mogvision / adl Goto Github PK

Hi,

Thanks for your amazing work.

I would like to ask a problem about reading model from checkpoint.

It seems that the code couldn't load the model correctly from some checkpoints and it always start to train from scratch.

For example, yesterday I trained the model for six hours (29,200 iters completed), and today I use the same training config.

The output states that it correctly store the model from step = 29,200, but the first iteration of this second training is still step: <<<<< 100/108650 >>>>>, and the validation and testing psnr isn't at the level of 29,200 iters too.

Is there any thing that I have to modify for checkpoint loading (e.g. in train_json ?) or I miss something here?

Thanks in advance.

Hello!
I have a question about training time.
How long does the training take?

I'm reproducing ADL.
I am training ADL with ImageNet, following the training details in ADL paper.
But according to my calculations, this training may take a few hundred days with default batch size=4, even though I modified the period of test from 200 to 4000000.

I think this training time doesn't make sense because it's too long.
Even if I modify the batch size from 4 to 32 or more, this training is expected to take at least several tens of days.

So, I wonder how long did your traing take on ImageNet? And what did you set the batch size to?
Or is there any other detailed setting for training?

Thank you for your reading!

File "E:\kuaiyaWJ\Misc\DNCNN+STARGAN\ADLcnn\trainer.py", line 133, in train
loss.backward()
File "F:\anaconda520\envs\pytorch\lib\site-packages\torch_tensor.py", line 255, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File "F:\anaconda520\envs\pytorch\lib\site-packages\torch\autograd_init_.py", line 149, in backward
allow_unreachable=True, accumulate_grad=True) # allow_unreachable flag
RuntimeError: the derivative for 'histc' is not implemented.

Just like the error reported above, when I was using 'torch. Histc' to perform 'backward', the problem appeared.
Could you help me with that, please? Thank you very, very much

Hello!

I can't find the pre-trained models, especially 2D model, in github of ADL.
Can you release the pre-trained 2D model?

I'm looking forward to hearing your answer.
Thank you for your reading!

Hello Team,
I was curious to know how can we denoise the already noisy image.
what changes are expected

I haven't been able to stabilise the GAN using the existing hyper parameters and any tweaks in the same.
Denoiser and discriminator train to a loss of ~0.05 during their respective warm ups. However, GAN training takes the discriminator loss on fake samples to ~1e-7 and then the denoiser stops training altogether.

Could someone share any insights on stabilising the GAN ?
Here's my implementation. https://github.com/jayantb1019/adl_seismic

Hi,
Could you share the pre-trained models? I am particularly interested in the 3D brain MRI ones for Pytorch.
Thanks!

Hi,

Great work. Please correct me if I am wrong, but shouldn't

img_dirs['y'] = [files.replace('/HR/', '/LR/') for files in x_files]

actually be:

img_dirs['y'] = [files.replace('/HR/', '/LR/') for files in img_dirs['x']]

in pytorch/utils/dataloader.py line 201

The x_files variable is not defined before.

Similar typo in line 204 and the corresponding dataloader_test file.

Cheers,
Bashir