def pred_candlestick(
self,
model,
output_png=None,
classes=["0", "1"],
img_size=[80, 80, 3],
):
"""ローソク足の画像作成して予測+gradcam"""
# ラベルと株価取得
df, label_price = self.get_pred_df()
# 画像出力
if self.is_2day_label or len(classes) == 4:
make_candlestick.make_mav_candlestick_mplfinance(df, output_png)
else:
make_candlestick.make_candlestick_mplfinance(df, output_png)
# 予測
y, pb = base_predict.pred_from_1img(model,
output_png,
img_size[0],
img_size[1],
classes=classes,
show_img=True)
# gradcam
_ = grad_cam.image2gradcam(model, output_png)
return y, pb, label_price, df
if p_args.mode == "predict":
args = get_class_fine_tuning_parameter_base()
pred_directory(args)
if p_args.grad_cam_model_path is not None and p_args.grad_cam_image_dir is not None:
args = get_class_fine_tuning_parameter_base()
for i, p in tqdm(
enumerate(util.find_img_files(p_args.grad_cam_image_dir))):
# 50枚ごとにモデル再ロード
if i % 50 == 0:
keras.backend.clear_session()
keras.backend.set_learning_phase(0)
model = keras.models.load_model(p_args.grad_cam_model_path,
compile=False)
# p_args.grad_cam_image_dirと同じディレクトリにGradCAM画像出力
grad_cam.image2gradcam(model, p, is_gradcam_plus=False)
if p_args.mode == "tuning":
os.makedirs(p_args.tuning_output_dir, exist_ok=True)
study = optuna.create_study(
direction="minimize",
study_name=p_args.study_name,
storage=
f"sqlite:///{p_args.tuning_output_dir}/{p_args.study_name}.db",
load_if_exists=True,
)
study.optimize(
Objective(p_args.tuning_output_dir),
n_trials=p_args.n_trials,
timeout=60 * 30, # 1trial30分
)