About the data set. about pytorch-multimodal_sarcasm_detection HOT 5 CLOSED

I don't think using the original split will make the model unable to train. You need to check your code.

from pytorch-multimodal_sarcasm_detection.

I found it also, and I changed to the original split, it run success, but F1 and acc decreased by 1.5% compare to original paper， is it some difference between tensorflow/pytorch

from pytorch-multimodal_sarcasm_detection.

I found it also, and I changed to the original split, it run success, but F1 and acc decreased by 1.5% compare to original paper， is it some difference between tensorflow/pytorch

Yes, it because I didn't do exactly the same as the original code:

1. I initialized all weight by zero, instead of truncated normal distribution.
2. In the LSTM part, the author did the dropout on hidden state, but I did it on the cell itself. It because PyTorch didn't support that yet, you may check here.

from pytorch-multimodal_sarcasm_detection.

I found it also, and I changed to the original split, it run success, but F1 and acc decreased by 1.5% compare to original paper， is it some difference between tensorflow/pytorch

Yes, it because I didn't do exactly the same as the original code:

1. I initialized all weight by zero, instead of truncated normal distribution.
2. In the LSTM part, the author did the dropout on hidden state, but I did it on the cell itself. It because PyTorch didn't support that yet, you may check here.

Thanks for your kind answer, I fix my code and it worked. By the way, I would like to know which part has been initialized by zero.

from pytorch-multimodal_sarcasm_detection.

I found it also, and I changed to the original split, it run success, but F1 and acc decreased by 1.5% compare to original paper， is it some difference between tensorflow/pytorch

Yes, it because I didn't do exactly the same as the original code:

1. I initialized all weight by zero, instead of truncated normal distribution.
2. In the LSTM part, the author did the dropout on hidden state, but I did it on the cell itself. It because PyTorch didn't support that yet, you may check here.

Thanks for your kind answer, I fix my code and it worked. By the way, I would like to know which part has been initialized by zero.

I may be incorrect, they might not be initialized by zero, but what I want to say is that I didn't do any of the work on the initialize part(except the "early fusion" which described in the paper). So all of the layers are initialized by PyTorch's default setting, which may differ from the original code.

from pytorch-multimodal_sarcasm_detection.