def main(_params):
global params
params = _params
train_df, test_df, dev_df, labels, num_labels, label_map, data_dir = prepare_data(
)
data_args, model_args, config, tokenizer = prepare_config_and_tokenizer(
data_dir, labels, num_labels, label_map)
# ## Create Dataset Objects
train_dataset = NerDataset(
data_dir=data_args['data_dir'],
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args['max_seq_length'],
overwrite_cache=data_args['overwrite_cache'], # True
mode=Split.train,
data_size=params["data_size"])
eval_dataset = NerDataset(data_dir=data_args['data_dir'],
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args['max_seq_length'],
overwrite_cache=data_args['overwrite_cache'],
mode=Split.dev,
data_size=params["data_size"])
print(train_dataset.__len__(), eval_dataset.__len__())
# Train top-model using the Trainer API
trainer, model = run_train(train_dataset, eval_dataset, config, model_args,
labels, num_labels, label_map)
gc.collect()
torch.cuda.empty_cache()
# ## Prepare test data, run trainer over test data and print metrics
# we can pass overwrite_cache as True since we might like to make new predictions by just changing test.txt
test_dataset = NerDataset(data_dir=data_args['data_dir'],
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args['max_seq_length'],
overwrite_cache=True,
mode=Split.test,
data_size=params["data_size"])
run_test(trainer, model, train_dataset, train_df, label_map)
run_test(trainer, model, eval_dataset, dev_df, label_map)
run_test(trainer, model, test_dataset, test_df, label_map)
def main(_params):
global params
params = _params
'''
params['seed_value'] = args.seed_value
params['set_seed'] = args.set_seed
'''
wb_run = wandb.init(project="NER", name=params['exp_name'] + "_init")
if params['set_seed']:
random_seed_set(params['seed_value'])
train_df, test_df, dev_df, labels, num_labels, label_map, data_dir, wt = prepare_data(
)
data_args, model_args, config, tokenizer = prepare_config_and_tokenizer(
data_dir, labels, num_labels, label_map)
if 'add_vocab' in params.keys():
process_entity(tokenizer, train_df)
process_entity(tokenizer, dev_df)
process_entity(tokenizer, test_df)
# ## Create Dataset Objects
xargs = {}
if params.get('xargs'):
xargs = params['xargs']
xargs['wt'] = wt
print('Got class weights')
xargs["top_model"] = params.get("top_model")
train_dataset = NerDataset(
data_dir=data_args['data_dir'],
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args['max_seq_length'],
overwrite_cache=data_args['overwrite_cache'], # True
mode=Split.train,
data_size=params["data_size"],
xargs=xargs)
eval_dataset = NerDataset(data_dir=data_args['data_dir'],
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args['max_seq_length'],
overwrite_cache=data_args['overwrite_cache'],
mode=Split.dev,
data_size=100)
# ## Prepare test data, run trainer over test data and print metrics
# we can pass overwrite_cache as True since we might like to make new predictions by just changing test.txt
test_dataset = NerDataset(data_dir=data_args['data_dir'],
tokenizer=tokenizer,
labels=labels,
model_type=config.model_type,
max_seq_length=data_args['max_seq_length'],
overwrite_cache=True,
mode=Split.test,
data_size=100)
print(train_dataset.__len__(), eval_dataset.__len__(),
test_dataset.__len__())
wb_run.finish()
# Train top-model using the Trainer API
if params.get("hyp"):
run_hyperp(train_dataset, eval_dataset, config, model_args, labels,
num_labels, label_map, tokenizer, xargs)
return
trainer, model = run_train(train_dataset, eval_dataset, config, model_args,
labels, num_labels, label_map, tokenizer, xargs)
gc.collect()
torch.cuda.empty_cache()
wb_run = wandb.init(project="NER", name=params['exp_name'] + "summary")
report = run_test(trainer, model, train_dataset, train_df, label_map)
wandb.run.summary["train_report"] = report
report = run_test(trainer, model, eval_dataset, dev_df, label_map)
wandb.run.summary["val_report"] = report
report = run_test(trainer, model, test_dataset, test_df, label_map)
wandb.run.summary["test_report"] = report
wandb.run.summary["model"] = model.__repr__()
wandb.run.summary["data"] = {
"train": train_dataset.__len__(),
"val": eval_dataset.__len__(),
"test": test_dataset.__len__(),
"wt": wt
}
params["model_type"] = params["model_type"].name
wandb.run.summary["params"] = params
wb_run.finish()
