config.py

# import os
# def get_data_dir_name():
# 	list_subdirs = os.listdir(os.getcwd() + '/data/')
# 	print('before cleaned list_subdirs:',  list_subdirs)
# 	list_subdirs = [i for i in list_subdirs if 'lvl' in i]
# 	assert len(list_subdirs) == 1
# 	return list_subdirs[0]

## training params 
classifier_type = 'twoHidden_3840_3840'

snippet_type = 'lvl_5' # snippet_type: 11 choices 
data_dir = f"data/{snippet_type}"

q_type = ['A1','A2','A3/A4','B1'] # the type of questions to use in this program
assert all(enum in ['A1','A2','A3/A4','B1'] for enum in q_type), "q_type does not exist"

# train params
train_type = 'choice_correct' # [choice_correct, choice_all]
train_file_name = 'test_18' # [train, test_17, test_18]
per_gpu_train_batch_size = 5 # must be a multiple of 5; set it to 5 if running on local machines
num_train_epochs = 3
assert train_type in ['choice_correct', 'choice_all'], "wrong train_type name"
train_file = f'{train_file_name}.tsv' if train_type == 'choice_all' else f'{train_file_name}_choice_correct.tsv'

# eval params
eval_files = ['test_17', 'test_18_choice_correct']
eval_epochs = range(1, num_train_epochs + 1)
per_gpu_eval_size = 5 


use_fine_tuned_model = False # set to True if using fined tuned bert model
use_multi_gpu = False # NOTE: MODIFY pbs file if using multi gpu on hpc
do_visualization = False
do_evaluation = False # deprecated! whether to evaluate models during trianing to see live results
use_multiprocessing = True


# others
bert_model = 'bert-base-chinese'
output_dir = 'outputs/'
report_dir = 'reports/'
cache_dir = 'cache/'
max_steps = -1 # default -1: run all steps
max_seq_length = 512
assert per_gpu_train_batch_size % 5 == 0 and per_gpu_eval_size % 5 == 0
learning_rate = 2e-5
random_seed = 42

gradient_accumulation_steps = 1
warmup_proportion = 0.1
output_mode = 'classification'
# classification
CONFIG_NAME = 'config.json'
WEIGHTS_NAME = 'pytorch_model.bin'
visdom_env_name = snippet_type

# file name has been modified
stats = f'train_type={train_type}, snippet_type={snippet_type}, use_multi_gpu={use_multi_gpu}, visdom_name={visdom_env_name}, use_fine_tuned_model={use_fine_tuned_model},  total_epochs={num_train_epochs}, train_file={train_file}, max_seq_length={max_seq_length}, per_gpu_train_batch_size={per_gpu_train_batch_size}, per_gpu_eval_batch_size={per_gpu_eval_size}'