search

Turdidae810 / HSCL

Code of Hierarchical Supervised Contrastive Learning for Multimodal Sentiment Analysis
MIT License
1 stars 0 forks source link

reproduce the results #1

Open 1060807523 opened 3 months ago

1060807523 commented 3 months ago

Thanks for your great work! However, I can't reproduce the performance that was reported in the paper. Would you like to provide more details about the code?

Turdidae810 commented 3 months ago

I've updated the README.md with details about the code. The experimental results in the paper were obtained on RTX3090. If you have any questions feel free to contact me.

ZJU-PLP commented 3 days ago

@Turdidae810 I train the code in the MOSI dataset by "python main.py --dataset mosi --data_path ../MOSI --bert_path ./bert-base-uncased --batch_size 64 --num_epochs 200 --patience 50". However, I can only get the F1 score all/non0: 0.8257/0.8443 over 686/656 while can not achieve 86.31% in your paper. Selection_735

Turdidae810 commented 20 hours ago

@ZJU-PLP Different environments, hyperparameters, and hardware may produce different results. On my machine (RTX 3090), using the hyperparameters specified in the paper (the default hyperparameters in this repository) and the environment in requirements.txt, I can reproduce the results in the paper. image

ZJU-PLP commented 56 minutes ago

@ZJU-PLP Different environments, hyperparameters, and hardware may produce different results. On my machine (RTX 3090), using the hyperparameters specified in the paper (the default hyperparameters in this repository) and the environment in requirements.txt, I can reproduce the results in the paper. image

Thanks for your nice reply. However, I have reproduced the experimental results in Nvidia GTX 1080Ti and RTX while still could not get your results. Maybe it really the reason that I should use the RTX 3090 to train?