def main():
# Check Gpu Enable
physical_devices = tf.config.experimental.list_physical_devices('GPU')
if len(physical_devices) > 0:
tf.config.experimental.set_memory_growth(physical_devices[0], True)
# parsing Arg
train_data_path = args.train_data
max_len = args.max_length
epoch = args.epochs
batch_size = args.batch_size
att_reg = args.att_reg
lr_decay = args.lr_decay
lr_rate = args.lr_rate
warmup_lr_rate = lr_rate*0.1
patience = args.patience
#period = args.save_period
weight_save_path = args.weight_save_path
document = args.document
label = args.label
channel = args.channel
steps_per_epoch = args.train_steps
# model weight pah
if os.path.isdir(weight_save_path) == False:
os.mkdir(weight_save_path)
# Read Data
if ".csv" in train_data_path:
read_data = pd.read_csv
elif ".xlsx" in train_data_path:
read_data = pd.read_excel
else:
read_data = pd.read_table
train_data = read_data(train_data_path)
# Make Tokenizer Token
tk = Token("Tokenizer", max_len)
train_data["Token"] = train_data[document].apply(lambda x: tk.make_token_ori(x))
# Using Keras Tokenizer
k_tokenizer = keras.preprocessing.text.Tokenizer(filters='')
k_tokenizer.fit_on_texts(train_data["Token"].values.tolist())
words_count = len(k_tokenizer.word_counts)
print("Save Keras tokenizer for validate in %s"%(weight_save_path))
with open(os.path.join(weight_save_path,"keras_tokenizer.pkl"), "wb") as f:
pickle.dump(k_tokenizer, f)
# Load Pre-trained embedding Word2Vec
w2v_model = word2vec.Word2Vec.load("w2v_pretrain_emb/w2v_20M_500.model")
init_weight = np.random.uniform(size=(words_count + 1, 500), low=-1, high=1)
words_lst = []
for i in range(1, len(k_tokenizer.index_word) + 1):
words = k_tokenizer.index_word[i]
try:
words_vector = w2v_model.wv[words]
except:
words_lst.append([i, words])
init_weight[i] = words_vector
# K_tokenizer Sequence
sequences = k_tokenizer.texts_to_sequences(train_data['Token'])
x_train = keras.preprocessing.sequence.pad_sequences(sequences, maxlen=max_len)
y_train = train_data[label].values
# Define validation set 분할
x_train2, x_val, y_train2, y_val = train_test_split(x_train, y_train, test_size=0.1, random_state=0)
# Build simple binary model
tf.keras.backend.clear_session()
amcnn = AMCNN(maxlen=max_len,
embed_dim=500,
words_count=words_count,
filter_size=50,
channel=channel,
mask_prob=0.5,
att_reg=att_reg)
model = amcnn.build(emb_trainable=False, emb_weight=init_weight)
model.compile(optimizer=tf.keras.optimizers.Adam(warmup_lr_rate), loss="binary_crossentropy",
metrics=["accuracy", k_precision, k_recall, k_f1score])
checkpoint_path = os.path.join(weight_save_path,"model-{epoch:04d}.h5")
# Define callbacks condition
callbacks = ModelCheckpoint(checkpoint_path, monitor='val_loss',
verbose=1,save_best_only=True, save_weights_only=True) # period=period,
# Define reduce Learning rate schedule
lr_reducer = ReduceLROnPlateau(monitor='val_loss',
factor=lr_decay,
cooldown=0,
patience=patience,
min_lr=lr_rate*0.01,
verbose=1)
# Train Warm up stage
print("===========Warm up %d Epoch Stage==========="%(int(epoch*0.1)))
# warm up embedding weight
model.fit(x_train2, y_train2, epochs=int(epoch*0.1), callbacks=[callbacks, lr_reducer], steps_per_epoch=steps_per_epoch,
batch_size=batch_size,verbose=2)
print("============Main %d Epoch Stage============="%(epoch-int(epoch*0.1)))
K.set_value(model.optimizer.learning_rate, lr_rate)
model.fit(x_train2, y_train2, epochs=epoch-int(epoch * 0.1), callbacks=[callbacks, lr_reducer], steps_per_epoch=steps_per_epoch
,validation_steps=steps_per_epoch,
batch_size=batch_size, validation_data=(x_val, y_val),verbose=2)
print("Complete Training Model")
print("Check Model Weight file in %s"%(weight_save_path))
def main():
# Check Gpu Enable
physical_devices = tf.config.experimental.list_physical_devices('GPU')
if len(physical_devices) > 0:
tf.config.experimental.set_memory_growth(physical_devices[0], True)
# parsing Arg
test_data_path = args.test_data
max_len = args.max_length
att_reg = args.att_reg
weight_save_path = args.weight_save_path
document = args.document
label = args.label
channel = args.channel
val_model_epoch = args.val_model_epoch
# Read Data
if ".csv" in test_data_path:
read_data = pd.read_csv
elif ".xlsx" in test_data_path:
read_data = pd.read_excel
else:
read_data = pd.read_table
test_data = read_data(test_data_path)
# Make Tokenizer Token
tk = Token("Tokenizer", max_len)
test_data["Token"] = test_data[document].apply(
lambda x: tk.make_token_ori(x))
# Using Keras Tokenizer
print("Load Keras tokenizer for validate in %s" % (weight_save_path))
with open(os.path.join(weight_save_path, "keras_tokenizer.pkl"),
"rb") as f:
k_tokenizer = pickle.load(f)
words_count = len(k_tokenizer.word_counts)
# K_tokenizer Sequence
sequences = k_tokenizer.texts_to_sequences(test_data['Token'])
x_test = keras.preprocessing.sequence.pad_sequences(sequences,
maxlen=max_len)
y_test = test_data[label].values
# Build simple binary model
tf.keras.backend.clear_session()
amcnn = AMCNN(maxlen=max_len,
embed_dim=500,
words_count=words_count,
filter_size=50,
channel=channel,
mask_prob=0.5,
att_reg=att_reg)
model = amcnn.build(pre_emb=False)
if val_model_epoch == -1:
model_lst = [i for i in os.listdir(weight_save_path) if ".h5" in i]
model_weight_path = model_lst[-1]
else:
model_weight_path = "model-%4d.h5" % (val_model_epoch)
model_weight_path = model_weight_path.replace(" ", "0")
model.load_weights(os.path.join(weight_save_path, model_weight_path))
print("Evaluate %s Test data" %
(os.path.join(weight_save_path, model_weight_path)))
pred_test = model.predict(x_test, verbose=1)
pred_test2 = np.int32(pred_test >= 0.5).reshape(-1)
print("==============Evaluate Result============")
print("f1_score :", f1_score(y_test, pred_test2))
print("acc_score :", accuracy_score(y_test, pred_test2))
print("recall_score :", recall_score(y_test, pred_test2))
print("precision_score :", precision_score(y_test, pred_test2))
