def main(logger):
df_train, df_test = load_data(INPUT_DIR, logger)
logger.info('Preprocess text')
df_train = preprocess_text(df_train.iloc[:200000])
df_test = preprocess_text(df_test)
seq_train, tokenizer = tokenize_text(df_train, logger)
seq_test, _ = tokenize_text(df_test, logger, tokenizer=tokenizer)
text_train = df_train['question_text'].values.tolist()
embedding_matrix = np.random.rand(len(tokenizer.word_index) + 1, 300)
scdv = SCDV(embedding_matrix,
tokenizer,
logger,
num_clusters=50,
gmm_path=DATA_DIR.joinpath('gmm_tmp.pkl'))
with logger.timer('SCDV computation on train data'):
scdv_train = scdv.fit_transform(text_train, seq_train)
logger.post(
f'Computing SCDV for train data has been done: shape = {scdv_train.shape}'
)
with logger.timer('SCDV computation on test data'):
scdv_test = scdv.transform(seq_test)
logger.post(
f'Computing SCDV for test data has been done: shape = {scdv_test.shape}'
)
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, 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])
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)
dataloader_train = DataLoader(dataset=dataset_train,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
worker_init_fn=worker_init_fn)
dataloader_valid = DataLoader(dataset=dataset_valid,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
worker_init_fn=worker_init_fn)
dataloaders = {'train': dataloader_train, 'valid': dataloader_valid}
model = StackedNormalizedRNNFM(embedding_matrix,
PADDING_LENGTH,
hidden_size=64,
out_hidden_dim=64,
embed_drop=0.2,
out_drop=0.3,
residual=False)
model.to(output_device)
criteria = [[nn.BCEWithLogitsLoss(reduction='mean')], [1.0]]
metric = f1_from_logits_optimized
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"]}'
)
def main(logger, args):
df_train, _ = load_data(INPUT_DIR, logger)
if args['debug']:
df_train = df_train.iloc[:30000]
logger.info('Extract nlp features')
df_train = extract_nlp_features(df_train)
else:
logger.info('Extract nlp features')
df_train = extract_nlp_features(df_train)
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)
if args['debug']:
embedding_matrix = np.random.rand(len(tokenizer.word_index) + 1,
300).astype(np.float32)
else:
logger.info('Load multiple embeddings')
embedding_matrices = load_multiple_embeddings(
tokenizer.word_index,
embed_types=[0, 2],
max_workers=args['max_workers'])
embedding_matrix = np.array(embedding_matrices).mean(0)
continuous_columns = [
'total_length', 'n_capitals', 'n_words', 'n_puncts', 'n_?', 'n_!',
'n_you'
]
for col in continuous_columns:
scaler = StandardScaler()
df_train[col] = scaler.fit_transform(df_train[col].values.astype(
np.float32).reshape(-1, 1)).reshape(-1, )
x_continuous = [
df_train[col].values.reshape(-1, 1) for col in continuous_columns
]
# ===== 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)
trigger = TRIGGER
if args['debug']:
epochs = 3
n_splits = 2
else:
epochs = EPOCHS
n_splits = KFOLD
logger.info('Start training and evaluation loop')
model_specs = [{
'nlp_dim': 16,
'nlp_dropout': 0.2,
'num_dense_layers': 2,
'mask': True
}]
model_name_base = 'NLPFeaturesRNN'
for spec_id, spec in enumerate(model_specs):
model_name = model_name_base + f'_specId={spec_id}_nlpdim={spec["nlp_dim"]}_nlpdrop={spec["nlp_dropout"]}'
model_name += f'_numlayers={spec["num_dense_layers"]}_mask={spec["mask"]}'
skf = StratifiedKFold(n_splits=n_splits,
shuffle=True,
random_state=SEED)
oof_mv_preds = np.zeros(len(seq_train))
oof_preds_proba = np.zeros(len(seq_train))
oof_opt_preds = np.zeros(len(seq_train))
oof_reopt_preds = np.zeros(len(seq_train))
results_list = []
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 = {
'text': seq_train[index_train].astype(int),
'continuous': [x[index_train] for x in x_continuous]
}
x_valid = {
'text': seq_train[index_valid].astype(int),
'continuous': [x[index_valid] for x in x_continuous]
}
y_train, y_valid = label_train[index_train].astype(
np.float32), label_train[index_valid].astype(np.float32)
model = NLPFeaturesRNN({'continuous': len(x_continuous)},
embedding_matrix,
PADDING_LENGTH,
hidden_size=64,
out_hidden_dim=64,
out_drop=0.3,
embed_drop=0.1,
dense_activate='relu',
nlp_hidden_dim=spec['nlp_dim'],
mask=spec['mask'],
nlp_dropout=spec['nlp_dropout'],
factorize=False,
num_dense_layers=spec['num_dense_layers'])
steps_per_epoch = seq_train[index_train].shape[0] // batch_size
scheduler_trigger_steps = steps_per_epoch * trigger
step_size = steps_per_epoch * (epochs - trigger) // NUM_SNAPSHOTS
config = {
'epochs': epochs,
'batch_size': batch_size,
'output_device': output_device,
'criterion_type': 'bce',
'criteria_weights': [1.0, 1.0],
'criterion_gamma': 2.0,
'criterion_alpha': 0.75,
'optimizer': 'adam',
'optimizer_lr': 0.003,
'num_snapshots': NUM_SNAPSHOTS,
'scheduler_type': 'cyclic',
'base_lr': 0.0005,
'max_lr': 0.003,
'step_size': step_size,
'scheduler_mode': 'triangular',
'scheduler_gamma': 0.9,
'scheduler_trigger_steps': scheduler_trigger_steps,
'sampler_type': 'normal',
'seed': SEED
}
trainer = Trainer(model, logger, config)
eval_results = trainer.train_and_eval_fold(x_train, y_train,
x_valid, y_valid, fold)
fold_results = calculate_fold_metrics(
eval_results, label_train[index_valid].reshape(-1, ))
results_list.append(fold_results)
message = f'Fold {fold + 1} / {KFOLD} has been done.\n'
message += f'Majority Voting - F1: {fold_results["oof_mv_f1"]}, '
message += f'Precision: {fold_results["oof_mv_precision"]}, Recall: {fold_results["oof_mv_recall"]}\n'
message += f'Optimized - F1: {fold_results["oof_opt_f1"]}, '
message += f'Precision: {fold_results["oof_opt_precision"]}, Recall: {fold_results["oof_opt_recall"]}\n'
message += f'Re-optimized - F1: {fold_results["oof_reopt_f1"]}, '
message += f'Precision: {fold_results["oof_reopt_precision"]}, Recall: {fold_results["oof_reopt_recall"]}\n'
message += f'Focal Loss: {fold_results["oof_focal_loss"]}, '
message += f'Optimized Threshold: {fold_results["oof_opt_threshold"]}, '
message += f'Re-optimized Threshold: {fold_results["oof_reopt_threshold"]}, '
logger.post(message)
eval_results_addition = {
'date': datetime.now(),
'script_name': SCRIPT_NAME,
'spec_id': spec_id,
'model_name': model_name,
'fold_id': fold
}
for res in eval_results:
res.update(eval_results_addition)
post_to_snapshot_metrics_table(data=res,
project_id=BQ_PROJECT_ID,
dataset_name=BQ_DATASET)
fold_results_addition = {
'date': datetime.now(),
'script_name': SCRIPT_NAME,
'spec_id': spec_id,
'model_name': model_name,
'fold_id': fold
}
fold_results.update(fold_results_addition)
post_to_fold_metrics_table(fold_results,
project_id=BQ_PROJECT_ID,
dataset_name=BQ_DATASET)
oof_mv_preds[index_valid] = fold_results['oof_mv_preds']
oof_opt_preds[index_valid] = fold_results['oof_opt_preds']
oof_reopt_preds[index_valid] = fold_results['oof_reopt_preds']
oof_preds_proba[index_valid] = fold_results['oof_preds_proba']
results = calculate_total_metrics(results_list)
results_addition = {
'date': datetime.now(),
'script_name': SCRIPT_NAME,
'spec_id': spec_id,
'model_name': model_name
}
results.update(results_addition)
post_to_total_metrics_table(results,
project_id=BQ_PROJECT_ID,
dataset_name=BQ_DATASET)
message = 'KFold training and evaluation has been done.\n'
message += f'Majority Voting - F1: avg = {results["mv_f1_avg"]}, std = {results["mv_f1_std"]}, '
message += f'Precision: {results["mv_precision_avg"]}, Recall: {results["mv_recall_avg"]}\n'
message += f'Optimized - F1: avg = {results["opt_f1_avg"]}, std = {results["opt_f1_std"]}, '
message += f'Precision: {results["opt_precision_avg"]}, Recall: {results["opt_recall_avg"]}\n'
message += f'Re-optimized - F1: avg = {results["reopt_f1_avg"]}, std = {results["reopt_f1_std"]}, '
message += f'Precision: {results["reopt_precision_avg"]}, Recall: {results["reopt_recall_avg"]}\n'
mv_thresholds = ",".join(
[str(th) for th in results["mv_thresholds_avg"]])
message += f'Focal Loss: {results["focal_loss_avg"]}, '
message += f'Optimized Threshold: {results["opt_threshold_avg"]}, '
message += f'Re-optimized Threshold: {results["reopt_threshold_avg"]}\n'
message += f'Majority Voting Thresholds: {mv_thresholds}'
logger.post(message)
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, 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_name = 'StackedCNNRNN'
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 = StackedCNNRNN(embedding_matrix,
PADDING_LENGTH,
hidden_size=32,
out_hidden_dim=32,
kernel_sizes=(3, 5),
seq_dropout=0.2,
out_drop=0.2,
embed_drop=0.1)
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': 1200,
'scheduler_mode': 'triangular',
'scheduler_gamma': 0.9,
'scheduler_trigger_steps': 4000,
'sampler_type': 'normal',
'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 main(logger, args):
df_train, _ = load_data(INPUT_DIR, logger)
if args['debug']:
df_train = df_train.iloc[:30000]
texts_train = df_train['question_text']
else:
logger.info('Preprocess text')
texts_train = preprocess_text(df_train, return_df=False)
seq_train, tokenizer = tokenize_texts(texts_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)
embed_types = [0, 1, 2]
logger.info(
'Start multiprocess nlp feature extraction and embedding matrices loading'
)
with mp.Pool(processes=2) as p:
results = p.map(parallel_apply,
[(extract_nlp_features, (df_train, )),
(load_multiple_embeddings,
(tokenizer.word_index, embed_types, args['debug']))])
df_train_extracted = results[0]
embedding_matrices = results[1]
embedding_matrix = np.concatenate(
[np.array([embedding_matrices[i] for i in [0, 1, 2]]).mean(0)] +
[embedding_matrices[j] for j in [1]],
axis=1)
nlp_columns = [
'total_length', 'n_capitals', 'n_words', 'n_puncts', 'n_?', 'n_!',
'n_you'
]
for col in nlp_columns:
scaler = StandardScaler()
df_train_extracted[col] = scaler.fit_transform(
df_train_extracted[col].values.astype(np.float32).reshape(
-1, 1)).reshape(-1, )
x_nlp = [
df_train_extracted[col].values.reshape(-1, 1) for col in nlp_columns
]
nlp_size = len(x_nlp)
# ===== 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)
trigger = TRIGGER
if args['debug']:
epochs = 3
n_splits = 2
else:
epochs = EPOCHS
n_splits = KFOLD
logger.info('Start training and evaluation loop')
model_specs = [
{
'nlp_layer_types': ({
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}, {
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}),
'rnn_layer_types': ({
'type': 'lstm',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}, {
'type': 'gru',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}),
'upper_layer_types': ({
'dim': 64,
'dropout': 0.5
}, {
'dim': 64,
'dropout': 0.3
})
},
{
'nlp_layer_types': ({
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}, {
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}),
'rnn_layer_types': ({
'type': 'lstm',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}, {
'type': 'gru',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}),
'upper_layer_types': ({
'dim': 128,
'dropout': 0.5
}, {
'dim': 64,
'dropout': 0.3
})
},
{
'nlp_layer_types': ({
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}, {
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}),
'rnn_layer_types': ({
'type': 'lstm',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}, {
'type': 'gru',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}),
'upper_layer_types': ({
'dim': 128,
'dropout': 0.5
}, {
'dim': 128,
'dropout': 0.3
})
},
{
'nlp_layer_types': ({
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}, {
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}),
'rnn_layer_types': ({
'type': 'lstm',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}, {
'type': 'gru',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}),
'upper_layer_types': ({
'dim': 256,
'dropout': 0.5
}, {
'dim': 64,
'dropout': 0.3
})
},
{
'nlp_layer_types': ({
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}, {
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}),
'rnn_layer_types': ({
'type': 'lstm',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}, {
'type': 'gru',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}),
'upper_layer_types': ({
'dim': 64,
'dropout': 0.5
}, {
'dim': 64,
'dropout': 0.5
}, {
'dim': 64,
'dropout': 0.3
})
},
{
'nlp_layer_types': ({
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}, {
'activation': 'relu',
'dim': 16,
'dropout': 0.2
}),
'rnn_layer_types': ({
'type': 'lstm',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}, {
'type': 'gru',
'dim': 64,
'num_layers': 1,
'dropout': 0.0
}),
'upper_layer_types': ({
'dim': 128,
'dropout': 0.5
}, {
'dim': 128,
'dropout': 0.5
}, {
'dim': 64,
'dropout': 0.3
})
},
]
model_name_base = 'NLPFeaturesDeepRNN'
for spec_id, spec in enumerate(model_specs):
model_name = model_name_base + f'_specId={spec_id}'
skf = StratifiedKFold(n_splits=n_splits,
shuffle=True,
random_state=SEED)
oof_mv_preds = np.zeros(len(seq_train))
oof_preds_proba = np.zeros(len(seq_train))
oof_opt_preds = np.zeros(len(seq_train))
oof_reopt_preds = np.zeros(len(seq_train))
results_list = []
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 = {
'text': seq_train[index_train].astype(int),
'nlp': [x[index_train] for x in x_nlp]
}
x_valid = {
'text': seq_train[index_valid].astype(int),
'nlp': [x[index_valid] for x in x_nlp]
}
y_train, y_valid = label_train[index_train].astype(
np.float32), label_train[index_valid].astype(np.float32)
model = NLPFeaturesDeepRNN(
embedding_matrix,
PADDING_LENGTH,
nlp_size,
embed_drop=0.2,
mask=True,
nlp_layer_types=spec['nlp_layer_types'],
rnn_layer_types=spec['rnn_layer_types'],
upper_layer_types=spec['upper_layer_types'])
steps_per_epoch = seq_train[index_train].shape[0] // batch_size
scheduler_trigger_steps = steps_per_epoch * trigger
step_size = steps_per_epoch * (epochs - trigger) // NUM_SNAPSHOTS
config = {
'epochs': epochs,
'batch_size': batch_size,
'output_device': output_device,
'criterion_type': 'bce',
'criteria_weights': [1.0, 1.0],
'criterion_gamma': 2.0,
'criterion_alpha': 0.75,
'optimizer': 'adam',
'optimizer_lr': 0.003,
'num_snapshots': NUM_SNAPSHOTS,
'scheduler_type': 'cyclic',
'base_lr': 0.0005,
'max_lr': 0.003,
'step_size': step_size,
'scheduler_mode': 'triangular',
'scheduler_gamma': 0.9,
'scheduler_trigger_steps': scheduler_trigger_steps,
'sampler_type': 'normal',
'seed': SEED
}
trainer = Trainer(model, logger, config)
eval_results = trainer.train_and_eval_fold(x_train, y_train,
x_valid, y_valid, fold)
fold_results = calculate_fold_metrics(
eval_results, label_train[index_valid].reshape(-1, ))
results_list.append(fold_results)
message = f'Fold {fold + 1} / {KFOLD} has been done.\n'
message += f'Majority Voting - F1: {fold_results["oof_mv_f1"]}, '
message += f'Precision: {fold_results["oof_mv_precision"]}, Recall: {fold_results["oof_mv_recall"]}\n'
message += f'Optimized - F1: {fold_results["oof_opt_f1"]}, '
message += f'Precision: {fold_results["oof_opt_precision"]}, Recall: {fold_results["oof_opt_recall"]}\n'
message += f'Re-optimized - F1: {fold_results["oof_reopt_f1"]}, '
message += f'Precision: {fold_results["oof_reopt_precision"]}, Recall: {fold_results["oof_reopt_recall"]}\n'
message += f'Focal Loss: {fold_results["oof_focal_loss"]}, '
message += f'Optimized Threshold: {fold_results["oof_opt_threshold"]}, '
message += f'Re-optimized Threshold: {fold_results["oof_reopt_threshold"]}, '
logger.post(message)
eval_results_addition = {
'date': datetime.now(),
'script_name': SCRIPT_NAME,
'spec_id': spec_id,
'model_name': model_name,
'fold_id': fold
}
for res in eval_results:
res.update(eval_results_addition)
# post_to_snapshot_metrics_table(data=res, project_id=BQ_PROJECT_ID, dataset_name=BQ_DATASET)
fold_results_addition = {
'date': datetime.now(),
'script_name': SCRIPT_NAME,
'spec_id': spec_id,
'model_name': model_name,
'fold_id': fold
}
fold_results.update(fold_results_addition)
post_to_fold_metrics_table(fold_results,
project_id=BQ_PROJECT_ID,
dataset_name=BQ_DATASET)
oof_mv_preds[index_valid] = fold_results['oof_mv_preds']
oof_opt_preds[index_valid] = fold_results['oof_opt_preds']
oof_reopt_preds[index_valid] = fold_results['oof_reopt_preds']
oof_preds_proba[index_valid] = fold_results['oof_preds_proba']
results = calculate_total_metrics(results_list)
results_addition = {
'date': datetime.now(),
'script_name': SCRIPT_NAME,
'spec_id': spec_id,
'model_name': model_name
}
results.update(results_addition)
if args['save_preds']:
save_path = DATA_DIR.joinpath(
f'predictions/{SCRIPT_NAME + "_" + model_name + ".pkl"}')
predictions = {
'proba': oof_preds_proba,
'mv': oof_mv_preds,
'opt': oof_opt_preds,
'reopt': oof_reopt_preds
}
joblib.dump(predictions, str(save_path))
post_to_total_metrics_table(results,
project_id=BQ_PROJECT_ID,
dataset_name=BQ_DATASET)
logger.post(f'Spec ID: {spec_id}\nModel Spec: {spec}')
message = 'KFold training and evaluation has been done.\n'
message += f'Majority Voting - F1: avg = {results["mv_f1_avg"]}, std = {results["mv_f1_std"]}, '
message += f'Precision: {results["mv_precision_avg"]}, Recall: {results["mv_recall_avg"]}\n'
message += f'Optimized - F1: avg = {results["opt_f1_avg"]}, std = {results["opt_f1_std"]}, '
message += f'Precision: {results["opt_precision_avg"]}, Recall: {results["opt_recall_avg"]}\n'
message += f'Re-optimized - F1: avg = {results["reopt_f1_avg"]}, std = {results["reopt_f1_std"]}, '
message += f'Precision: {results["reopt_precision_avg"]}, Recall: {results["reopt_recall_avg"]}\n'
mv_thresholds = ", ".join(
[str(th) for th in results["mv_thresholds_avg"]])
message += f'Focal Loss: {results["focal_loss_avg"]}, '
message += f'Optimized Threshold: {results["opt_threshold_avg"]}, '
message += f'Re-optimized Threshold: {results["reopt_threshold_avg"]}\n'
message += f'Majority Voting Thresholds: {mv_thresholds}'
logger.post(message)
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,
padding='post',
truncating='post')
pos_train = np.repeat([np.arange(PADDING_LENGTH) + 1],
seq_train.shape[0],
axis=0)
pos_train = pos_train * np.not_equal(seq_train, 0)
label_train = df_train['target'].values.reshape(-1, 1)
if args['debug']:
embedding_matrix = np.random.rand(len(tokenizer.word_index) + 1,
300).astype(np.float32)
else:
logger.info('Load multiple embeddings')
embedding_matrices = load_multiple_embeddings(
tokenizer.word_index,
embed_types=[0, 2],
max_workers=args['max_workers'])
embedding_matrix = np.array(embedding_matrices).mean(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
trigger = TRIGGER
logger.info('Start training and evaluation loop')
model_specs = [{
'attention_type': 'general',
'num_layers': 2
}, {
'attention_type': 'dot',
'num_layers': 2
}, {
'attention_type': 'general',
'num_layers': 1
}, {
'attention_type': 'dot',
'num_layers': 1
}]
model_name_base = 'TransformerRNN'
for spec_id, spec in enumerate(model_specs):
model_name = model_name_base + f'_specId={spec_id}_attentiontype={spec["attention_type"]}'
model_name += f'_numlayers={spec["num_layers"]}'
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 = {
'sequence': seq_train[index_train].astype(int),
'position': pos_train[index_train].astype(int)
}
x_valid = {
'sequence': seq_train[index_valid].astype(int),
'position': pos_train[index_valid].astype(int)
}
y_train, y_valid = label_train[index_train].astype(
np.float32), label_train[index_valid].astype(np.float32)
model = TransformerRNN(embedding_matrix,
PADDING_LENGTH,
hidden_dim=64,
out_hidden_dim=64,
out_drop=0.3,
embed_drop=0.2,
attention_type=spec['attention_type'],
num_layers=spec['num_layers'])
steps_per_epoch = seq_train[index_train].shape[0] // batch_size
scheduler_trigger_steps = steps_per_epoch * trigger
step_size = steps_per_epoch * (epochs - trigger) // NUM_SNAPSHOTS
config = {
'epochs': epochs,
'batch_size': batch_size,
'output_device': output_device,
'criterion_type': 'bce',
'criteria_weights': [0.5, 0.5],
'criterion_gamma': 2.0,
'criterion_alpha': 0.75,
'optimizer': 'adam',
'optimizer_lr': 0.001,
'num_snapshots': NUM_SNAPSHOTS,
'scheduler_type': 'cyclic',
'base_lr': 0.0005,
'max_lr': 0.001,
'step_size': step_size,
'scheduler_mode': 'triangular',
'scheduler_gamma': 0.9,
'scheduler_trigger_steps': scheduler_trigger_steps,
'sampler_type': 'normal',
'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 main(logger, args):
df_train, _ = load_data(INPUT_DIR, logger)
if args['debug']:
df_train = df_train.iloc[:200000]
logger.info('Extract nlp features')
df_train = extract_nlp_features(df_train)
else:
logger.info('Extract nlp features')
df_train = extract_nlp_features(df_train)
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)
if args['debug']:
embedding_matrix = np.random.rand(len(tokenizer.word_index) + 1,
300).astype(np.float32)
else:
logger.info('Load multiple embeddings')
embedding_matrices = load_multiple_embeddings(
tokenizer.word_index,
embed_types=[0, 2],
max_workers=args['max_workers'])
embedding_matrix = np.array(embedding_matrices).mean(0)
continuous_columns = [
'total_length', 'n_capitals', 'n_words', 'n_puncts', 'n_?', 'n_!',
'n_you'
]
for col in continuous_columns:
scaler = StandardScaler()
df_train[col] = scaler.fit_transform(df_train[col].values.astype(
np.float32).reshape(-1, 1)).reshape(-1, )
x_continuous = [
df_train[col].values.reshape(-1, 1) for col in continuous_columns
]
# ===== 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
trigger = TRIGGER
logger.info('Start training and evaluation loop')
model_specs = [
{
'nlp_dim': 64,
'nlp_dropout': 0.2,
'num_dense_layers': 1,
'mask': False
},
{
'nlp_dim': 32,
'nlp_dropout': 0.2,
'num_dense_layers': 1,
'mask': False
},
{
'nlp_dim': 16,
'nlp_dropout': 0.2,
'num_dense_layers': 2,
'mask': False
},
{
'nlp_dim': 32,
'nlp_dropout': 0.2,
'num_dense_layers': 2,
'mask': False
},
{
'nlp_dim': 64,
'nlp_dropout': 0.5,
'num_dense_layers': 2,
'mask': False
},
{
'nlp_dim': 32,
'nlp_dropout': 0.5,
'num_dense_layers': 1,
'mask': False
},
{
'nlp_dim': 64,
'nlp_dropout': 0.2,
'num_dense_layers': 1,
'mask': True
},
{
'nlp_dim': 32,
'nlp_dropout': 0.2,
'num_dense_layers': 2,
'mask': True
},
]
model_name_base = 'NLPFeaturesRNN'
for spec_id, spec in enumerate(model_specs):
model_name = model_name_base + f'_specId={spec_id}_nlpdim={spec["nlp_dim"]}_nlpdrop={spec["nlp_dropout"]}'
model_name += f'_numlayers={spec["num_dense_layers"]}_mask={spec["mask"]}'
skf = StratifiedKFold(n_splits=KFOLD, shuffle=True, random_state=SEED)
oof_preds_optimized = np.zeros(len(seq_train))
oof_preds_majority = np.zeros(len(seq_train))
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 = {
'text': seq_train[index_train].astype(int),
'continuous': [x[index_train] for x in x_continuous]
}
x_valid = {
'text': seq_train[index_valid].astype(int),
'continuous': [x[index_valid] for x in x_continuous]
}
y_train, y_valid = label_train[index_train].astype(
np.float32), label_train[index_valid].astype(np.float32)
model = NLPFeaturesRNN({'continuous': len(x_continuous)},
embedding_matrix,
PADDING_LENGTH,
hidden_size=64,
out_hidden_dim=64,
out_drop=0.3,
embed_drop=0.1,
dense_activate='relu',
nlp_hidden_dim=spec['nlp_dim'],
mask=spec['mask'],
nlp_dropout=spec['nlp_dropout'],
factorize=False,
num_dense_layers=spec['num_dense_layers'])
steps_per_epoch = seq_train[index_train].shape[0] // batch_size
scheduler_trigger_steps = steps_per_epoch * trigger
step_size = steps_per_epoch * (epochs - trigger) // NUM_SNAPSHOTS
config = {
'epochs': epochs,
'batch_size': batch_size,
'output_device': output_device,
'criterion_type': 'bce',
'criteria_weights': [1.0, 1.0],
'criterion_gamma': 2.0,
'criterion_alpha': 0.75,
'optimizer': 'adam',
'optimizer_lr': 0.003,
'num_snapshots': NUM_SNAPSHOTS,
'scheduler_type': 'cyclic',
'base_lr': 0.0005,
'max_lr': 0.003,
'step_size': step_size,
'scheduler_mode': 'triangular',
'scheduler_gamma': 0.9,
'scheduler_trigger_steps': scheduler_trigger_steps,
'sampler_type': 'normal',
'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)