abdo-eldesokey / nconv Goto Github PK

Hey there.
I'm a little confused by which datasets you used for training exp_guided_nconv_cnn_l2, it seems to be train+selval according to run-nconv-cnn.py#L67 .
But it's validate on selval too.

Also what's the difference between nconv/workspace/exp_guided_nconv_cnn_l2/checkpoints/CNN_ep0039.pth.tar and the one for test set, which datasets do you train it on.

Hi,
I have a question about the code.
When I train the model on KittiDataset, I find that the program puts a lot of data to the cached memory until the memory is full.
when i use "top" command in Terminal, I see that: (I use the exp=exp_guided_enc_dec)

Have you ever encountered this situation? Did you notice that? (Although it might not affect training? )
Many thanks!

Thank you for sharing this wonderful work,

When I go to the nyu repo, ican't find the nyu pretrained weight. Can you provide it?

Hi, I have a few questions about your implementation.

1. I saw that you already had a pre-trained model for unguided_network.py. How to re-produce this pretrained model?
2. I didn't see you use the SoftPlus function in unguided_network.py. Can you explain it to me? In your paper, your unguided network was proposed to use this function.

This is the exp I use : exp_unguided_depth
But got this:

TypeError: init() missing 1 required positional argument: 'padding_mode'

Could you help us figure it out?

The demo works very well when I choose -exp exp_guided_nconv_cnn_l2

Hi,
Is there an error in the evaluate function in kittidepthtrainer?

outputs *= self.params['data_normalize_factor']/256
labels *= self.params['data_normalize_factor']/256

why do you divide outputs by 256? I think the output should be just multiplied by data_normmalize_factor.

Many thanks!

Hi!
When I run:

 python3 run-nconv-cnn.py -mode eval -exp /home/chen/nconv/workspace/exp_guided_nconv_cnn_l1

This is a error:

Traceback (most recent call last):
  File "run-nconv-cnn.py", line 56, in <module>
    dataloaders, dataset_sizes = KittiDepthDataloader(params)
  File "/home/chen/nconv/dataloader/KittiDepthDataloader.py", line 44, in KittiDepthDataloader
    dataloaders['train'] = DataLoader(image_datasets['train'], shuffle=True, batch_size=params['train_batch_sz'], num_workers=4)
  File "/home/chen/.local/lib/python3.5/site-packages/torch/utils/data/dataloader.py", line 802, in __init__
    sampler = RandomSampler(dataset)
  File "/home/chen/.local/lib/python3.5/site-packages/torch/utils/data/sampler.py", line 64, in __init__
    "value, but got num_samples={}".format(self.num_samples))
ValueError: num_samples should be a positive integeral value, but got num_samples=0

Could you tell me why?
Thanks!

Hello,

I am very interested in running your evaluation code with the trained model. Will you provide it soon?

Thanks,
Adam

hi, recently I looked into your code and got doubt on how to obtain a confidence map from the sparse depth map. Please help me to go through this doubt, Thanks in advance.

Hello ! When retraining the code for NYU dataset, the original file nconv.py is as followed:

but when running, it meets a problem that reminds me that missing argument:

And after I add the argument "padding_mode = 'zeros' ",it reminds me size dismatch.

So I'm confused and wondering how to fix this problem. Looking forward to your reply.

Hi,
I have a question about the EnforcePos part in file ncon.py.
Why do you choose to use a softplus function as a forward_pre_hook, instead of just use softplus in forward function?
My understanding is that, after lots of training iteration, the forward_pre_hook(softplus) will work on weights multiple times, which makes W = softplus(••••••softplus(W)) ? (althouth you set the beta=10, very close to RELU? it will be almost an identity map when W > 0?) And for last training iteration, after the backward optimization, we do not apply the softplus to weights. (Which I may think that the weights are already converged?)
But I am still confused about that...
Can we just simply apply softplus in forward function, for both train and eval phase?
Many thanks!

Hi, Abdelrahman!
I notice that in normalized convolution the basis is set to only one basis constant function f(x)=1. I wonder why we don't have 2 or more basis functions and what would happen than. I would be grateful if you would like to share your view on the basis setting.