def compute_loss(ouputs, y, criterion):
"""Compute the loss given the ouput, ground truth y, and the criterion function.
Hint: criterion should be cross entropy.
Hint: the input is a list of tensors, each has shape as (batch_size, vocab_size)
Hint: you need concat the tensors, and reshape its size to (batch_size*num_step, vocab_size).
Then compute the loss with y.
Returns:
output: A 0-d tensor--averaged loss (scalar)
"""
### YOUR CODE HERE
#print("output in compute_loss length: ", len(ouputs))
print("initial single output shape in compute_loss length: ", ouputs[0].shape)
for element in range(len(ouputs)):
if element == 0:
output = ouputs[element]
else:
output = torch.cat((output,ouputs[element]),0)
output = torch.reshape(output,(y.size(dim=0)*len(ouputs),ouputs[0].size(dim=1)))
print("y in compute_loss: ", y.shape)
print("Final output shape:",output.shape)
print("\n",output[0].shape)
print("criterion in compute_loss: ", criterion(output,y))
### END YOUR CODE
return loss
I am getting the output as below but with an error:
initial single output shape in compute_loss length: torch.Size([64, 10000])
y in compute_loss: torch.Size([64, 10])
Final output shape: torch.Size([640, 10000])
The error is :
/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py in cross_entropy(input, target, weight, size_average, ignore_index, reduce, reduction, label_smoothing)
3012 if size_average is not None or reduce is not None:
3013 reduction = _Reduction.legacy_get_string(size_average, reduce)
-> 3014 return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing)
3015
3016
ValueError: Expected input batch_size (640) to match target batch_size (64).
I can see the final output shape is torch.Size([640, 10000]) and I am sending that to criterion(output,y) but it is still asking for a batch of 640 whereas the output shape is torch.Size([640, 10000])
