def train():
"""Renders the train page and performs training if request method is POST."""
global PROGRESS, TRAINING
warnings, errors = [], []
if request.method == 'GET':
return render_train()
# Get arguments
data_name, epochs, ensemble_size, checkpoint_name = \
request.form['dataName'], int(request.form['epochs']), \
int(request.form['ensembleSize']), request.form['checkpointName']
gpu = request.form.get('gpu')
data_path = os.path.join(app.config['DATA_FOLDER'], f'{data_name}.csv')
dataset_type = request.form.get('datasetType', 'regression')
# Create and modify args
args = TrainArgs().parse_args([
'--data_path', data_path, '--dataset_type', dataset_type, '--epochs',
str(epochs), '--ensemble_size',
str(ensemble_size)
])
# Check if regression/classification selection matches data
data = get_data(path=data_path)
targets = data.targets()
unique_targets = {
target
for row in targets for target in row if target is not None
}
if dataset_type == 'classification' and len(unique_targets - {0, 1}) > 0:
errors.append(
'Selected classification dataset but not all labels are 0 or 1. Select regression instead.'
)
return render_train(warnings=warnings, errors=errors)
if dataset_type == 'regression' and unique_targets <= {0, 1}:
errors.append(
'Selected regression dataset but all labels are 0 or 1. Select classification instead.'
)
return render_train(warnings=warnings, errors=errors)
if gpu is not None:
if gpu == 'None':
args.cuda = False
else:
args.gpu = int(gpu)
current_user = request.cookies.get('currentUser')
if not current_user:
# Use DEFAULT as current user if the client's cookie is not set.
current_user = app.config['DEFAULT_USER_ID']
ckpt_id, ckpt_name = db.insert_ckpt(checkpoint_name, current_user,
args.dataset_type, args.epochs,
args.ensemble_size, len(targets))
with TemporaryDirectory() as temp_dir:
args.save_dir = temp_dir
process = mp.Process(target=progress_bar, args=(args, PROGRESS))
process.start()
TRAINING = 1
# Run training
logger = create_logger(name='train',
save_dir=args.save_dir,
quiet=args.quiet)
task_scores = run_training(args, logger)
process.join()
# Reset globals
TRAINING = 0
PROGRESS = mp.Value('d', 0.0)
# Check if name overlap
if checkpoint_name != ckpt_name:
warnings.append(
name_already_exists_message('Checkpoint', checkpoint_name,
ckpt_name))
# Move models
for root, _, files in os.walk(args.save_dir):
for fname in files:
if fname.endswith('.pt'):
model_id = db.insert_model(ckpt_id)
save_path = os.path.join(app.config['CHECKPOINT_FOLDER'],
f'{model_id}.pt')
shutil.move(os.path.join(args.save_dir, root, fname),
save_path)
return render_train(trained=True,
metric=args.metric,
num_tasks=len(args.task_names),
task_names=args.task_names,
task_scores=format_float_list(task_scores),
mean_score=format_float(np.mean(task_scores)),
warnings=warnings,
errors=errors)
"""Trains a model on a dataset."""
from chemprop.parsing import parse_train_args
from chemprop.train import cross_validate, run_training
from chemprop.utils import create_logger
import pickle
import os
if __name__ == '__main__':
args = parse_train_args()
logger = create_logger(name='train',
save_dir=args.save_dir,
quiet=args.quiet)
test_avg_score, test_preds, test_smiles = run_training(args, logger)
with open(os.path.join(args.save_dir, 'test_preds.pickle'), 'wb') as preds:
pickle.dump(test_preds, preds)
with open(os.path.join(args.save_dir, 'test_smiles.pickle'),
'wb') as smiles:
pickle.dump(test_smiles, smiles)