def main():
utils.seed_python(config.seed)
utils.seed_torch(config.seed)
train_eval_data = pd.read_csv(os.path.join(args.dataset_path, 'train.csv'),
converters={'EncodedPixels': str})
train_eval_data['root'] = os.path.join(args.dataset_path, 'train_images')
test_data = pd.read_csv(os.path.join(args.dataset_path,
'sample_submission.csv'),
converters={'EncodedPixels': str})
test_data['root'] = os.path.join(args.dataset_path, 'test_images')
if args.lr_search:
lr = lr_search(train_eval_data)
print('lr_search: {}'.format(lr))
gc.collect()
return
if args.fold is None:
folds = FOLDS
else:
folds = [args.fold]
if not args.infer:
for fold in folds:
train_fold(fold, train_eval_data)
update_transforms(1.) # FIXME:
temp = find_temp_for_folds(folds, train_eval_data)
gc.collect()
build_submission(folds, test_data, temp)
def main():
utils.seed_python(config.seed)
utils.seed_torch(config.seed)
train_eval_data = pd.read_csv(os.path.join(args.dataset_path, 'train.csv'))
train_eval_data['root'] = os.path.join(args.dataset_path, 'train')
test_data = pd.read_csv(os.path.join(args.dataset_path, 'test.csv'))
test_data['root'] = os.path.join(args.dataset_path, 'test')
infer_image_transform.reset(tta=False)
if args.lr_search:
lr = lr_search(train_eval_data)
print('lr_search: {}'.format(lr))
gc.collect()
return
if args.fold is None:
folds = FOLDS
else:
folds = [args.fold]
if not args.infer:
for fold in folds:
train_fold(fold, train_eval_data, test_data)
infer_image_transform.reset(tta=True)
update_transforms(1.) # FIXME:
temp = find_temp_for_folds(folds, train_eval_data)
gc.collect()
build_submission(folds, test_data, temp)
def main(model_paths, dataset_path, submission_path):
utils.seed_python(config.seed)
utils.seed_torch(config.seed)
test_data = load_test_data(dataset_path, 'test')
predictions, ids = build_submission(model_paths, FOLDS, test_data)
predictions = predictions.cpu()
submission = {
'fname': ids,
**{ID_TO_CLASS[i]: predictions[:, i] for i in range(NUM_CLASSES)}
}
submission = pd.DataFrame(submission)
submission.to_csv(submission_path, index=False)
def main():
utils.seed_python(config.seed)
utils.seed_torch(config.seed)
train_eval_data = load_train_eval_data(args.dataset_path, 'train_curated')
train_noisy_data = load_train_eval_data(args.dataset_path, 'train_noisy')
test_data = load_test_data(args.dataset_path, 'test')
noisy_meta = pd.read_csv('./noisy_meta.csv')
noisy_indices = np.argsort(-noisy_meta.score)[:config.noisy_topk]
for data in [train_eval_data, train_noisy_data, test_data]:
if args.debug:
data['path'] = './frees/sample.wav'
if config.opt.lr is None:
lr = find_lr(train_eval_data, train_noisy_data.iloc[noisy_indices])
gc.collect()
else:
lr = config.opt.lr
if args.fold is None:
folds = FOLDS
else:
folds = [args.fold]
for fold in folds:
train_fold(fold, train_eval_data, train_noisy_data.iloc[noisy_indices],
lr)
# TODO: check and refine
# TODO: remove?
evaluate_folds(folds, train_eval_data)
scores, ids = evaluate_noisy(folds, train_noisy_data)
noisy_meta = pd.DataFrame({'fname': ids, 'score': scores})
noisy_meta.to_csv(os.path.join(args.experiment_path, 'noisy_meta.csv'),
index=False)
predictions, ids = build_submission(folds, test_data)
predictions = predictions.cpu()
submission = {
'fname': ids,
**{ID_TO_CLASS[i]: predictions[:, i]
for i in range(NUM_CLASSES)}
}
submission = pd.DataFrame(submission)
submission.to_csv(os.path.join(args.experiment_path, 'submission.csv'),
index=False)
def main():
utils.seed_python(config.seed)
utils.seed_torch(config.seed)
train_eval_data = pd.read_csv(os.path.join(args.dataset_path, 'train.csv'))
train_eval_data['root'] = os.path.join(args.dataset_path, 'train')
test_data = pd.read_csv(os.path.join(args.dataset_path, 'test.csv'))
test_data['root'] = os.path.join(args.dataset_path, 'test')
data = pd.concat([train_eval_data, test_data])
if args.fold is None:
folds = FOLDS
else:
folds = [args.fold]
infer_image_transform.reset(tta=True)
update_transforms(1.) # FIXME:
compute_features(folds, data)
def main():
# TODO: refactor seed
utils.seed_python(config.seed)
if config.opt.lr is None:
lr = find_lr()
print('find_lr: {}'.format(lr))
gc.collect()
else:
lr = config.opt.lr
if args.fold is None:
folds = FOLDS
else:
folds = [args.fold]
for fold in folds:
train_fold(fold, lr)
# TODO: check and refine
threshold = find_threshold_for_folds(folds)
build_submission(folds, threshold)
def worker_init_fn(_):
utils.seed_python(torch.initial_seed() % 2**32)
def main():
utils.seed_python(config.seed)
utils.seed_torch(config.seed)
train()
