a PR for this after having solved this error), I get a RuntimeError in the current version of Chapter6.ipynb.

After running

torch.save(audionet.state_dict(), "audionet.pth")
optimizer = optim.Adam(audionet.parameters(), lr=0.001)
### added: , train_loader, device="cuda"
logs,losses = find_lr(audionet, nn.CrossEntropyLoss(), optimizer, train_loader, device="cuda")
plt.plot(logs,losses)

I get this error:

---------------------------------------------------------------------------

RuntimeError                              Traceback (most recent call last)

<ipython-input-47-d21b94029709> in <module>()
      2 optimizer = optim.Adam(audionet.parameters(), lr=0.001)
      3 ### added: , train_loader, device="cuda"
----> 4 logs,losses = find_lr(audionet, nn.CrossEntropyLoss(), optimizer, train_loader, device="cuda")
      5 plt.plot(logs,losses)

4 frames

<ipython-input-41-3824bd6072de> in find_lr(model, loss_fn, optimizer, train_loader, init_value, final_value, device)
     49         targets = targets.to(device)
     50         optimizer.zero_grad()
---> 51         outputs = model(inputs)
     52         loss = loss_fn(outputs, targets)
     53 

/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    530             result = self._slow_forward(*input, **kwargs)
    531         else:
--> 532             result = self.forward(*input, **kwargs)
    533         for hook in self._forward_hooks.values():
    534             hook_result = hook(self, input, result)

<ipython-input-44-db3925c72c5e> in forward(self, x)
     31         x = F.relu(self.bn4(x))
     32         x = self.pool4(x)
---> 33         x = self.avgPool(x)
     34         x = x.squeeze(-1)
     35         x = self.fc1(x)

/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    530             result = self._slow_forward(*input, **kwargs)
    531         else:
--> 532             result = self.forward(*input, **kwargs)
    533         for hook in self._forward_hooks.values():
    534             hook_result = hook(self, input, result)

/usr/local/lib/python3.6/dist-packages/torch/nn/modules/pooling.py in forward(self, input)
    484         return F.avg_pool1d(
    485             input, self.kernel_size, self.stride, self.padding, self.ceil_mode,
--> 486             self.count_include_pad)
    487 
    488 

RuntimeError: Given input size: (512x1x1). Calculated output size: (512x1x0). Output size is too small

The same happens when I run

train(audionet, optimizer, torch.nn.CrossEntropyLoss(),train_loader, valid_loader, epochs=20)

I ran the code in Colab.

[falloutdurham]

Hi - I have a little update on quantizing models hopefully coming at the weekend, so I'll take a look into this then as well (and upload the casper pic!).

[falloutdurham]

My email is ian@snappishproductions.com - I didn't realize it was being translated!