def get_arguments():
# Initialize the parser
parser = argparse.ArgumentParser(
description='Input arguments for which dataset to select')
## Program Arguments
parser.add_argument('--seed',
type=int,
help='Seed value. Default=0',
default=0)
parser.add_argument('--task_name',
choices=['qnli', 'sst2', 'mrpc'],
help='Which dataset to select. Default=qnli',
default='sst2')
parser.add_argument('--model',
choices=['bart', 'xlnet'],
help='Which model to select. Default=bart',
default='xlnet')
parser.add_argument(
'--out_dir',
help='Output dir where to save the dataset. Default=./out_dir',
default='./out_dir')
## Parse the arguments
args = parser.parse_args()
args.out_dir = eutils.mkdir_p(args.out_dir)
return args
def compute_dev_s_scores():
## TODO: Write the values to the dataframe dynamically instead of at the very end
key_name = 'dev_s'
model_names, task_names, ngrams_list, seeds_list = get_settings_list()
args = get_arguments()
args.inp_dir = eutils.mkdir_p(f'./out_dir/{key_name}')
avg_scores = []
for outer_idx, (task, model, ngram) in enumerate(itertools.product(task_names, model_names, ngrams_list)):
print(f'\nKey_name: {key_name}, Model: {model}, task: {task}, ngram: {ngram}\n')
args.task_name = task
args.model = model
args.ngram = ngram
tmp_res = []
for inner_idx, seed in enumerate(seeds_list):
args.seed = seed ## Set the seed
tmp_res.append(get_model_scores(args))
dict_of_list = merge_with(list, *tmp_res)
dict_of_avg_val = {key: statistics.mean(val) for key, val in dict_of_list.items()}
avg_scores.append({'model': model, 'task_name': task, 'ngram': ngram, 'word_order_sensitivity':word_order_sensitivity(dict_of_avg_val['accuracy']),
**dict_of_avg_val})
df = pd.DataFrame(avg_scores)
file_name = args.out_dir / f'{key_name}_scores.csv'
df.to_csv(file_name, index=False)
def main(args):
## Debugging Mode
if sys.gettrace() is not None:
print('In debug mode')
args.profiler = 'simple'
args.num_workers = 0
args.pin_memory = False
args.default_root_dir = mkdir_p(abs_path('./tmp_dir'))
args.stochastic_weight_avg = True
args.limit_train_batches = 0.001
args.limit_val_batches = 0.001
args.num_sanity_val_steps = 0
args.terminate_on_nan = True
args.deterministic = True
args.auto_select_gpus = False
args.fast_dev_run = False # Quick Check Working
args.progress_bar_refresh_rate = 0
args.gpus = 1
args.precision = 16
args.train_batch_size = 32
# args.val_batch_size=4
args.freeze_encoder = True
args.verbose = True
args.max_epochs = 2
## Model
dict_vars = vars(args)
model = BertForRegression(**dict_vars)
print(f'Model Init Done')
## TODO: Early Stopping Callback
## Callbacks
checkpoint_callback = pl.callbacks.ModelCheckpoint(
filename='{epoch:02d}-{val_loss:.5f}-{step_count}',
dirpath=None,
prefix="best_model_checkpoint",
monitor="val_loss",
mode="min",
save_top_k=args.save_top_k,
)
# lr_monitor = pl.callbacks.LearningRateMonitor(logging_interval='step')
trainer = pl.Trainer.from_argparse_args(
args,
callbacks=[LoggingCallback(), checkpoint_callback],
)
trainer.fit(model)
best_model_path = trainer.checkpoint_callback.best_model_path
print(f'\nBEST MODEL PATH IS {best_model_path}')
def compute_dev_r_scores():
key_name = 'dev_r'
model_names, task_names, ngrams_list, seeds_list = get_settings_list()
args = get_arguments()
args.inp_dir = eutils.mkdir_p(f'./out_dir/{key_name}')
avg_scores = []
for outer_idx, (task, model) in enumerate(itertools.product(task_names, model_names)):
print(f'\nKey_name: {key_name}, Model: {model}, task: {task}\n')
args.task_name = task
args.model = model
tmp_res = get_model_scores(args)
avg_scores.append({'model': model, 'task_name': task, **tmp_res})
df = pd.DataFrame(avg_scores)
file_name = args.out_dir / f'{key_name}_scores.csv'
df.to_csv(file_name, index=False)
def save_dataset(dataset, args):
## Save as CSV file (dev-r)
out_dir = eutils.mkdir_p(args.out_dir.joinpath('dev_r'))
file_name = os.path.join(
out_dir, f'model_{args.model}_task_name_{args.task_name}.csv')
dataset.to_csv(path_or_buf=file_name, index=False)
def load_dataset_from_csv_with_key(args, data_key):
if data_key == 'dev_r':
return load_dataset_from_csv(args)
elif data_key == 'dev_s':
args_copy = copy.deepcopy(args)
args_copy.inp_dir = args_copy.inp_dir / f'seed_{args.seed}' / f'ngram_{args.ngram}'
return load_dataset_from_csv(args_copy)
else:
print(f'Not yet implemented')
sys.exit(0)
def get_model_scores(args):
dataset = load_dataset_from_csv_with_key(args, args.inp_dir.stem)
eval_results, model_tuple = eutils.get_model_scores(dataset, args)
return {
'accuracy': eval_results['eval_accuracy'],
'avg_confidence_score': eval_results['eval_avg_confidence_score']
}
if __name__ == '__main__':
args = get_arguments()
if sys.gettrace() is not None:
args.inp_dir = eutils.mkdir_p('./out_dir/dev_s')
args.out_dir = args.inp_dir
aa = get_model_scores(args)
print(f'Done')
def save_dataset(dataset, args):
## Save as CSV file (dev-s)
out_dir = eutils.mkdir_p(args.out_dir.joinpath(f'dev_s/seed_{args.seed}/ngram_{args.ngram}/'))
file_name = os.path.join(out_dir, f'model_{args.model}_task_name_{args.task_name}_seed_{args.seed}_ngram_{args.ngram}.csv')
dataset.to_csv(path_or_buf=file_name, index=False)
