skip this point data_size = 1

[rharadgithub]

Dear sir,When I run the script:bash ./scripts/train_edges2shoes.sh,the following RuntimeError occurs.
(epoch: 1, iters: 49400, time: 0.311) , z_encoded_mag: 0.577, G_total: 4.293, G_L1_encoded: 2.367, z_L1: 0.259, KL: 0.076, G_GAN: 1.002, D_GAN: 0.498, G_GAN2: 0.589, D_GAN2: 0.988
(epoch: 1, iters: 49600, time: 0.322) , z_encoded_mag: 0.409, G_total: 2.001, G_L1_encoded: 0.385, z_L1: 0.302, KL: 0.069, G_GAN: 0.794, D_GAN: 0.960, G_GAN2: 0.450, D_GAN2: 1.137
(epoch: 1, iters: 49800, time: 0.311) , z_encoded_mag: 0.939, G_total: 3.441, G_L1_encoded: 1.597, z_L1: 0.373, KL: 0.079, G_GAN: 0.833, D_GAN: 0.774, G_GAN2: 0.560, D_GAN2: 1.015
skip this point data_size = 1
Traceback (most recent call last):
File "./train.py", line 28, in
model.update_G()
File "/home/rharad/junyanz/BicycleGAN/models/bicycle_gan_model.py", line 148, in update_G
self.backward_EG()
File "/home/rharad/junyanz/BicycleGAN/models/bicycle_gan_model.py", line 114, in backward_EG
self.loss_G.backward(retain_graph=True)
File "/home/rharad/anaconda3/lib/python3.6/site-packages/torch/autograd/variable.py", line 167, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, retain_variables)
File "/home/rharad/anaconda3/lib/python3.6/site-packages/torch/autograd/init.py", line 99, in backward
variables, grad_variables, retain_graph)
RuntimeError: Trying to backward through the graph a second time, but the buffers have already been freed. Specify retain_graph=True when calling backward the first time.

How can I solve this problem.....

[junyanz]

Could you report your PyTorch version? The current code only works well with PyTorch 0.1-0.3.

[rharadgithub]

@junyanz ,Dear
In [1]: import torch

In [2]: print(torch.version)
0.3.1.post2
I try
Firstly ,torch version=0.3.1.post2 with Anaconda3(python 3.6.4)
Secondly,torch version=0.2.0_1 with python 2.7.12
Thirdly,torch version=0.2.0_0 with anaconda2(python 2.7.14)
Fourthly,torch version=0.3.0.post4 with anaconda3(python 3.6.4)

But the problem remains same like before.

I am eagerly waiting for your reply.
Thank you.

[HelenMao]

I meet the same problems with @rharadgithub

[HelenMao]

Is the problem of the number of the datasets of edge2shoes
the number of trainsets in edge2shoes is 49825
however, the batch_size = 2
so the code in def forward
self.skip = self.opt.isTrain and self.input_A.size(0) < self.opt.batchSize if self.skip: print('skip this point data_size = %d' % self.input_A.size(0)) return
will return, so the bug is that there is no forward function. @rharadgithub

In fact the code
self.real_A_random = self.real_A[half_size:]
and
self.real_B_random = self.real_B[half_size:]
are not used at all
I suggest to delete the codes so that it will not be any bug when training the edge2shoes codes.
@junyanz

[rharadgithub]

Dear @HelenMao ,yes I am trying for datasets of edge2shoes. According to your suggestion,I just delete the L38 and L40 and run the scripts:bash ./scripts/train_edges2shoes.sh but still the following error
Traceback (most recent call last):
File "./train.py", line 27, in
model.update_D(data)
File "/home/hr/BicycleGAN/models/bicycle_gan_model.py", line 117, in update_D
self.forward()
File "/home/hr/BicycleGAN/models/bicycle_gan_model.py", line 45, in forward
self.z_random = self.get_z_random(self.real_A_random.size(0), self.opt.nz, 'gauss')
AttributeError: 'BiCycleGANModel' object has no attribute 'real_A_random'

On the other hand,When I just delete " self.skip = self.opt.isTrain and self.input_A.size(0) < self.opt.batchSize if self.skip: print('skip this point data_size = %d' % self.input_A.size(0)) return" keeping the L38 and L40 and run the scripts:bash ./scripts/train_edges2shoes.sh but still the following error:
Traceback (most recent call last):
File "./train.py", line 27, in
model.update_D(data)
File "/home/hr/BicycleGAN/models/bicycle_gan_model.py", line 119, in update_D
self.forward()
File "/home/hr/BicycleGAN/models/bicycle_gan_model.py", line 38, in forward
self.real_A_random=self.real_A[half_size:]
File "/home/hr/anaconda3/lib/python3.6/site-packages/torch/autograd/variable.py", line 78, in getitem
return Index.apply(self, key)
File "/home/hr/anaconda3/lib/python3.6/site-packages/torch/autograd/_functions/tensor.py", line 89, in forward
result = i.index(ctx.index)
ValueError: result of slicing is an empty tensor

Am I wrong with your suggestion?

[HelenMao]

you can’ delete the L38 and L40 directly, the most convenient way is that you delete one image from train dataset from 49825 into 49824.

the number of the dataset should be even @rharadgithub

[rharadgithub]

@HelenMao ,I tried to train maps dataset,it has even number of samples and its working. But for odd number of samples,the problem remains(AttributeError: 'BiCycleGANModel' object has no attribute 'real_A_random') even after deleting L38 and L40.

[HelenMao]

If you delete the L38 and L40 directly, the latter codes which use of these variables will be effected. And this is why it has arrtibuteError. @rharadgithub

I don't get the point why Junyan write the codes including the variable self.real_A_random. @junyanz

[twak]

I think I fixed it in this patch pull request #24

[junyanz]

Thanks for the bug report. I think I fixed the issue with the latest commit. self.real_A_random is sometimes used in conditional D case. I refactored the code. self.real_A_random will not be used for unconditional D.

[HelenMao]

@junyanz thanks for your replying, I am also wondering why in the unconditional D
self.real_data_random = self.real_B_random
why we need use self.real_B_random? I think it is another picture, why we need another picture in using the pair, self.real_A_encoded and self.real_B_encoded
@junyanz

[WorkingCC]

I have the same question as HelenMao. By the way, why is the batchsize set to two instead of one?
@junyanz

[rharadgithub]

Thanks to @junyanz

[junyanz]

@HelenMao @WorkingCC Sorry for the late reply. This is a minor implementation detail that should not significantly affect the results. We chose to use different images for training cVAE-GAN and training cLR-GAN. That is also the reason for batchSize=2.

[WorkingCC]

@junyanz Thank you very much for your reply .