Hi, your work is very interesting! But I met some question when I followed the readme's file to reproduct the project. I run the train_eae.sh file and the file's CONFIG was replaced to config_eae_ace05e_AMRBart+prefixgen+copy_enccrossprefix_tp40.json, and I just change the json's pretrainModel dir path. But I run this project, the result is lower that your paper's result.
The recent results section log information is as follows：
[2023-06-27 15:41:31] - __main__ - Epoch 45
Train 45: 100%|████████████████████████| 1051/1051 [05:07<00:00, 3.42it/s]
[2023-06-27 15:46:38] - __main__ - Average training loss : 0.039728544652462006...
Dev 45: 100%|██████████████████████████████| 38/38 [00:26<00:00, 1.45it/s]
[2023-06-27 15:47:05] - __main__ - --------------------------Dev Scores---------------------------------
[2023-06-27 15:47:05] - __main__ - Role I - P: 70.79 ( 429/ 606), R: 54.79 ( 429/ 783), F: 61.77
[2023-06-27 15:47:05] - __main__ - Role C - P: 66.01 ( 400/ 606), R: 51.09 ( 400/ 783), F: 57.60
[2023-06-27 15:47:05] - __main__ - ---------------------------------------------------------------------
[2023-06-27 15:47:05] - __main__ - {'epoch': 45, 'dev_scores': {'arg_id': (0.7079207920792079, 0.5478927203065134, 0.6177105831533477), 'arg_cls': (0.6600660066006601, 0.5108556832694764, 0.5759539236861051)}}
[2023-06-27 15:47:05] - __main__ - Current best
[2023-06-27 15:47:05] - __main__ - {'best_epoch': 36, 'best_scores': {'arg_id': (0.7207357859531772, 0.5504469987228607, 0.6241853729181752), 'arg_cls': (0.6822742474916388, 0.5210727969348659, 0.5908761766835626)}}
[2023-06-27 15:47:05] - __main__ - ./output/eae_ace05e_AMRBart+prefixgen+copy_enccrossprefix_tp40/20230627_101723/train.log
[2023-06-27 15:47:05] - __main__ - Done!
{
"dataset": "ace05e",
"gpu_device": 0,
"seed": 100,
"train_file": "./processed_data/ace05e_bart/train.w1.oneie.json",
"dev_file": "./processed_data/ace05e_bart/dev.w1.oneie.json",
"test_file": "./processed_data/ace05e_bart/test.w1.oneie.json",
"finetune_dir": "./data/eae_ace05e/",
"train_finetune_file": "./data/eae_ace05e/train_all.pkl",
"dev_finetune_file": "./data/eae_ace05e/dev_all.pkl",
"test_finetune_file": "./data/eae_ace05e/test_all.pkl",
"vocab_file": "./data/eae_ace05e/vocab.json",
"model_type": "AMR+prefixgen+copy",
"output_dir": "./output/eae_ace05e_AMRBart+prefixgen+copy_enccrossprefix_tp40",
"cache_dir": "./cache",
"model_name": "/media/ubuntu/projects/pretrainModel/bart-large",
"input_style": ["event_type_sent", "triggers", "template", "na_token"],
"output_style": ["argument:sentence"],
"max_epoch": 45,
"warmup_epoch": 5,
"train_batch_size": 4,
"eval_batch_size": 12,
"accumulate_step": 1,
"learning_rate": 1e-05,
"weight_decay": 1e-05,
"grad_clipping": 5.0,
"beam_size": 1,
"max_length": 250,
"max_output_length": 50,
"ignore_first_header": true,
"use_encoder_prefix": true,
"use_cross_prefix": true,
"use_decoder_prefix": false,
"AMR_model_path": "/media/ubuntu/projects/pretrainModel/AMRBART-large-finetuned-AMR3.0-AMR2Text-v2",
"latent_dim": 1024,
"prefix_length": 40,
"freeze_AMR": false,
"freeze_bart": false,
"freeze_prefixprojector": false,
"pretrained_model_path": null
}
Hi, your work is very interesting! But I met some question when I followed the readme's file to reproduct the project. I run the
train_eae.shfile and the file'sCONFIGwas replaced toconfig_eae_ace05e_AMRBart+prefixgen+copy_enccrossprefix_tp40.json, and I just change the json's pretrainModel dir path. But I run this project, the result is lower that your paper's result.The recent results section log information is as follows：
config_eae_ace05e_AMRBart+prefixgen+copy_enccrossprefix_tp40.jsoninformation is as follows: