def main(logger, args):
df_train, _ = load_data(INPUT_DIR, logger)
logger.info('Preprocess text')
if args['debug']:
df_train = df_train.iloc[:200000]
else:
df_train = preprocess_text(df_train)
seq_train, tokenizer = tokenize_text(df_train, logger)
logger.info('Pad train text data')
seq_train = pad_sequences(seq_train, maxlen=PADDING_LENGTH)
label_train = df_train['target'].values.reshape(-1, 1)
logger.info('Load multiple embeddings')
if args['debug']:
embedding_matrix = np.random.rand(len(tokenizer.word_index) + 1, 300)
else:
embedding_matrix = load_multiple_embeddings(
tokenizer.word_index,
embed_types=[0, 1, 2],
max_workers=args['max_workers'])
embedding_matrix = np.array(embedding_matrix).mean(axis=0)
# ===== training and evaluation loop ===== #
device_ids = args['device_ids']
output_device = device_ids[0]
torch.cuda.set_device(device_ids[0])
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.deterministic = True
batch_size = args['batch_size'] * len(device_ids)
epochs = EPOCHS
logger.info('Start training and evaluation loop')
model_specs = []
for sampler_type in ['over']:
for over_sample_factor in [[5, 2540], [10, 3170]]:
model_specs.append({
'sampler_type': sampler_type,
'over_sample_factor': over_sample_factor[0],
'steps_per_epoch': over_sample_factor[1]
})
model_name_base = 'StackedRNNFM'
for spec_id, spec in enumerate(model_specs):
model_name = model_name_base + f'_specId={spec_id}_factor={spec["over_sample_factor"]}'
skf = StratifiedKFold(n_splits=KFOLD, shuffle=True, random_state=SEED)
oof_preds_optimized = np.zeros(seq_train.shape[0])
oof_preds_majority = np.zeros(seq_train.shape[0])
results = []
for fold, (index_train, index_valid) in enumerate(
skf.split(label_train, label_train)):
logger.info(
f'Fold {fold + 1} / {KFOLD} - create dataloader and build model'
)
x_train, x_valid = seq_train[index_train].astype(
int), seq_train[index_valid].astype(int)
y_train, y_valid = label_train[index_train].astype(
np.float32), label_train[index_valid].astype(np.float32)
model = StackedRNNFM(embedding_matrix,
PADDING_LENGTH,
hidden_size=64,
out_hidden_dim=64,
embed_drop=0.2,
out_drop=0.3)
config = {
'epochs': epochs,
'batch_size': batch_size,
'output_device': output_device,
'optimizer': 'adam',
'optimizer_lr': 0.003,
'num_snapshots': NUM_SNAPSHOTS,
'scheduler_type': 'cyclic',
'base_lr': 0.00001,
'max_lr': 0.003,
'step_size': int(spec['steps_per_epoch'] * 3 / NUM_SNAPSHOTS),
'scheduler_mode': 'triangular',
'scheduler_gamma': 0.9,
'scheduler_trigger_steps': int(spec['steps_per_epoch'] * 2),
'sampler_type': spec['sampler_type'],
'over_sample_factor': spec['over_sample_factor'],
'seed': SEED
}
trainer = Trainer(model, logger, config)
eval_results = trainer.train_and_eval_fold(x_train, y_train,
x_valid, y_valid, fold)
oof_preds_majority[index_valid] = np.array(
[res['preds_binary'] for res in eval_results]).mean(0) > 0.5
oof_majority_f1 = f1_score(
label_train.reshape(-1, )[index_valid],
oof_preds_majority[index_valid])
oof_preds_proba = np.array(
[res['preds_proba'] for res in eval_results]).mean(0)
oof_threshold_mean: float = np.mean(
[res['best_threshold'] for res in eval_results])
oof_preds_optimized[
index_valid] = oof_preds_proba > oof_threshold_mean
oof_optimized_f1 = f1_score(
label_train.reshape(-1, )[index_valid],
oof_preds_optimized[index_valid])
message = f'Fold {fold + 1} / {KFOLD} has been done.\n'
message += f'Score: majority voting - {oof_majority_f1:.6f}, optimized threshold - {oof_optimized_f1:.6f}'
logger.post(message)
post_to_snapshot_spreadsheet(
logger,
SPREADSHEET_SNAPSHOT_URL,
eval_type='SNAPSHOT',
tag='SCORE',
script_name=SCRIPT_NAME,
model_name=model_name,
fold=fold,
snapshot_info=[res['f1'] for res in eval_results])
post_to_snapshot_spreadsheet(
logger,
SPREADSHEET_SNAPSHOT_URL,
eval_type='SNAPSHOT',
tag='THRESHOLD',
script_name=SCRIPT_NAME,
model_name=model_name,
fold=fold,
snapshot_info=[res['best_threshold'] for res in eval_results])
post_to_main_spreadsheet(logger,
SPREADSHEET_MAIN_URL,
eval_type='SNAPSHOT',
script_name=SCRIPT_NAME,
model_name=model_name,
fold=fold,
f1_majority=oof_majority_f1,
f1_optimized=oof_optimized_f1,
threshold=oof_threshold_mean)
results.append({
'f1_majority': oof_majority_f1,
'f1_optimized': oof_optimized_f1,
'threshold': oof_threshold_mean
})
f1_majority_mean = np.mean([res['f1_majority'] for res in results])
f1_majority_std = np.std([res['f1_majority'] for res in results])
f1_optimized_mean = np.mean([res['f1_optimized'] for res in results])
f1_optimized_std = np.std([res['f1_optimized'] for res in results])
threshold_mean = np.mean([res['threshold'] for res in results])
total_metrics = [
f1_majority_mean, f1_majority_std, f1_optimized_mean,
f1_optimized_std, threshold_mean
]
post_to_main_spreadsheet(logger,
SPREADSHEET_MAIN_URL,
eval_type='SNAPSHOT',
script_name=SCRIPT_NAME,
model_name=model_name,
fold=-1,
f1_majority=-1,
f1_optimized=-1,
threshold=-1,
others=total_metrics)
message = 'KFold training and evaluation has been done.\n'
message += f'F1 majority voting - Avg: {f1_majority_mean}, Std: {f1_majority_std}\n'
message += f'F1 optimized - Avg: {f1_optimized_mean}, Std: {f1_optimized_std}\n'
message += f'Threshold - Avg: {threshold_mean}'
logger.post(message)
def train(embedding_matrix):
skf = StratifiedKFold(n_splits=KFOLD, shuffle=True, random_state=SEED)
oof_preds = np.zeros(seq_train.shape[0])
results = []
for fold, (index_train, index_valid) in enumerate(
skf.split(label_train, label_train)):
logger.info(
f'Fold {fold + 1} / {KFOLD} - create dataloader and build model'
)
x_train, x_valid = seq_train[index_train].astype(
int), seq_train[index_valid].astype(int)
y_train, y_valid = label_train[index_train].astype(
np.float32), label_train[index_valid].astype(np.float32)
dataset_train = SimpleDataset(x_train, y_train)
dataset_valid = SimpleDataset(x_valid, y_valid)
dataloader_train = DataLoader(dataset=dataset_train,
batch_size=batch_size,
shuffle=True,
pin_memory=True)
dataloader_valid = DataLoader(dataset=dataset_valid,
batch_size=batch_size,
shuffle=False,
pin_memory=True)
dataloaders = {
'train': dataloader_train,
'valid': dataloader_valid
}
model = StackedRNNFM(embedding_matrix,
PADDING_LENGTH,
hidden_size=64)
model.to(output_device)
criteria = [[nn.BCEWithLogitsLoss(reduction='mean')], [1.0]]
metric = f1_from_logits
optimizer = optim.Adam(model.parameters(), lr=0.001)
scheduler = None
model_save_path = str(
DATA_DIR.joinpath(
f'models/{Path(__file__).stem}_fold_{fold}.model'))
model_name = model._get_name()
config = {
'epochs': epochs,
'loss_names': ['BCE Loss'],
'metric_type': 'batch',
'model_save_path': model_save_path,
'output_device': output_device,
'mode': 'max',
'early_stopping': 200,
'model_name': model_name,
'reg_lambda': None,
'fold': fold
}
model, valid_score, best_epoch = train_model(
model, criteria, metric, optimizer, scheduler, dataloaders,
logger, config)
results.append({
'fold': fold,
'best_score': valid_score,
'best_epoch': best_epoch
})
message = f'Training and evaluation for the fold {fold + 1} / {KFOLD} has been done.\n'
message += f'Validation F1 score: {valid_score}\n'
logger.post(message)
dataloader_valid = DataLoader(dataset=dataset_valid,
batch_size=batch_size,
shuffle=False,
pin_memory=True)
oof_preds[index_valid] = sp.special.expit(
predict(model, dataloader_valid, config).reshape(-1, ))
logger.post(f'K-Fold train and evaluation results: {results}')
return oof_preds
def main(logger, args):
df_train, _ = load_data(INPUT_DIR, logger)
logger.info('Preprocess text')
df_train = preprocess_text(df_train)
seq_train, tokenizer = tokenize_text(df_train, logger)
logger.info('Pad train text data')
seq_train = pad_sequences(seq_train, maxlen=PADDING_LENGTH)
label_train = df_train['target'].values.reshape(-1, 1)
logger.info('Load multiple embeddings')
embedding_matrix = load_multiple_embeddings(
tokenizer.word_index,
embed_types=[0, 2],
max_workers=args['max_workers'])
embedding_matrix = np.array(embedding_matrix).mean(axis=0)
# ===== training and evaluation loop ===== #
device_ids = args['device_ids']
output_device = device_ids[0]
torch.cuda.set_device(device_ids[0])
set_seed(SEED)
batch_size = args['batch_size'] * len(device_ids)
max_workers = args['max_workers']
if args['debug']:
epochs = 2
else:
epochs = EPOCHS
logger.info('Start training and evaluation loop')
skf = StratifiedKFold(n_splits=KFOLD, shuffle=True, random_state=SEED)
oof_preds = np.zeros(seq_train.shape[0])
results = []
for fold, (index_train,
index_valid) in enumerate(skf.split(label_train, label_train)):
logger.info(
f'Fold {fold + 1} / {KFOLD} - create dataloader and build model')
x_train, x_valid = seq_train[index_train].astype(
int), seq_train[index_valid].astype(int)
y_train, y_valid = label_train[index_train].astype(
np.float32), label_train[index_valid].astype(np.float32)
dataset_train = SimpleDataset(x_train, y_train)
dataset_valid = SimpleDataset(x_valid, y_valid)
sampler = BinaryOverSampler(y_train,
over_sample_factor=2,
shuffle=True)
dataloader_train = DataLoader(dataset=dataset_train,
sampler=sampler,
batch_size=batch_size,
pin_memory=True)
dataloader_valid = DataLoader(dataset=dataset_valid,
batch_size=batch_size,
shuffle=False,
pin_memory=True)
dataloaders = {'train': dataloader_train, 'valid': dataloader_valid}
model = StackedRNNFM(embedding_matrix,
PADDING_LENGTH,
hidden_size=64,
out_hidden_dim=64,
embed_drop=0.2,
out_drop=0.3)
model.to(output_device)
criteria = [[nn.BCEWithLogitsLoss(reduction='mean')], [1.0]]
metric = f1_from_logits
optimizer = optim.Adam(model.parameters(), lr=0.001)
scheduler = None
model_save_path = str(
DATA_DIR.joinpath(
f'models/{Path(__file__).stem}_fold_{fold}.model'))
model_name = model._get_name()
config = {
'epochs': epochs,
'loss_names': ['BCE Loss'],
'metric_type': 'batch',
'model_save_path': model_save_path,
'output_device': output_device,
'mode': 'max',
'early_stopping': 200,
'model_name': model_name,
'reg_lambda': None,
'fold': fold
}
model, valid_score, best_epoch = train_model(model, criteria, metric,
optimizer, scheduler,
dataloaders, logger,
config)
results.append({
'fold': fold,
'best_score': valid_score,
'best_epoch': best_epoch
})
message = f'Training and evaluation for the fold {fold + 1} / {KFOLD} has been done.\n'
message += f'Validation F1 score: {valid_score}\n'
logger.post(message)
dataloader_valid = DataLoader(dataset=dataset_valid,
batch_size=batch_size,
shuffle=False,
pin_memory=True)
oof_preds[index_valid] = sp.special.expit(
predict(model, dataloader_valid, config).reshape(-1, ))
logger.post(f'K-Fold train and evaluation results: {results}')
logger.info(
'Training and evaluation loop has been done. Start f1 threshold search.'
)
search_result = threshold_search(label_train.reshape(-1, ), oof_preds)
logger.post(
f'Threshold search result - f1: {search_result["f1"]}, threshold: {search_result["threshold"]}'
)
