vinhkhuc / pytorch-mini-tutorials Goto Github PK

doing .forward I believe is not recommended because it doesn't end up using masks or handle or something like that I forget...just heads up

Hi @vinhkhuc ,

Would you be interested in a PR for the cuda version of these 4 models? I made some simple modifications to make them work on GPUs, while trying to preserve the flow of the original non-GPU versions.

Example 1 & 2 probably doesn't need a GPU version since they train quite fast anyway.

Thx

Hi Vinh Khuc,

Thanks for creating this wonderful tutorial for pytorch. I was trying your examples to learn pytorch. For linear regression, I was trying a variant so that I understand better the inner working of pytorch. So, instead of using single independent variable, I'm using multiple independent variable (2 Xs, X1 and X2). I created them as following in the training function. Rest of the code is exactly the same as you have created.

X1 = torch.linspace(-2, 2, 101)
X2 = torch.linspace(-1, 1, 101)
X = torch.cat(([X1, X2])).resize_(100,2)

But when I'm using batch_size = 10, I'm getting the following error:

RuntimeError Traceback (most recent call last)
in ()
4 for k in range(num_batches):
5 start, end = k * batch_size, (k + 1) * batch_size
----> 6 cost += train(model, loss, optimizer, X[start:end], Y[start:end])
7 print("Epoch = %d, cost = %s" % (i + 1, cost / num_batches))
8

in train(model, loss, optimizer, x, y)
7
8 # Forward
----> 9 fx = model.forward(x.view(len(x), 1))
10 output = loss.forward(fx, y)
11

C:\Program Files\Anaconda3\lib\site-packages\torch\autograd\variable.py in view(self, *sizes)
466
467 def view(self, *sizes):
--> 468 return View.apply(self, sizes)
469
470 def view_as(self, tensor):

C:\Program Files\Anaconda3\lib\site-packages\torch\autograd_functions\tensor.py in forward(ctx, i, sizes)
87 ctx.new_sizes = sizes
88 ctx.old_size = i.size()
---> 89 result = i.view(*sizes)
90 ctx.mark_shared_storage((i, result))
91 return result

RuntimeError: size '[10 x 1]' is invalid for input of with 20 elements at D:\Downloads\pytorch-master-1\torch\lib\TH\THStorage.c:59

So, I'm not understanding, why I'm getting this error. If you could guide me then it will be a great help to me.

Thank you!

The pytorch document shows:

class torch.nn.CrossEntropyLoss(weight=None, size_average=True)[source]
This criterion combines LogSoftMax and NLLLoss in one single class.

SO could I remove softmax layer from the model sequential?

model.add_module("softmax", torch.nn.Softmax())

Hi there,

Thanks for sharing this great set of tutorials. I made a few tiny changes to make it work for Python 3.x. Would you be interested in a pull request?

Changes:

1. use integer division for calculating minibatch size
2. use urllib.reqeust.urlretrieve() if python major version is > 2
3. download MNIST files if any of the data files is missing, instead of just checking folder existence.

Thanks.