def train(self, batch_size=50, epochs=100, weights_name="default_weights"):
self.X_img, self.X_control, self.y = shuffle(self.X_img,
self.X_control,
self.y,
random_state=0)
weights_path = self.model_path / "weights" / weights_name
try_make_dirs(weights_path)
logs_path = weights_path / "logs"
try_make_dirs(logs_path)
logs_path_str = str(logs_path.absolute())
tb_callback = keras.callbacks.TensorBoard(log_dir=logs_path_str,
histogram_freq=0,
write_graph=True,
write_images=True)
self.model.compile(loss="mean_squared_error",
optimizer=optimizers.adam(lr=0.001))
launch_tensorboard(logs_path_str)
self.model.fit([self.X_img, self.X_control],
self.y,
batch_size=batch_size,
epochs=epochs,
shuffle=False,
validation_split=0.2,
callbacks=[tb_callback])
self.model.save_weights(weights_path / "weights.h5")
with open(weights_path / "info.json", "w") as info_file:
json.dump(self.info, info_file)
def start_recording(self):
self.set_save_dir()
self.refresh_image = False
self.predict_button.setEnabled(False)
self.frame = 1
try_make_dirs(self.save_dir / "images")
try_make_dirs(self.save_dir / "key-events")
info_dict = {"key_labels": self.key_labels}
with open(self.save_dir / "info.json", "w") as fp:
json.dump(info_dict, fp)
self.recording = True
def process(self, data_folder, input_channels_mask=None, img_update_callback=None):
with open(data_folder / "info.json") as info_file:
data_info = json.load(info_file)
key_labels = np.asarray(data_info["key_labels"])[input_channels_mask]
self.info = {
"key_labels": key_labels.tolist()
}
self.X, self.y = self.stack_arrays(data_folder / "key-events",
data_folder / "images",
img_update_callback)
if input_channels_mask:
self.y = self.y[:, input_channels_mask]
save_path = self.model_path / "data" / data_folder.name
try_make_dirs(save_path)
np.save(save_path / "y", self.y)
np.save(save_path / "X", self.X)
with open(save_path / "info.json", "w") as info_file:
json.dump(self.info, info_file)