About UR_FUNNY dataset

[yangshuang]

Hi, When I use the UR_FUNNY dataset, the error is as follows:

Traceback (most recent call last):
File "main.py", line 166, in
train_loader = get_loader(hyp_params, train_config, shuffle=True)
File "/content/drive/MyDrive/BBFN-main/src/data_loader.py", line 62, in get_loader
dataset = MSADataset(config)
File "/content/drive/MyDrive/BBFN-main/src/data_loader.py", line 27, in init
dataset = UR_FUNNY(config)
File "/content/drive/MyDrive/BBFN-main/src/create_dataset.py", line 481, in init
words.append(word2id[np.array(word)])
TypeError: unhashable type: 'numpy.ndarray'

Could you give me some advice? Thank you very much!

[yangshuang]

I found that it seems to be the dataset problem. Have you ever encountered a similar problem?

[Wei Han (Henry)]

No I haven't. I found your code segment is different from the main branch. You can clone the main branch and try again.

[yangshuang]

Thanks a lot! I ran it again, but it still reported two errors:

Traceback (most recent call last):
File "/content/drive/MyDrive/BBFN-main/src/create_dataset.py", line 415, in init
self.train = load_pickle(DATA_PATH + '/train.pkl')
File "/content/drive/MyDrive/BBFN-main/src/create_dataset.py", line 21, in load_pickle
return pickle.load(f)
EOFError: Ran out of input

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "main.py", line 166, in
train_loader = get_loader(hyp_params, train_config, shuffle=True)
File "/content/drive/MyDrive/BBFN-main/src/data_loader.py", line 61, in get_loader
dataset = MSADataset(config)
File "/content/drive/MyDrive/BBFN-main/src/data_loader.py", line 27, in init
dataset = UR_FUNNY(config)
File "/content/drive/MyDrive/BBFN-main/src/create_dataset.py", line 478, in init
words.append(word2id[word])
TypeError: unhashable type: 'numpy.ndarray'

I can't solve this problem at the moment😫, could you share the preprocessed data of the UR_FUNNY?

[Wei Han (Henry)]

Could you please check what are the contents in the variable 'word'?

[yangshuang]

Thank you very much! I tried to modify it, and now it can be trained, but the log showed as follows:

Found video that doesn't bolong to any splits: XXX

And the performance isn't very well, the Accuracy is around 60.

[Wei Han (Henry)]

You need to add 'use_bert' option in command line to enable the BERT encoder, then you can tune the hyperparameters for better performance.