Source code for ICASSP2022 paper: EmotionFlow: Capture the Dialogue Level Emotion Transitions
transformers=4.14.1
torch=1.8
vocab=0.0.5
numpy
tqdm
sklearn
pickle
pandas
download MELD dataset from https://github.com/declare-lab/MELD/ and save to ./MELD
python train.py -tr -wp 0 -bsz 1 -acc_step 8 -lr 1e-4 -ptmlr 1e-5 -dpt 0.3 -bert_path roberta-[base, large] -epochs [20, 5]
python train.py -te -ft -bsz 1 -dpt 0.3 -bert_path roberta-[base, large]
| model | weighted-F1 | Checkpoint |
|---|---|---|
| EmotionFlow-roberta-base | 65.05 | roberta-base-meld.pkl |
| EmotionFlow-roberta-large | 66.50 | roberta-large-meld.pkl |
Checkpoints are produced on a single V100 GPU.