skelemoa/synse-zsl

Official PyTorch code for the ICIP 2021 paper 'Syntactically Guided Generative Embeddings For Zero Shot Skeleton Action Recognition'

synse-zsl's Issues

RuntimeError: DataLoader worker (pid(s) 27447) exited unexpectedly

Hi, I'm yet to train SynSE from scratch because this error keep happening intermittently, but always before reaching epoch 1700, often around 1200. Did it ever happen to you?

python3 synse_training.py --ntu=60 --phase=train --ss=5 --st=r --dataset=/home/dario/temp/synse_resources/ntu_results/shift_5_r/ --ve=shift --wdir=/home/dario/temp/synse_resources/language_modelling/synse_5_r_retrain --le=bert --mode=train --num_cycles=10 --num_epoch_per_cycle=1700 --latent_size=100 --gpu=0

at synse/synse_training.py

def train_one_cycle(cycle_num, cycle_length):
...
            (inputs, target) = next(iter(train_loader))

RuntimeError: DataLoader worker (pid(s) 27447) exited unexpectedly

What is gtest_out.npy?

In synse_resources/ntu_results/shift_10_r/ we have
g_label.npy - 50906 NTU120 action labels
gtest.npy - 50906 ShiftGCN 256D features
and gtest_out.npy - a 50906x120 array between [-19.3308 39.108074]

What is gtest_out.npy? The output of the ShiftGCN classification layer (can't be, it's 120d, and you said you trained ShiftGCN only on the 110 seen classes)?
Why is the size 50906? (NTU120 has in total 52 940 samples, and only 302 are samples with missing skeletons)

What is synse_10_r_unseen_zs.npy and synse_10_r_seen_zs.npy

In synse_resources/language_modelling/repo_test_synse_10_r_val/bert/ we have
synse_10_r_unseen_zs.npy - of shape (4821, 10) and min max [ 0.0077835037 , 0.47233814 ]
and synse_10_r_seen_zs.npy - of shape (3240, 10) and min max [ 0.009033936 , 0.7786457 ]

My guess for synse_10_r_unseen_zs.npy is the unseen classes probabilities as output of softmax (SynSEclassifier(sequence_encoder(ShiftGCN_embeddings))) passed by (is this essentially correct?)
But I have no guess for synse_10_r_seen_zs.npy, given it's also of width 10

What is val*.npy, where does it come from, and how do you determined the sample size of it?

Hi,

I wonder where does val_label.npy, val_out.npy, and val.npy came from?
My guess is that they are randomly drawn from the validation split of the seen classes in the val setting.
For example, for NTU 55/5, we also take the 50/5 split in the 55 seen classes for training the gate module, and the val*.npys are taken from the validation split of the 50 seen classes.

I have another question about the size of val*.npys:

• for NTU 55/5, you take 2000 samples
• for NTU 48/12, you take 1440 samples
• for NTU 110/10, you take 3240 samples
• for NTU 96/24, you take 7200 samples.

I wonder if I could know how do you decide how many samples to take?
Thanks in advance!

[Question] Why did ShiftGCN perform so much better than MS-G3D?

Hi,

You performed ablation tests comparing using ShiftGCN vs MS-G3D, and the resulting accuracy drop was striking.

However MS-G3D published results slightly outperform ShiftGCN on the NTU-60 action recognition task.

Do you have thoughts on why ShiftGCN performed so much better in this case? (Was MS-G3D at least faster?)

Is there any Inference to test on an Image or a video?

Extra spaces in examples

At https://github.com/skelemoa/synse-zsl/blob/master/synse/README.md there is an extra space at the dataset argument and the ve argument of each example
-- dataset 'synse_resources/ntu_results/shift_5_r/' --wdir 'synse_resources/language_modelling/repo_test_synse_5_r' --le bert -- ve

[Questions] Only calling Dataloader once per epoch? Training classifier from scratch at each cycle?

at synse/synse_training.py

def train_one_cycle(cycle_num, cycle_length):
...
    for epoch in range(s_epoch, e_epoch):
...
        (inputs, target) = next(iter(train_loader))


if __name__ == "__main__":
...
            train_one_cycle(num_cycle, cycle_length)
...
            zsl_acc, val_out_embs, cls = train_classifier()

the dataloader is only called once per epoch, so only 64 samples are used for training each epoch, and a classifier is trained from scratch after each cycle.

• Did you test training with all training samples per epoch and concluded only training with a few samples per epoch gave better results?

• Did you test training the classifier from where it left off last cycle and concluded training from scratch each time gave better results?

Rename num_epoch_per_cycle to num_cycle_per_epoch

At https://github.com/skelemoa/synse-zsl/blob/master/synse/README.md and in the code it says

Argument	Possible Values	Description
num_epoch_per_cycle	Integer	Number of epochs per cycle 1700 for 5 random and 1900 for others

but in the examples we have: --num_cycle_per_epoch

Please do that refactoring in the code (renaming num_epoch_per_cycle to num_cycle_per_epoch) or revert the change in the examples in the README.md

Confused about some folders

in /synse_resources/ntu_results,what is the different between shift_val_24 and shif_24.
And the test_label in the two folders are not the same.
the labels in shift_val_24 match with the ReadMe,but the other one is not.

NTU split

Hi, how you split the NTU-60 dataset. Did you use the files in the CS split or the CV split? Did you remove any samples?
Thanks

How to replicate ablation test with MS-G3D?

How to replicate ablation test with MS-G3D?

Running this command:
synse_training.py --ntu=60 --phase=train --ss=5 --st=r --wdir=/home/dario/temp/synse_resources/language_modelling/repo_test_synse_5_r --le=bert --load_classifier --num_cycles=10 --num_epoch_per_cycle=1700 --latent_size=100 --load_epoch=3399 --mode=eval --gpu=0
--dataset=/home/dario/temp/synse_resources/ntu_results/msg3d_5_r/
--ve=msg3d

yields at

def load_models(load_epoch):
    sequence_encoder.load_state_dict(torch.load(se_checkpoint)['state_dict'])

Exception has occurred: RuntimeError
Error(s) in loading state_dict for Encoder:
	size mismatch for mu_transformation.0.weight: copying a param with shape torch.Size([100, 256]) from checkpoint, the shape in current model is torch.Size([100, 384]).
	size mismatch for logvar_transformation.0.weight: copying a param with shape torch.Size([100, 256]) from checkpoint, the shape in current model is torch.Size([100, 384]).

se_checkpoint == '/home/dario/temp/synse_resources/language_modelling/repo_test_synse_5_r/bert/se_3399.pth.tar'

In synse_resources rs10.npy == rs100.npy

Just a heads up:

In synse_resources/resources/label_splits rs10.npy is exactly the same as rs100.npy

some questions about visual feature extractor

Hello, I follow this paper and use shift-gcn to extract the visual features.
There are some questions.
1、I use the seen classes to train the shift-gcn, but without the test data, how can i choose the best epoch?
2、For each stream in shift-gcn, i can get a type of visual embedding. So how to get the 4s embeddding? just add them up in a ratio of [0.6,0.6,0.4,0.4]?
3、the best.pt in synse_resources/ntu_results/shift_5_r/weights/ is the weight for all the 4s shift-gcn models?
(maybe each stream has a best.pt?)

thank you for your work