def handler(context):
print(
f'start training with parameters : {Parameters.as_dict()}, context : {context}'
)
X_train, y_train, cols_train = train_data_loader(DATALAKE_CHANNEL_ID,
DATALAKE_TRAIN_FILE_ID,
LABEL_FIELD, INPUT_FIELDS)
dtrain = lgb.Dataset(X_train, y_train)
if DATALAKE_VAL_FILE_ID:
X_val, y_val, _ = train_data_loader(DATALAKE_CHANNEL_ID,
DATALAKE_VAL_FILE_ID, LABEL_FIELD,
INPUT_FIELDS)
else:
X_val, y_val = None, None
extraction_cb = ModelExtractionCallback()
tensorboard_cb = TensorBoardCallback(statistics, writer)
tensorboard_cb.set_valid(X_val, y_val, Parameters.IS_CLASSIFICATION,
IS_MULTI, Parameters.NUM_CLASS)
callbacks = [
extraction_cb,
tensorboard_cb,
]
lgb.cv(PARAMS,
dtrain,
nfold=Parameters.NFOLD,
early_stopping_rounds=Parameters.EARLY_STOPPING_ROUNDS,
verbose_eval=Parameters.VERBOSE_EVAL,
stratified=STRATIFIED,
callbacks=callbacks,
metrics=Parameters.METRIC,
seed=Parameters.SEED)
models = extraction_cb.raw_boosters
for i, model in enumerate(models):
model.save_model(
os.path.join(ABEJA_TRAINING_RESULT_DIR, f'model_{i}.txt'))
di = {**(Parameters.as_dict()), 'cols_train': cols_train}
lgb_env = open(os.path.join(ABEJA_TRAINING_RESULT_DIR, 'lgb_env.json'),
'w')
json.dump(di, lgb_env)
lgb_env.close()
writer.close()
dtrain,
nfold=Parameters.NFOLD,
early_stopping_rounds=Parameters.EARLY_STOPPING_ROUNDS,
verbose_eval=Parameters.VERBOSE_EVAL,
stratified=STRATIFIED,
callbacks=callbacks,
metrics=Parameters.METRIC,
seed=Parameters.SEED)
# In[10]:
models = extraction_cb.raw_boosters
for i, model in enumerate(models):
model.save_model(os.path.join(ABEJA_TRAINING_RESULT_DIR, f'model_{i}.txt'))
di = {**(Parameters.as_dict()), 'cols_train': cols_train}
lgb_env = open(os.path.join(ABEJA_TRAINING_RESULT_DIR, 'lgb_env.json'), 'w')
json.dump(di, lgb_env)
lgb_env.close()
writer.close()
# In[ ]:
def handler(context):
print("finish.")
# In[ ]:
parameters = Parameters.as_env()
def handler(context):
print(
f'start training with parameters : {Parameters.as_dict()}, context : {context}'
)
X_train, y_train, cols_train = train_data_loader(DATALAKE_CHANNEL_ID,
DATALAKE_TRAIN_FILE_ID,
LABEL_FIELD, INPUT_FIELDS)
models = []
pred = np.zeros(len(X_train))
if DATALAKE_VAL_FILE_ID:
X_val, y_val, _ = train_data_loader(DATALAKE_CHANNEL_ID,
DATALAKE_VAL_FILE_ID, LABEL_FIELD,
INPUT_FIELDS)
if IS_MULTI:
pred_val = np.zeros((len(X_val), NUM_CLASS))
else:
pred_val = np.zeros(len(X_val))
else:
X_val, y_val, pred_val = None, None, None
for i, (train_index, valid_index) in enumerate(skf.split(X_train,
y_train)):
model = classifier(**PARAMS)
model.fit(X_train.iloc[train_index], y_train[train_index])
pred[valid_index] = model.predict(X_train.iloc[valid_index])
score, loss = evaluator(y_train[valid_index], pred[valid_index])
score_val = 0.0
loss_val = 0.0
filename = os.path.join(ABEJA_TRAINING_RESULT_DIR, f'model_{i}.pkl')
pickle.dump(model, open(filename, 'wb'))
models.append(model)
if DATALAKE_VAL_FILE_ID:
pred_val_cv = model.predict(X_val)
if IS_MULTI:
pred_val += np.identity(NUM_CLASS)[pred_val_cv]
else:
pred_val += pred_val_cv
score_val, loss_val = evaluator(y_val, pred_val_cv)
print('-------------')
print(
'cv {} || score:{:.4f} || loss:{:.4f} || val_score:{:.4f} || val_loss:{:.4f}'
.format(i + 1, score, loss, score_val, loss_val))
writer.add_scalar('main/acc', score, i + 1)
writer.add_scalar('main/loss', loss, i + 1)
writer.add_scalar('test/acc', score_val, i + 1)
writer.add_scalar('test/loss', loss_val, i + 1)
statistics(i + 1, loss, score, loss_val, score_val)
writer.flush()
score, loss = evaluator(y_train, pred)
score_val = 0.0
loss_val = 0.0
if DATALAKE_VAL_FILE_ID:
if IS_MULTI:
pred_val = np.argmax(pred_val, axis=1)
else:
pred_val /= len(models)
score_val, loss_val = evaluator(y_val, pred_val)
print('-------------')
print(
'cv total score:{:.4f} || cv total loss:{:.4f} || cv total val_score:{:.4f} || cv total val_loss:{:.4f}'
.format(score, loss, score_val, loss_val))
statistics(Parameters.NFOLD, None, score, None, score_val)
writer.add_scalar('main/acc', score, Parameters.NFOLD)
writer.add_scalar('main/loss', loss, Parameters.NFOLD)
writer.add_scalar('test/acc', score_val, Parameters.NFOLD)
writer.add_scalar('test/loss', loss_val, Parameters.NFOLD)
writer.close()
di = {**(Parameters.as_dict()), 'cols_train': cols_train}
skf_env = open(os.path.join(ABEJA_TRAINING_RESULT_DIR, 'skf_env.json'),
'w')
json.dump(di, skf_env)
skf_env.close()
return
