def train_directory(args):
"""指定ディレクトリについてgenerator作ってモデル学習"""
print("train_directory")
# ### train validation data load ### #
d_cls = get_train_valid_test.LabeledDataset(
[args["img_rows"], args["img_cols"], args["channels"]],
args["batch_size"],
valid_batch_size=args["batch_size"],
train_samples=len(util.find_img_files(args["train_data_dir"])),
valid_samples=len(util.find_img_files(args["validation_data_dir"])),
)
if args["is_flow"]:
# 指定ディレクトリの前処理済み画像、ラベル、ファイルパスロード
d_cls.X_train, d_cls.y_train, train_paths = base_dataset.load_my_data(
args["train_data_dir"],
classes=args["classes"],
img_height=args["img_rows"],
img_width=args["img_cols"],
channel=args["channels"],
is_pytorch=False,
)
d_cls.X_valid, d_cls.y_valid, valid_paths = base_dataset.load_my_data(
args["validation_data_dir"],
classes=args["classes"],
img_height=args["img_rows"],
img_width=args["img_cols"],
channel=args["channels"],
is_pytorch=False,
)
d_cls.X_train, d_cls.X_valid = d_cls.X_train * 255.0, d_cls.X_valid * 255.0
d_cls.create_my_generator_flow(my_IDG_options=args["my_IDG_options"])
elif args["is_flow_from_directory"]:
d_cls.create_my_generator_flow_from_directory(
args["train_data_dir"],
args["classes"],
valid_data_dir=args["validation_data_dir"],
color_mode=args["color_mode"],
class_mode=args["class_mode"],
my_IDG_options=args["my_IDG_options"],
)
# d_cls.train_gen_augmentor = d_cls.create_augmentor_util_from_directory(args['train_data_dir']
# , args['batch_size']
# , augmentor_options=args['train_augmentor_options'])
# binaryラベルのgeneratorをマルチタスクgeneratorに変換するラッパー
if args["n_multitask"] > 1 and args["multitask_pred_n_node"] == 1:
d_cls.train_gen = get_train_valid_test.binary_generator_multi_output_wrapper(
d_cls.train_gen
)
d_cls.valid_gen = get_train_valid_test.binary_generator_multi_output_wrapper(
d_cls.valid_gen
)
# ### model ### #
if args["model_path"] is not None:
print(f"INFO: load model: {args['model_path']}")
model = keras.models.load_model(args["model_path"], compile=False)
else:
os.makedirs(args["output_dir"], exist_ok=True)
if args["choice_model"] == "model_paper":
model = model_paper.create_paper_cnn(
input_shape=(args["img_cols"], args["img_rows"], args["channels"]),
num_classes=args["num_classes"],
activation=args["activation"],
)
else:
model, orig_model = define_model.get_fine_tuning_model(
args["output_dir"],
args["img_rows"],
args["img_cols"],
args["channels"],
args["num_classes"],
args["choice_model"],
trainable=args["trainable"],
fcpool=args["fcpool"],
fcs=args["fcs"],
drop=args["drop"],
activation=args["activation"],
weights=args["weights"],
)
optim = define_model.get_optimizers(
choice_optim=args["choice_optim"], lr=args["lr"], decay=args["decay"]
)
model.compile(loss=args["loss"], optimizer=optim, metrics=args["metrics"])
cb = my_callback.get_base_cb(
args["output_dir"],
args["num_epoch"],
early_stopping=args["num_epoch"] // 4,
monitor="val_" + args["metrics"][0],
metric=args["metrics"][0],
)
# lr_finder
if args["is_lr_finder"] == True:
# 最適な学習率確認して関数抜ける
lr_finder.run(
model,
d_cls.train_gen,
args["batch_size"],
d_cls.init_train_steps_per_epoch,
output_dir=args["output_dir"],
)
return
# 各クラスのlossを加重平均にする場合
class_weight = {c: 1.0 for c in args["classes"]}
if args["is_class_weight"]:
class_weight = my_class_weight.cal_weight(
args["classes"], args["train_data_dir"]
)
# ### train ### #
start_time = time.time()
hist = model.fit(
d_cls.train_gen,
steps_per_epoch=d_cls.init_train_steps_per_epoch,
epochs=args["num_epoch"],
validation_data=d_cls.valid_gen,
validation_steps=d_cls.init_valid_steps_per_epoch,
verbose=1, # 1:ログをプログレスバーで標準出力 2:最低限の情報のみ出す
callbacks=cb,
class_weight=class_weight,
)
end_time = time.time()
print("Elapsed Time : {:.2f}sec".format(end_time - start_time))
model.save(os.path.join(args["output_dir"], "model_last_epoch.h5"))
plot_log.plot_results(
args["output_dir"],
os.path.join(args["output_dir"], "tsv_logger.tsv"),
acc_metric=args["metrics"][0],
)
return hist
def trial_train_directory(self, trial, args):
keras.backend.clear_session()
# ### train validation data load ### #
d_cls = get_train_valid_test.LabeledDataset(
[args["img_rows"], args["img_cols"], args["channels"]],
args["batch_size"],
valid_batch_size=args["batch_size"],
train_samples=len(util.find_img_files(args["train_data_dir"])),
valid_samples=len(util.find_img_files(args["validation_data_dir"])),
)
if args["is_flow"]:
# 指定ディレクトリの前処理済み画像、ラベル、ファイルパスロード
d_cls.X_train, d_cls.y_train, train_paths = base_dataset.load_my_data(
args["train_data_dir"],
classes=args["classes"],
img_height=args["img_rows"],
img_width=args["img_cols"],
channel=args["channels"],
is_pytorch=False,
)
d_cls.X_valid, d_cls.y_valid, valid_paths = base_dataset.load_my_data(
args["validation_data_dir"],
classes=args["classes"],
img_height=args["img_rows"],
img_width=args["img_cols"],
channel=args["channels"],
is_pytorch=False,
)
d_cls.X_train, d_cls.X_valid = d_cls.X_train * 255.0, d_cls.X_valid * 255.0
d_cls.create_my_generator_flow(my_IDG_options=args["my_IDG_options"])
elif args["is_flow_from_directory"]:
d_cls.create_my_generator_flow_from_directory(
args["train_data_dir"],
args["classes"],
valid_data_dir=args["validation_data_dir"],
color_mode=args["color_mode"],
class_mode=args["class_mode"],
my_IDG_options=args["my_IDG_options"],
)
# d_cls.train_gen_augmentor = d_cls.create_augmentor_util_from_directory(args['train_data_dir']
# , args['batch_size']
# , augmentor_options=args['train_augmentor_options'])
# binaryラベルのgeneratorをマルチタスクgeneratorに変換するラッパー
if args["n_multitask"] > 1 and args["multitask_pred_n_node"] == 1:
d_cls.train_gen = get_train_valid_test.binary_generator_multi_output_wrapper(
d_cls.train_gen
)
d_cls.valid_gen = get_train_valid_test.binary_generator_multi_output_wrapper(
d_cls.valid_gen
)
# ### model ### #
if args["model_path"] is not None:
model = keras.models.load_model(args["model_path"], compile=False)
else:
os.makedirs(args["output_dir"], exist_ok=True)
if args["choice_model"] == "model_paper":
model = model_paper.create_paper_cnn(
input_shape=(args["img_cols"], args["img_rows"], args["channels"]),
num_classes=args["num_classes"],
activation=args["activation"],
)
else:
model, orig_model = define_model.get_fine_tuning_model(
args["output_dir"],
args["img_rows"],
args["img_cols"],
args["channels"],
args["num_classes"],
args["choice_model"],
trainable=args["trainable"],
fcpool=args["fcpool"],
fcs=args["fcs"],
drop=args["drop"],
activation=args["activation"],
weights=args["weights"],
)
optim = define_model.get_optimizers(
choice_optim=args["choice_optim"], lr=args["lr"], decay=args["decay"]
)
model.compile(loss=args["loss"], optimizer=optim, metrics=args["metrics"])
cb = my_callback.get_base_cb(
args["output_dir"],
args["num_epoch"],
early_stopping=20,
monitor="val_" + args["metrics"][0],
metric=args["metrics"][0],
) # args['num_epoch']//3
cb.append(OptunaCallback(trial, True))
# 各クラスのlossを加重平均にする場合
class_weight = {c: 1.0 for c in args["classes"]}
if args["is_class_weight"]:
class_weight = my_class_weight.cal_weight(
args["classes"], args["train_data_dir"]
)
# ### train ### #
hist = model.fit(
d_cls.train_gen,
steps_per_epoch=d_cls.init_train_steps_per_epoch,
epochs=args["num_epoch"],
validation_data=d_cls.valid_gen,
validation_steps=d_cls.init_valid_steps_per_epoch,
verbose=2, # 1:ログをプログレスバーで標準出力 2:最低限の情報のみ出す
callbacks=cb,
class_weight=class_weight,
)
return hist