def __init__(self, **kwargs):
self.latent_size = 100
self.initial_epoch = 1
self.epochs = 50000
self.batch_size = 64
self.lr = 0.0002
self.margin = 1
self.pt_ratio = 0.1
self.ckpt_path = kwargs["ckpt_path"]
self.ckpt_epoch = kwargs["ckpt_epoch"]
self.log_folder, self.ckpt_folder, self.image_folder = create_folder(
kwargs["result_folder"])
self.training_result_file = self.log_folder / "training_result.txt"
self.gen = make_generator(self.latent_size)
self.disc = make_discriminator(self.latent_size)
# kwargs 값 저장
msg = ""
for k, v in list(kwargs.items()):
msg += "{} = {}\n".format(k, v)
msg += "new model checkpoint path = {}\n".format(self.ckpt_folder)
with (self.log_folder / "model_settings.txt").open(
"w", encoding="utf-8") as fp:
fp.write(msg)
self.dataset, self.num_train = self.get_dataset(self.batch_size)
save_model_info({
"Generator": self.gen,
"Discriminator": self.disc
}, self.log_folder)
def __init__(self, **kwargs) :
dataset_folder = Path(kwargs["dataset_folder"]).resolve()
check_valid_path(dataset_folder)
result_folder = kwargs["result_folder"]
self.initial_epoch = 1
self.test_mode = kwargs["test"]
self.epochs = kwargs["epochs"]
self.hidden_size = kwargs["hidden_size"]
self.num_heads = kwargs["heads"]
self.use_label_smoothing = kwargs["label_smoothing"]
self.ckpt_path = kwargs["ckpt_path"]
self.ckpt_epoch = kwargs["ckpt_epoch"]
# model에 필요한 폴더 및 파일 생성
self.log_folder, self.ckpt_folder, self.image_folder = create_folder(result_folder)
if not self.test_mode :
self.training_result_file = self.log_folder / "training_result.txt"
self.test_result_file = None
# kwargs 값 저장
msg = ""
for k, v in list(kwargs.items()) :
msg += "{} = {}\n".format(k, v)
msg += "new model checkpoint path = {}\n".format(self.ckpt_folder)
with (self.log_folder / "model_settings.txt").open("w", encoding = "utf-8") as fp :
fp.write(msg)
# 필요한 data를 불러옴
self.src_word2id, self.src_id2word, self.src_vocab_size = load_word_dic(dataset_folder / "src_word2id.pkl")
self.tar_word2id, self.tar_id2word, self.tar_vocab_size = load_word_dic(dataset_folder / "tar_word2id.pkl")
if not self.test_mode :
# encoder data : <END> tag 추가
# decoder data : 1) input = <START> tag만 추가 2)output = <END> tag만 추가
train_src, num_train_src = get_dataset(self.src_word2id, dataset_folder / "train_src.txt", False, True, True)
train_tar, num_train_tar = get_dataset(self.tar_word2id, dataset_folder / "train_tar.txt", True, True, True)
if num_train_src != num_train_tar :
raise Exception("한글 데이터셋({})과 영어 데이터셋({})의 크기가 다릅니다.".format(
num_train_src, num_train_tar))
self.num_train = num_train_src
self.train_dataset = tf.data.Dataset.from_generator(lambda: zip(train_src, train_tar), (tf.int32, tf.int32))
self.train_dataset = self.train_dataset.cache().shuffle(self.num_train + 1).padded_batch(
batch_size = kwargs["batch_size"], padded_shapes = (tf.TensorShape([None]), tf.TensorShape([None])),
padding_values = (self.src_word2id["<PAD>"], self.tar_word2id["<PAD>"])).prefetch(1)
test_src_path = dataset_folder / "test.txt"
if test_src_path.exists() :
test_src, self.num_test = get_dataset(self.src_word2id, test_src_path, False, True, False)
self.test_dataset = tf.data.Dataset.from_generator(lambda: test_src, tf.int32)
self.test_dataset = self.test_dataset.cache().batch(1).prefetch(1)
self.test_result_file = self.log_folder / "test_result.txt"
elif self.test_mode :
raise FileNotFoundError("[ {} ] 경로가 존재하지 않습니다.".format(test_src_path))
self.transformer = Transformer(self.src_vocab_size, self.tar_vocab_size, self.src_word2id["<PAD>"],
kwargs["num_layers"], kwargs["heads"], kwargs["embedding_size"], kwargs["hidden_size"],
kwargs["dropout_rate"], kwargs["use_conv"])
def __init__(self, **kwargs):
dataset_folder = Path(kwargs["dataset_folder"]).resolve()
check_valid_path(dataset_folder)
result_folder = kwargs["result_folder"]
self.initial_epoch = 1
self.test_mode = kwargs["test"]
self.epochs = kwargs["epochs"]
self.use_label_smoothing = kwargs["label_smoothing"]
self.ckpt_path = kwargs["ckpt_path"]
self.ckpt_epoch = kwargs["ckpt_epoch"]
# model에 필요한 폴더 및 파일 생성
self.log_folder, self.ckpt_folder, self.image_folder = create_folder(
result_folder)
if not self.test_mode:
self.training_result_file = self.log_folder / "training_result.txt"
self.test_result_file = None
# kwargs 값 저장
msg = ""
for k, v in list(kwargs.items()):
msg += "{} = {}\n".format(k, v)
msg += "new model checkpoint path = {}\n".format(self.ckpt_folder)
with (self.log_folder / "model_settings.txt").open(
"w", encoding="utf-8") as fp:
fp.write(msg)
# 필요한 data를 불러옴
self.src_word2id, self.src_id2word, self.src_vocab_size = load_word_dic(
dataset_folder / "src_word2id.pkl")
self.tar_word2id, self.tar_id2word, self.tar_vocab_size = load_word_dic(
dataset_folder / "tar_word2id.pkl")
if not self.test_mode:
train_src, num_train_src = get_dataset(
self.src_word2id, dataset_folder / "train_src.txt", False,
True, True)
train_tar, num_train_tar = get_dataset(
self.tar_word2id, dataset_folder / "train_tar.txt", True, True,
True)
if num_train_src != num_train_tar:
raise Exception(
"source 데이터셋({})과 target 데이터셋({})의 크기가 다릅니다.".format(
num_train_src, num_train_tar))
self.num_train = num_train_src
self.train_dataset = tf.data.Dataset.from_generator(
lambda: zip(train_src, train_tar), (tf.int32, tf.int32))
self.train_dataset = self.train_dataset.cache().shuffle(
self.num_train + 1).padded_batch(
batch_size=kwargs["batch_size"],
padded_shapes=(tf.TensorShape([None]),
tf.TensorShape([None])),
padding_values=(self.src_word2id["<PAD>"],
self.tar_word2id["<PAD>"])).prefetch(1)
test_src_path = dataset_folder / "test.txt"
if test_src_path.exists():
test_src, self.num_test = get_dataset(self.src_word2id,
test_src_path, False, True,
False)
# self.test_src_max_len = max([len(sentence) for sentence in test_src])
# padded_test_src = tf.keras.preprocessing.sequence.pad_sequences(
# test_src, maxlen = self.test_src_max_len, padding = 'post',
# dtype = 'int32', value = self.src_word2id["<PAD>"])
self.test_dataset = tf.data.Dataset.from_generator(
lambda: test_src, tf.int32)
self.test_dataset = self.test_dataset.cache().batch(1).prefetch(1)
self.test_result_file = self.log_folder / "test_result.txt"
elif self.test_mode:
raise FileNotFoundError(
"[ {} ] 경로가 존재하지 않습니다.".format(test_src_path))
self.encoder = Encoder(self.src_vocab_size, kwargs["embedding_size"],
kwargs["hidden_size"], kwargs["dropout_rate"],
kwargs["gru"], kwargs["bi"])
self.decoder = Decoder(self.tar_vocab_size, kwargs["embedding_size"],
kwargs["hidden_size"], kwargs["attention_size"],
kwargs["dropout_rate"], kwargs["gru"],
kwargs["bi"])
# 아래 line 6줄은 colab에서 한글 깨짐을 방지하기 위한 부분으로 생략해도 됩니다.
# %config InlineBackend.figure_format = 'retina'
# !apt -qq -y install fonts-nanum
fontpath = '/usr/share/fonts/truetype/nanum/NanumBarunGothic.ttf'
font = fm.FontProperties(fname=fontpath, size=9)
plt.rc('font', family='NanumBarunGothic')
mpl.font_manager._rebuild()
def train_model(model,
trgen,
valgen,
prefix,
epochs=1,
verbose=1,
loss='binary_crossentropy',
history_folder='./history',
model_folder='./models',
optimizer=SGD(lr=0.02,
decay=1e-6,
momentum=0.9,
nesterov=True,
clipnorm=5),
steps_per_epoch=100,
validation_steps=10,
metrics=['acc'],
save_period=0,
sample_weight_mode=None,
report_stats=False,
class_names=[]):
model_path = os.path.join(model_folder,
prefix + '.{epoch:02d}-{val_loss:.2f}.hdf5')
pickle_path = os.path.join(history_folder, prefix + '.pkl')
create_folder(model_folder)
create_folder(history_folder)
callbacks = []
if save_period > 0:
callbacks.append(
ModelCheckpoint(filepath=model_path, period=save_period,
verbose=1))
if loss == 'ctc':
ctcmodel = add_ctc(model)
ctcmodel.compile(loss={
'ctc': lambda y_true, y_pred: y_pred
},
optimizer=optimizer,
sample_weight_mode=sample_weight_mode)
hist = ctcmodel.fit_generator(generator=trgen,
validation_data=valgen,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
callbacks=callbacks,
verbose=verbose,
shuffle=False)
else:
model.compile(optimizer=optimizer,
loss=loss,
metrics=metrics,
sample_weight_mode=sample_weight_mode)
hist = model.fit_generator(generator=trgen,
validation_data=valgen,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
callbacks=callbacks,
verbose=verbose,
shuffle=False)
if report_stats:
cnf = get_model_class_stats(model, valgen, validation_steps,
class_names)
dump_data(hist.history, pickle_path)
return hist