Hi @carlpe ,

I think you should move the value of expected key in the dict to device, instead of moving dict itself.
Thanks.

It seems to be working now!

However I have run into a new issue, this time it is to do with the size parameters in the code.

Thanks for suggestions and feedback on this.

Hi @ericspod , could you please help take a look this issue?
It's related to the GAN networks.

Thanks.

The discriminator's in_shape parameter states the size of input images, your images are different from the 64x64 images the example uses so you'll have to adjust that value. The start_shape parameter of the generator is the shape of the tensor after the input latent vector is forwarded through the linear layer, this is going to be starting shape of the image which is upsampled by the layers specified in channels and strides. Since strides is [2, 2, 2, 1] this means the starting image will be upsampled by 3 transpose convolutions with strides of 2, so it will be 8 times the size of start_shape. You want to choose a size for start_shape which is 1/8th the size of the final output image you want in the spatial dimensions.

Hi Eric, thanks for reply.

I am not sure I understand everything yet. But does it make sense that because the inputs are 3D images with 400,400,400, I would need more to get it to work?

I assume that Discriminator in-shape should be 1,400,400 and start_shape 1,50,50 for the Generator. Given that I keep strides [2, 2, 2, 1]. I did try to run this with 1,400,400 and 1,50,50 .. and I got a new error:

RuntimeError: Expected 4-dimensional input for 4-dimensional weight [8, 1, 5, 5], but got 5-dimensional input of size [1, 1, 1, 64, 64] instead.

If you have 3D inputs the discriminator and generator need to be given 3D sizes. The in_shape value should then be (1,400,400,400) which is a volume 400 cubed with a single channel. The start_shape for the generator should then be (1,50,50,50) or maybe even with a channel size larger than 1. I would also choose a much larger latent_size than what was in the original notebook, for a larger image size you want a larger latent size so that it has enough capacity to represent variation.

Good. I will try it tomorrow.

Thanks 👍

I did manage to get around that error now. Using 1,400,400,400 and 1,50,50,50

However, it is too demanding for video memory. How would I proceed if I wanted to preprocess the images to smaller sizes before I ship them in?

RuntimeError: CUDA out of memory. Tried to allocate 74.51 GiB (GPU 0; 23.65 GiB total capacity; 2.67 MiB already allocated; 22.79 GiB free; 4.00 MiB reserved in total by PyTorch)

So, I resized all inputs to 64,64,64 - and now everything seems to be working. Although I wish there was some way to do this with the original size.,.

Thanks for help

Memory is often a challenge with 3D inputs, if you can't fit a batch of 1 into memory the alternative is to train on patches. One workflow is https://github.com/Project-MONAI/MONAI/blob/master/examples/workflows/unet_training_dict.py which uses RandCropByPosNegLabeld uring training and SlidingWindowInferer during evaluation.

Would that work with GAN also?

Depends on the exact GAN configuration you're using. If you train the discriminator on patches I could see it learning to tell real from synthetic effectively, but a generator would have a hard time learning how to create a coherent image patch-wise.

Yes, I see...

I will try to find a balanced size that fits into memory for now I think..

It seems to be working now!
How you cope up with this prob??
However I have run into a new issue, this time it is to do with the size parameters in the code.

Thanks for suggestions and feedback on this.

I think this is from from the in_shape value of the Discriminator which needs to be the size of the input images, so I think that's (1,400,400) for you.