tongxiao2002
commented
1 week ago Maybe it's caused by the physical environments? To mitigate these issues, we further release the checkpoints in the data link (under "checkpoints" folder). Hope this would help you.
My experiment: acc: 0.6352604166666667 loss: 1.1794924835364025
My config.json: { "dataset_reader": { "type": "s2s_manual_reader", "source_token_indexer": { "tokens": { "type": "pretrained_transformer", "do_lowercase": false, "model_name": "./chinese-roberta-wwm-ext" } }, "target_token_indexer": { "tokens": { "type": "single_id" } }, "tokenizer": { "word_splitter": { "type": "just_spaces" } } }, "iterator": { "type": "basic", "batch_size": 64 }, "model": { "type": "geo_s2s", "beam_size": 10, "encoder": { "dropout": 0.5, "emb_dim": 768, "hid_dim": 512, "input_dim": 21128 }, "knowledge_explanation_file": "./GeoQA-Data/GeoQA-Pro/geometry-knowledges-chinese.json", "knowledge_points_ratio": 0, "max_decoding_steps": 20, "scheduled_sampling_ratio": 0, "source_embedder": { "token_embedders": {} }, "target_embedding_dim": 512 }, "train_data_path": "./GeoQA-Data/GeoQA-Pro/pro_train.pk", "validation_data_path": "./GeoQA-Data/GeoQA-Pro/pro_dev.pk", "test_data_path": "./GeoQA-Data/GeoQA-Pro/pro_test.pk", "trainer": { "cuda_device": 0, "grad_norm": 10, "learning_rate_scheduler": { "type": "slanted_triangular", "cut_frac": 0.1, "num_epochs": 100, "num_steps_per_epoch": 108 }, "num_epochs": 100, "num_serialized_models_to_keep": 2, "optimizer": { "type": "adam", "lr": 0.001, "parameter_groups": [ [ [ "mcan", "merge_att", "channel_transform", "attflat_img", "attflat_lang", "decode_transform", "inference_transform", "goal_generation_module" ], { "lr": 1e-05 } ], [ [ "resnet" ], { "lr": 1e-05 } ], [ [ "source_embedder", "encoder.embedding" ], { "lr": 2e-05 } ], [ [ "encoder.concat_trans", "encoder.concattrans", "encoder.lstm_embedding", "encoder.trans", "encoder.norm", "encoder.concat_norm" ], { "lr": 0.001 } ] ] }, "validation_metric": "+acc" }, "evaluate_on_test": true }