Some problems when training on PCN

[leejingbei]

1. When data goes trough PCN network, it comes out the error 'Out of Memory'. And this error is located at class 'global_network', 'out = out.repeat(1, 1, int(self.image_size / n), int(self.image_size / n))' in PCN.py. Do you mean convert the feature with size [1, 512, 256, 256] to a feature [1, 512, 32, 32] in 'global512' block? Why don't you use upsampling method?

2. I changed the 'global_network' in PCN.py to let PCN network work and counted the total 'd_loss' and 'g_loss' for each epoch. But 'd_loss' didn't change since iteration 16 and 'g_loss' changed very very tiny. Any idea to fix this problem?

[iamxinxin]

Hi, I also met this question "cuda out of memory" when trainning PCN. Have you solved this problem? And could you share how did you change the 'global_network' to make it work to me?

[leejingbei]

Hi, I also met this question "cuda out of memory" when trainning PCN. Have you solved this problem? And could you share how did you change the 'global_network' to make it work to me?

I commented the code 'out = out.repeat(1, 1, int(self.image_size / n), int(self.image_size / n))' and change the definition of convolution layer 'self.twoD', 'self.threeD' , 'self.fourD' in global_network, to apply the size of the output of 'conv4', 'conv8', 'conv9'.

Hope it would help you.

[iamxinxin]

Hi, I also met this question "cuda out of memory" when trainning PCN. Have you solved this problem? And could you share how did you change the 'global_network' to make it work to me?

I commented the code 'out = out.repeat(1, 1, int(self.image_size / n), int(self.image_size / n))' and change the definition of convolution layer 'self.twoD', 'self.threeD' , 'self.fourD' in global_network, to apply the size of the output of 'conv4', 'conv8', 'conv9'.

Hope it would help you.

Thanks for your reply.
I have changed the code in the way similar to your recommendation, but the d_loss and g_loss almost unchanged after few iteration same as your description.
Have you fixed it?

[leejingbei]

Hi, I also met this question "cuda out of memory" when trainning PCN. Have you solved this problem? And could you share how did you change the 'global_network' to make it work to me?

I commented the code 'out = out.repeat(1, 1, int(self.image_size / n), int(self.image_size / n))' and change the definition of convolution layer 'self.twoD', 'self.threeD' , 'self.fourD' in global_network, to apply the size of the output of 'conv4', 'conv8', 'conv9'.
Hope it would help you.

Thanks for your reply.
I have changed the code in the way similar to your recommendation, but the d_loss and g_loss almost unchanged after few iteration same as your description.
Have you fixed it?

Not yet. So I am wondering if something wrong in this part...

[FloatingFowl]

@leejingbei I used functional.interpolate and the code works, however the training process is very slow (estimated time for one epoch on the bird256 dataset is around a day...). Have you managed to get it working somehow?

[AhmedMohamedAliAli]

@leejingbei change the definition of convolution layer 'self.twoD', 'self.threeD' , 'self.fourD' in global_network, to apply the size of the output of 'conv4', 'conv8', 'conv9'.how to do this

[zjqccccz]

嗨，我在训练PCN时也遇到了“ CUDA内存不足”这个问题。你解决了这个问题吗？您能否分享您如何更改“ global_network”以使其对我有用？

我注释了代码'out = out.repeat（1，1，int（self.image_size / n），int（self.image_size / n））'并更改了卷积层'self.twoD'，'self的定义。 global_network中的threeD'和'self.fourD'，以应用'conv4'，'conv8'，'conv9'的输出大小。
希望对您有帮助。

感谢您的回复。
我已经按照与您的建议类似的方式更改了代码，但是经过几次迭代后，与您的描述相同，d_loss和g_loss几乎没有变化。
你修好了吗？

还没。所以我想知道这部分是否有问题...

HI ,Have you fixed this problem now?

[Aliang-CN]

i have fixed this problem。

[israelpf]

i have fixed this problem。

How??

[HyunJai]

Hi, How can I solve the below problems?

1. UserWarning: Using a target size (torch.Size([32])) that is different to the input size (torch.Size([32,1])) is deprecates.
2. RuntimeError: CUDA error: out of memory

[HyunJai]

i have fixed this problem。

How did yo fix it ??

[dinosaur279]

i have fixed this problem。

How did yo fix it ??

Did you fix that and i had the same problem as you

[JuanseHevia]

hey guys, is there an official solution to this?