the JGA is only 0.27....

[Janeey99]

Firstly, thanks for the author's good works!
I use your original code and train in this hyperparmeters:
attention_probs_dropout_prob=0.1, batch_size=16, bert_ckpt_path='assets/bert-base-uncased-pytorch_model.bin', bert_config_path='assets/bert_config_base_uncased.json', data_root='data/mwz2.0', dec_lr=0.0001, dec_warmup=0.1, decoder_teacher_forcing=0.5, dev_data='dev_dials.json', dev_data_path='data/mwz2.0/dev_dials.json', dropout=0.1, enc_lr=4e-05, enc_warmup=0.1, eval_epoch=1, exclude_domain=False, hidden_dropout_prob=0.1, max_seq_length=256, msg=None, n_epochs=30, n_history=1, not_shuffle_state=False, num_workers=4, ontology_data='data/mwz2.0/ontology.json', op_code='4', random_seed=42, save_dir='outputs', shuffle_p=0.5, shuffle_state=True, slot_token='[SLOT]', test_data='test_dials.json', test_data_path='data/mwz2.0/test_dials.json', train_data='train_dials.json', train_data_path='data/mwz2.0/train_dials.json', vocab_path='assets/vocab.txt', word_dropout=0.1
actually, I only changed batch_size, and other same as yours.
But, the results were a little off the mark.
op_code: 4, is_gt_op: False, is_gt_p_state: False, is_gt_gen: False
Epoch 20 joint accuracy : 0.2733441910966341
Epoch 20 slot turn accuracy : 0.9547819399203203
Epoch 20 slot turn F1: 0.8490504095046196
Epoch 20 op accuracy : 0.9708785740137087
Epoch 20 op F1 : {'delete': 0.0886426592797784, 'update': 0.7394006048941435, 'dontcare': 0.03618649965205289, 'carryover': 0.984787856819641}
Epoch 20 op hit count : {'delete': 16, 'update': 6723, 'dontcare': 26, 'carryover': 207838}
Epoch 20 op all count : {'delete': 339, 'update': 11265, 'dontcare': 1401, 'carryover': 208035}
Final Joint Accuracy : 0.08108108108108109
Final slot turn F1 : 0.8280395869977584
Latency Per Prediction : 21.975781 ms

I dont know what's wrong with it?