Shape of the input features

[dangkh]

Hi, I truly like what you did and I'm reproducing your work with another dataset. But, I have problems with the shape of the input features.

At file "gcnet/train_gcnet.py", line 223, you mention the shape after reading in dataloader
audio_host, text_host, visual_host: [seqlen, batch, dim]

However, the shape of the data in the dataloader is usually [batch, seqlen, dim] and without any shape transpose in your code. Whether there are any piece of code missing from your work?

[zeroQiaoba]

Actually, based on pytorch=1.8.0, https://pytorch.org/docs/1.8.0/generated/torch.nn.LSTM.html?highlight=nn%20lstm#torch.nn.LSTM, torch.nn.LSTM takes (seq_len, batch, input_size) as the input.