sgugger/deep-learning

A few notebooks about deep learning in pytorch

deep-learning's Issues

DeepPainterlyHarmonization notebook error: IndexError: too many indices for tensor of dimension 1

Hi Sylvain, thanks for the awesome exposition in your DeepPainterlyHarmonization notebook. I am however running into an issue during the second training phase. Not changing anything, was able to reproduce the code up until In[84].
With the following warning:

IndexError: too many indices for tensor of dimension 1

Here's the detailed warning:

---

IndexError Traceback (most recent call last)
in ()
1 n_iter=0
----> 2 while n_iter <= max_iter: optimizer.step(partial(step,final_loss))

~/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/torch/optim/lbfgs.py in step(self, closure)
101
102 # evaluate initial f(x) and df/dx
--> 103 orig_loss = closure()
104 loss = float(orig_loss)
105 current_evals = 1

in step(loss_fn)
2 global n_iter
3 optimizer.zero_grad()
----> 4 loss = loss_fn(opt_img_v)
5 loss.backward()
6 n_iter += 1

in final_loss(opt_img_v)
4 c_loss = content_loss(out_ftrs[-1])
5 s_loss = style_loss(out_ftrs)
----> 6 h_loss = hist_loss([out_ftrs[0], out_ftrs[3]])
7 t_loss = tv_loss(opt_img_v[0])
8 return c_loss + w_s * s_loss + w_h * h_loss + w_tv * t_loss

in hist_loss(out_ftrs)
6 mask = V(torch.Tensor(mf).contiguous().view(1, -1), requires_grad=False)
7 of_masked = of * mask
----> 8 of_masked = torch.cat([of_masked[i][mask>=0.1].unsqueeze(0) for i in range(of_masked.size(0))])
9 loss += F.mse_loss(of_masked, V(remap_hist(of_masked, sh), requires_grad=False))
10 return loss / 2

in (.0)
6 mask = V(torch.Tensor(mf).contiguous().view(1, -1), requires_grad=False)
7 of_masked = of * mask
----> 8 of_masked = torch.cat([of_masked[i][mask>=0.1].unsqueeze(0) for i in range(of_masked.size(0))])
9 loss += F.mse_loss(of_masked, V(remap_hist(of_masked, sh), requires_grad=False))
10 return loss / 2

IndexError: too many indices for tensor of dimension 1

to_gpu(x, *args, **kwargs) 'numpy.ndarray' object has no attribute 'cuda'

Hi, I try to run the code of the website (https://github.com/sgugger/Deep-Learning/blob/master/DeepPainterlyHarmonization.ipynb).

When in the cell 22, I met a error.
In [22]: m_vgg(VV(input_tfm[None]))
input_ftrs = [s.features for s in sfs]
[sf.shape for sf in input_ftrs]

d:\python3.6.5\lib\site-packages\fastai\core.py in VV(x)
20 def V(x): return [V_(o) for o in x] if isinstance(x,list) else V_(x)
21 def VV_(x): return to_gpu(x, async=True) if isinstance(x, Variable) else Variable(to_gpu(x, async=True), volatile=True)
---> 22 def VV(x): return [VV_(o) for o in x] if isinstance(x,list) else VV_(x)
23
24 def to_np(v):

d:\python3.6.5\lib\site-packages\fastai\core.py in VV_(x)
19 def V_(x): return to_gpu(x, async=True) if isinstance(x, Variable) else Variable(to_gpu(x, async=True))
20 def V(x): return [V_(o) for o in x] if isinstance(x,list) else V_(x)
---> 21 def VV_(x): return to_gpu(x, async=True) if isinstance(x, Variable) else Variable(to_gpu(x, async=True), volatile=True)
22 def VV(x): return [VV_(o) for o in x] if isinstance(x,list) else VV_(x)
23

d:\python3.6.5\lib\site-packages\fastai\core.py in to_gpu(x, *args, **kwargs)
28 def to_gpu(x, *args, **kwargs):
29 if torch.cuda.is_available():
---> 30 return x.cuda(*args, **kwargs)
31 else:
32 return x

AttributeError: 'numpy.ndarray' object has no attribute 'cuda'

the programe is runing in the windows 10 with GTX Titan X.
The software environment is : fastai 0.6 troch 0.4.0 cuda 9.1 + v7.1

I would like to know why ?
Did I installed the software with the wrong version?
How can I fix it ?

1cycle Policy. Unfamiliar results

Hey,

I was implementing 1 cycle policy as an exercise. And I have a few observations from my experiments.
I have a
Model : Resnet18.
Batch size for training = 128
Batch size for testing = 100

Optimser : optim.SGD(net.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4)
Total number of epochs 26

1 cycle policy : Learning rate goes from 0.01 to 0.1 and back till 24 epochs

Then model is trained for 2 epochs at 0.001 learning rate.

No cyclic momentum used or adamw.

I achieved a test set accuracy of 93.4%in 26 epochs.

This seems like a big difference from the 70 epochs at 512 batch size that is quoted in your blog post.

Am I doing something wrong ? Is the number of epochs a good metric to base your results on, as those are dependant on the batch size ? .

The whole point of using super convergence is using high learning rates to converge quicker , but it seems like using low learning rates (0.01- 0.1 < 0.8-3) is faster to train.

Use with pytorch model

Can you give me an example of how to use with with my own dataset function. For instance your getdata() function uses ImageClassifierData module of fastai. I have created a pytorch dataset but I'm not sure how to wrap it for this function

ImageClassifierDataset no longer exists?

when I try to fun your notebook I get:

NameError: name 'ImageClassifierData' is not defined

what is the cause of this issue?

Thanks for this incredible work

I wish more people published notebooks like this. Thanks a lot.