Open pcc03 opened 3 months ago
Do you load the wlasl_all_attn-187.pt at the begining of training, or resume from it to restart the training?
I restarted the training from the beginning. Also, I didn't find the wlasl_all_attn-187.pt in the repository. Could you please let me know where I can download it?
BTW, I also tested the pretrained model "./pretrained_models/pretrained_model_for_WLASL2000.pt". The accuracy is also very low. I am confused that whether the provided one is a workable model.
Could you give some hints that whether I did wrong to train and test the WLASL?
I will check the code and respond later.
Thank you! Appreciate it.
I have trained on WLASL2000 and WLASL300, and obtain correct results. The log files are log_wlasl300_joint.txt and log_wlasl2000_joint.txt. But when i test with the saved weights, i also get 0.1% accuracy. I check the code and still don't find any problem. I will keep tracking the problems.
Thank you for your reply! By comparing the log files, I found that I need to set the phase=train. Otherwise, it is the test phase by default. I can successfully train the WLASL300 now.
But one more questions is that, when I try to train on the WLASL100 using the provided npy and pkl files, the following error appears. Could you please help me with this?
/home/mssn/DSTA-SLR/model/fstgan.py:114: UserWarning: nn.init.constant is now deprecated in favor of nn.init.constant_. nn.init.constant(self.Linear_bias, 1e-6) Attention Enabled! Attention Enabled! Attention Enabled! Attention Enabled! /home/mssn/DSTA-SLR/model/fstgan.py:791: UserWarning: nn.init.normal is now deprecated in favor of nn.init.normal_. nn.init.normal(self.fc.weight, 0, math.sqrt(2.0 / num_class)) 1442 Traceback (most recent call last): File "main.py", line 801, in <module> processor = Processor(arg) File "main.py", line 238, in __init__ self.load_data() File "main.py", line 270, in load_data dataset=Feeder( File "/home/mssn/DSTA-SLR/feeders/feeder.py", line 70, in __init__ self.load_data() File "/home/mssn/DSTA-SLR/feeders/feeder.py", line 129, in load_data self.data = np.load(self.data_path, mmap_mode="r") File "/home/mssn/anaconda3/envs/DSTA-SLR/lib/python3.8/site-packages/numpy/lib/npyio.py", line 438, in load raise ValueError("Cannot load file containing pickled data " ValueError: Cannot load file containing pickled data when allow_pickle=False
You may change line 129 in ./feeders/feeder.py into self.data = np.load(self.data_path, mmap_mode="r", allow_pickle=True)
to test it.
I changed this line, but the new error came out as the following.
Could you let me know how you get the train_labels.pkl and train_data_joint.npy for the WLASL100? I think it would be good if I could reproduce the generated datasets.
Traceback (most recent call last):
File "/home/mssn/anaconda3/envs/DSTA-SLR/lib/python3.8/site-packages/numpy/lib/npyio.py", line 441, in load
return pickle.load(fid, **pickle_kwargs)
_pickle.UnpicklingError: could not find MARK
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "main.py", line 801, in <module>
processor = Processor(arg)
File "main.py", line 238, in __init__
self.load_data()
File "main.py", line 270, in load_data
dataset=Feeder(
File "/home/mssn/DSTA-SLR/feeders/feeder.py", line 70, in __init__
self.load_data()
File "/home/mssn/DSTA-SLR/feeders/feeder.py", line 130, in load_data
self.data = np.load(self.data_path, mmap_mode="r", allow_pickle=True)
File "/home/mssn/anaconda3/envs/DSTA-SLR/lib/python3.8/site-packages/numpy/lib/npyio.py", line 443, in load
raise pickle.UnpicklingError(
_pickle.UnpicklingError: Failed to interpret file './data/WLASL100/val_data_joint.npy' as a pickle
I get the data for WLASL100 by squeezing the data from WLASL2000 data according to the classes of WLASL100. The occurred error seems strange, because the input data is not a pickle file but a npy file.
I see. Thank you for the reply! I will try to regenerate the WLASL100 files by myself.
Hi author,
It's a great work! I tried to reproduce your result on WLASL dataset. But I found that the trained result is bad (top1 per instance = 1.66%, top5 per instance= 1.77%) using your preprocessed train_label.pkl and train_data_joint.npy of WLASL300.
The train.yaml is as below. Could you please give me a suggestion on how to modify the configuration to reproduce your results?