SemSup-XC / cleaned_code /configs /final_wiki_descs_500random750.yml
Pranjal2041's picture
Initial Commit
4014562
raw
history blame
3.1 kB
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