File size: 3,095 Bytes
4014562
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97

DATA:
    task_name: wiki1m
    dataset_name: wiki1m
    dataset_config_name: null
    max_seq_length: 512
    overwrite_output_dir: false # Set to false, if using one_hour_job
    overwrite_cache: false
    pad_to_max_length: true
    load_from_local: true
    max_train_samples: null
    max_eval_samples: null
    max_predict_samples: null
    train_file: datasets/Wiki1M/train.jsonl

    # validation_file: datasets/Wiki1M/test_unseen.jsonl
    # test_file: datasets/Wiki1M/test_unseen.jsonl

    validation_file: datasets/Wiki1M/test.jsonl
    test_file: datasets/Wiki1M/test.jsonl

    label_max_seq_length: 128 # For names baseline
    descriptions_file: datasets/Wiki1M/wiki_all_final_descs_fixed_dedup.json

    all_labels : datasets/Wiki1M/all_labels.txt
    test_labels: datasets/Wiki1M/unseen_labels.txt

    large_dset: true
    tokenized_descs_file: datasets/Wiki1M/tokenized_ner_descs_final_fixed_dedup128.npy

    train_tfidf_short: datasets/Wiki1M/train_shortlists_4K_500.h5

    # test_tfidf_short: datasets/Wiki1M/test_unseen_shortlists_4K.h5
    # test_tfidf_short: datasets/Wiki1M/test_unseen_shortlists_4K_fixed.h5
    test_tfidf_short: datasets/Wiki1M/test_shortlists_4K_1000_bak.h5

    tok_format: 1

    # max_descs_per_label: 5
    contrastive_learning_samples: 750
    # cl_min_positive_descs: 1
    # bm_short_file: datasets/eurlex4.3k/train_bmshort.txt
    # ignore_pos_labels_file: datasets/Amzn13K/ignore_train_split6500_fs5.txt
    coil_cluster_mapping_path: bert_coil_map_dict_lemma255K_isotropic.json

MODEL:
    model_name_or_path: bert-base-uncased
    # pretrained_model_path: final_wiki_descs_short500/checkpoint-16000/pytorch_model.bin
    config_name: null
    tokenizer_name: null
    cache_dir: null
    use_fast_tokenizer: true
    model_revision: main
    use_auth_token: false
    ignore_mismatched_sizes: false
    negative_sampling: "none"
    semsup: true
    # label_model_name_or_path: bert-base-uncased # prajjwal1/bert-small
    label_model_name_or_path: prajjwal1/bert-small
    encoder_model_type: bert
    use_custom_optimizer: adamw
    output_learning_rate: 1.e-4
    arch_type : 2
    add_label_name: false
    normalize_embeddings: false
    tie_weights: false
    coil: true
    colbert: false
    # use_precomputed_embeddings: datasets/eurlex4.3k/heir_withdescriptions_4.3k_v1_embs_bert_9_128.npy
    token_dim: 16
    label_frozen_layers: 2

TRAINING:
    do_train: true
    do_eval: true
    do_predict: false
    per_device_train_batch_size: 2
    gradient_accumulation_steps: 4
    per_device_eval_batch_size: 2
    learning_rate: 5.e-5 # Will point to input encoder lr, if user_custom_optimizer is False
    num_train_epochs: 3
    save_steps: 5000
    evaluation_strategy: steps
    eval_steps: 5000
    fp16: true
    fp16_opt_level: O1
    lr_scheduler_type: "linear" # defaults to 'linear'
    dataloader_num_workers: 16
    label_names: [labels]
    scenario: "unseen_labels"

    ddp_find_unused_parameters: false
    max_eval_samples: 50000
    ignore_data_skip: true
    # one_hour_job: true
    random_sample_seed: 69