sunnydreamrain / indrnn_pytorch Goto Github PK

Hi ... Thanks alot for this fabulous work. I'm still trying to understand the code and and trying to run this in my machine.
The main issue I'm having is, I'm not able to prepare the dataset as I required.

I'm trying to understand the following code

datasets=train_datasets
dataname=datasets+'.npy'
labelname=datasets+'_label.npy'
lenname=datasets+'_len.npy'
data_handle=np.load(dataname)
label_handle=np.load(labelname)
len_handle=np.load(lenname)

What is this train_datasets should be? What's it shape need to be? Can you elaborate more on this. Because I need to arrange it to feed to the network

Hi, is there any chance to implement a cuda version of IndRNNCell? the purpose is to speed up processing variable length sequences.

RuntimeError: Legacy autograd function with non-static forward method is deprecated. Please use new-style autograd function with static forward method. (Example: https://pytorch.org/docs/stable/autograd.html#torch.autograd.Function)

Hi, thanks for your great work! But I have encountered something wrong while reproducing the code.
During "train", the grad of the most layers is "None". Except layers "classify_weight"、"classify_bias"、"RNN5_weight"、"RNN5_bias" have not-None grad, others have grad which are None. As a result, error happens when running to "grad_climp", as showed in the following figure.
Maybe something goes wrong with layer "RNN5_weight_hh" during loss.backward() I think.

I wonder how to address this problem. Looking forward to the reply, thank you!

Hi,

Could u please explain how to replace RNN with LSTM?

Thank u

Hi,

Thanks for the wonderful work. I tried to reproduce the results for Action Recognition task on NTU RGB+D dataset on the subject split, by running the provided command.

• python -u Indrnn_action_train.py --dropout 0.25 --use_weightdecay_nohiddenW
  The maximum Accuracy I am able to reach is 72.48%.
  Is there something needed to be changed to get the desired results?

First of all, thanks a lot for this great work!

I am trying to reproduce results with resIndRNN on word level PTB data, however following the recommended settings in the paper I was only able to get around 60 perplexity (in the paper it is around 59). Would it be possible for you to share also the configuration similar to the denseIndRNN case? Thanks a lot in advance!

By the way, to have deterministic behavior, I would also add the following three lines in train_language.py, see PyTorch note on reproducibility.

np.random.seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

Also, please correct me if I am wrong, but while reading your Theano implementation, it seems that the resIndRNN implemented there uses the original resnet configuration (dense then activation layer) while this version uses the new pre-activation configuration (activation then dense layer). Would this be the reason for the different results?

Hello,
Thank you verymuch for sharing your work.
I have a question about input shape
[48000, 300,50,3] input shape --> As per I understood ,300 represents sequence length of one .skeleton file . When we have .skeletion files such as 154 ,155 sequence lengths ,did you fill the rest of the length with zero padding to make it as 300 sequence length ?
Will it affects the accuracy of action recognition?

Thank you and waiting for your reply.

hello, thanks for your excellent work. I notice that you implementation on action recognition is different from the paper formulated. I want to know why you have those changes in your codebase?
ps: your code is return F.relu(input + hx * self.weight_hh.unsqueeze(0).expand(hx.size(0), len(self.weight_hh))) can you explain this code ? hope for your reply

I noticed that there are word-level PTB results in the paper but I only find character level in the repo. Is there any folder for the word level PTB?