Cannot input grayscale images

[gvenkat21]

Hi,

I have a folder of grayscale images that I want to use for training. I am using the following command:

python train.py --dataroot ./datasets/sample --name sample --loadSize 1536 --fineSize 512 --no_flip --batchSize 3 --niter 3000 --gpu_ids 1 --no_instance --resize_or_crop resize_and_crop --label_nc 0 --n_local_enhancers 1 --norm batch --use_dropout --input_nc 1 --output_nc 1

But I am getting the following error:

RuntimeError: Given groups=1, weight of size [64, 1, 7, 7], expected input[3, 3, 518, 518] to have 1 channels, but got 3 channels instead

What is the issue?

[leweohlsen]

Hi,

the issue is that in the DataLoader, inputs and outputs are always converted to RGB like this, even when you specify --input_nc 1 --output_nc 1.

You could remove the converting or make it conditional but RGB images seem to be required for the Vgg19 model. You could probably go with --no_vgg_loss.

Looks like we need a custom DataLoader for grayscale images?

[gvenkat21]

Ok...thanks...I’ll try your suggestion

[TravisFar]

Hi, I will be very grateful if you tell how prepare your dataset as grayscale images to feed the pix2pixhd?
I am working on medical images so I have 3 modalities and I wanna synthesize another one but I dont know how I should prepare dataset for this. By preparing, I mean, creating input matrix. thank you for your time,

[WenliangDu]

Hi,

I have a folder of grayscale images that I want to use for training. I am using the following command:

python train.py --dataroot ./datasets/sample --name sample --loadSize 1536 --fineSize 512 --no_flip --batchSize 3 --niter 3000 --gpu_ids 1 --no_instance --resize_or_crop resize_and_crop --label_nc 0 --n_local_enhancers 1 --norm batch --use_dropout --input_nc 1 --output_nc 1

But I am getting the following error:

RuntimeError: Given groups=1, weight of size [64, 1, 7, 7], expected input[3, 3, 518, 518] to have 1 channels, but got 3 channels instead

What is the issue?

Hi

I solved the --input_nc 1 problem by editing the data/base_dataset.py (for transforms), data/aligned_dataset.py (for transforms) and models/pix2pixHD_model.py (for VGG loss). You can get the updated version in https://github.com/WenliangDu/pix2pixHD, and find the changes in the three codes with comments of "# for --input_nc 1"

[1999kevin]

Hi,
I have a folder of grayscale images that I want to use for training. I am using the following command:
python train.py --dataroot ./datasets/sample --name sample --loadSize 1536 --fineSize 512 --no_flip --batchSize 3 --niter 3000 --gpu_ids 1 --no_instance --resize_or_crop resize_and_crop --label_nc 0 --n_local_enhancers 1 --norm batch --use_dropout --input_nc 1 --output_nc 1
But I am getting the following error:
RuntimeError: Given groups=1, weight of size [64, 1, 7, 7], expected input[3, 3, 518, 518] to have 1 channels, but got 3 channels instead
What is the issue?

Hi

I solved the --input_nc 1 problem by editing the data/base_dataset.py (for transforms), data/aligned_dataset.py (for transforms) and models/pix2pixHD_model.py (for VGG loss). You can get the updated version in https://github.com/WenliangDu/pix2pixHD, and find the changes in the three codes with comments of "# for --input_nc 1"

Do you have any idea for input_nc 3(RGB) and output_nc 1(gray)? I have tried your code and it doesn't work.

[WenliangDu]

Hi,
I have a folder of grayscale images that I want to use for training. I am using the following command:
python train.py --dataroot ./datasets/sample --name sample --loadSize 1536 --fineSize 512 --no_flip --batchSize 3 --niter 3000 --gpu_ids 1 --no_instance --resize_or_crop resize_and_crop --label_nc 0 --n_local_enhancers 1 --norm batch --use_dropout --input_nc 1 --output_nc 1
But I am getting the following error:
RuntimeError: Given groups=1, weight of size [64, 1, 7, 7], expected input[3, 3, 518, 518] to have 1 channels, but got 3 channels instead
What is the issue?

Hi
I solved the --input_nc 1 problem by editing the data/base_dataset.py (for transforms), data/aligned_dataset.py (for transforms) and models/pix2pixHD_model.py (for VGG loss). You can get the updated version in https://github.com/WenliangDu/pix2pixHD, and find the changes in the three codes with comments of "# for --input_nc 1"

Do you have any idea for input_nc 3(RGB) and output_nc 1(gray)? I have tried your code and it doesn't work.

Thank you for pointing the bug. I modified Line 54 in aligned_dataset.py. You could try this code for your issue.

"transform_B = get_transform(self.opt, params, grayscale=(self.opt.input_nc == 1))"=>"transform_B = get_transform(self.opt, params, grayscale=(self.opt.output_nc == 1))"

[1999kevin]

Thank you very much!

Hi,

I have a folder of grayscale images that I want to use for training. I am using the following command:

python train.py --dataroot ./datasets/sample --name sample --loadSize 1536 --fineSize 512 --no_flip --batchSize 3 --niter 3000 --gpu_ids 1 --no_instance --resize_or_crop resize_and_crop --label_nc 0 --n_local_enhancers 1 --norm batch --use_dropout --input_nc 1 --output_nc 1

But I am getting the following error:

RuntimeError: Given groups=1, weight of size [64, 1, 7, 7], expected input[3, 3, 518, 518] to have 1 channels, but got 3 channels instead

What is the issue?

Hi

I solved the --input_nc 1 problem by editing the data/base_dataset.py (for transforms), data/aligned_dataset.py (for transforms) and models/pix2pixHD_model.py (for VGG loss). You can get the updated version in https://github.com/WenliangDu/pix2pixHD, and find the changes in the three codes with comments of "# for --input_nc 1"

Do you have any idea for input_nc 3(RGB) and output_nc 1(gray)? I have tried your code and it doesn't work.

Thank you for pointing the bug. I modified Line 54 in aligned_dataset.py. You could try this code for your issue.

"transform_B = get_transform(self.opt, params, grayscale=(self.opt.input_nc == 1))"=>"transform_B = get_transform(self.opt, params, grayscale=(self.opt.output_nc == 1))"