- Pytorch==1.8.1
- Transformers==4.3.3
- Torch-geometric==1.7.0
Edge attributes (extract files and put them into bert_data/DATASET_NAME/):
IEMOCAP
DailyDialog
EmoryNLP
MELD
We additionally provide the utterance representations and edge attributes of MELD to perform the training style like COSMIC.
The data is in here meld. Extract the files and put them in cosmic_data/meld/
Training SKAIG models on datasets can use:
IEMOCAP: python gnn_train.py -index 1 -lr 1e-5 -choice 'cn' -pretrain 'roberta-base' -cn_num_layer 5 -sent_dim 300 -cn_nhead 6 -edge_dim 300 -cn_ff_dim 600 -cn_dropout 0.1 -hip 7 -seed 7 -residual_type 'none' -dataset_name 'IEMOCAP'
DailyDialog: python gnn_train.py -index 1 -lr 1e-5 -choice 'cn' -pretrain 'roberta-large' -cn_num_layer 5 -sent_dim 300 -cn_nhead 6 -edge_dim 300 -cn_ff_dim 600 -cn_dropout 0.1 -hip 2 -seed 7 -residual_type 'none' -dataset_name 'DailyDialog'
EmoryNLP: python gnn_train.py -index 1 -lr 1e-5 -choice 'cn' -pretrain 'roberta-large' -cn_num_layer 5 -sent_dim 300 -cn_nhead 6 -edge_dim 300 -cn_ff_dim 600 -cn_dropout 0.1 -hip 4 -seed 7 -residual_type 'none' -dataset_name 'EmoryNLP'
MELD: python gnn_train.py -index 1 -lr 8e-6 -choice 'cn' -pretrain 'roberta-large' -cn_num_layer 2 -sent_dim 200 -cn_nhead 4 -edge_dim 200 -cn_ff_dim 200 -cn_dropout 0.1 -hip 2 -seed 7 -residual_type 'none' -dataset_name 'MELD'
For COSMIC style training, the command is:
python cosmic_style_train.py -index 1 --edge_mapping --shuffle --beta --root_weight --use_gpu -dataset_name 'meld' -batch_size 8 -sent_dim 200 -ff_dim 200 -heads 4 -dropout 0.5 -num_layers 2 -edge_dim 200 -hop 3 -seed 7 -lr 1e-4 -l2 3e-4
The model is inserted into MentalModel, which can be activated by -choice 'tr' in gnn_train.py. Additionally, the hyper-params for it is started with 'tr_'.
The commands for training TRMSM: Pending.