def train(lang='pt'):
params = PARAMS.copy()
initial_epoch = 0
X, Y = util.get_X_Y(data_type='keras_tokenized_tri', lang=lang, file_type="dump")
X = np.asarray(X)
params['embedding_matrix'] = load_embedding_matrix(name="fasttext_sg_tri_8", tokenizer='keras_tokenized_tri',lang=lang, model_type="dump")
params["vocab_size"] = params['embedding_matrix'].shape[0]
params["embedding_dim"] = params['embedding_matrix'].shape[1]
if not os.path.exists(PATH):
os.makedirs(PATH)
if not os.path.exists(PATH+'log_dir'):
os.makedirs(PATH+'log_dir')
#params["loss"] = util.focal_loss(gamma=5.,alpha=1588)
lastest_model = load_lastest(lang=lang)
if(lastest_model == None):
model, params = generate_model(params)
else:
model = lastest_model[0]
initial_epoch = lastest_model[1]
print(model.metrics_names)
params['sampler'] = FunctionSampler(func=balance_dataset,
kw_args={'cut_off': 0.5,
'random_state': 42})
data_generator = DataGenerator(X,Y, lang=lang, process_x=process_x, process_y=process_y, batch_size=PARAMS['batch_size'], sampler=params['sampler'])
#data_generator.remove_reliable_0(pct=1.0)
validation_data = data_generator.get_validation_data()
print('data_generator.x: ', data_generator.__getitem__(0)[0][0:5])
print('data_generator.y: ', data_generator.__getitem__(0)[1][0:5])
#params["class_weights"]= data_generator.get_classes_weights()
reduce_lr = ReduceLROnPlateau(monitor='val_categorical_accuracy', factor=0.2, patience=3, verbose=1)
early_stopping = EarlyStopping(monitor='val_categorical_accuracy', min_delta=0.02, patience=10, verbose=1)
csv_logger = CSVLogger(PATH+'traning.log', append=True)
tensorboard_callback = TensorBoard(log_dir=PATH+'log_dir', batch_size=params["batch_size"])
model_checkpoint = ModelCheckpoint(filepath=PATH+'weights-{epoch:03d}-{val_categorical_accuracy:.4f}-'+lang+'.hdf5',
monitor='val_categorical_accuracy',
verbose=1,
mode='max')
params["callbacks"] = [model_checkpoint, early_stopping, tensorboard_callback, csv_logger, reduce_lr]
model.fit_generator(data_generator,
epochs=params["epochs"],
verbose=1,
callbacks=params["callbacks"],
validation_data=validation_data,
#workers=7,
#use_multiprocessing=True,
class_weight=params["class_weights"],
initial_epoch=initial_epoch)
def train(lang='pt'):
params = PARAMS.copy()
initial_epoch = 0
X, Y = util.get_X_Y(data_type='keras_tokenized_tri', lang=lang, file_type="dump")
X = np.asarray(X)
params['embedding_matrix'] = load_embedding_matrix(name="fasttext_sg_tri_8", tokenizer='keras_tokenized_tri',lang=lang, model_type="dump")
params["vocab_size"] = params['embedding_matrix'].shape[0]
params["embedding_dim"] = params['embedding_matrix'].shape[1]
if not os.path.exists(PATH):
os.makedirs(PATH)
if not os.path.exists(PATH+'log_dir'):
os.makedirs(PATH+'log_dir')
kfold_count = 1
skf = StratifiedKFold(n_splits=params['k-folds'], shuffle=True)
for train_index, test_index in skf.split(X, Y):
print("TRAIN:", train_index.shape, "TEST:", test_index.shape)
X_train, Y_train = X[train_index], Y[train_index]
X_test, Y_test = X[test_index], Y[test_index]
params['sampler'] = FunctionSampler(func=balance_dataset,
kw_args={'cut_off': 0.5,
'random_state': np.random.randint(0, 100000)})
model, params = generate_model(params)
print(model.metrics_names)
data_g_train = DataGenerator(X_train, Y_train, lang=lang, process_x=process_x, process_y=process_y, sampler=params['sampler'], batch_size=PARAMS['batch_size'], separate_val=False)
data_g_val = DataGenerator(X_test, Y_test, lang=lang, process_x=process_x, process_y=process_y, batch_size=PARAMS['batch_size'], separate_val=False)
print('data_generator.x: ', data_g_train.__getitem__(0)[0][0:5])
print('data_generator.y: ', data_g_train.__getitem__(0)[1][0:5])
#params["class_weights"]= data_generator.get_classes_weights()
reduce_lr = ReduceLROnPlateau(monitor='val_categorical_accuracy', factor=0.2, patience=3, verbose=1)
early_stopping = EarlyStopping(monitor='val_categorical_accuracy', min_delta=0.02, patience=10, verbose=1)
csv_logger = CSVLogger(PATH+'traning.log', append=True)
tensorboard_callback = TensorBoard(log_dir=PATH+'log_dir', batch_size=params["batch_size"])
model_checkpoint = ModelCheckpoint(filepath=PATH+'weights-{epoch:03d}-{val_categorical_accuracy:.4f}-'+lang+'.hdf5',
monitor='val_categorical_accuracy',
verbose=1,
mode='max')
clr = CyclicLR(base_lr=1e-3, max_lr=2e-3,
step_size=300., mode='exp_range',
gamma=0.99994)
params["callbacks"] = [tensorboard_callback, csv_logger, clr]
model.fit_generator(data_g_train,
epochs=params["epochs"],
verbose=1,
callbacks=params["callbacks"],
#workers=7,
#use_multiprocessing=True,
class_weight=params["class_weights"],
initial_epoch=initial_epoch)
batch_val = 1000
data_g_val.set_batch_size(batch_val)
y_pred = np.zeros(Y_test.shape)
y_val = np.zeros(Y_test.shape)
for i, (x, y) in enumerate(data_g_val):
y_pred[i*batch_val:(i+1)*batch_val] = np.argmax(model.predict(x), axis=1)
y_val[i*batch_val:(i+1)*batch_val] = np.argmax(y, axis=1)
result = util.evaluate(y_val, y_pred)
print('Model '+NAME+' val score on '+lang+', k-fold-'+str(kfold_count)+': ', result)
model.save(PATH+'weights-{epoch:03d}-kfold{fold}-{result:.4f}-{lang}.hdf5'.format(epoch=params["epochs"], result=result, lang=lang, fold=kfold_count))
del data_g_train, data_g_val
del model
del X_train, Y_train, X_test, Y_test
K.clear_session()
gc.collect()
kfold_count += 1
