def run_episode(self,sess):
# Get feed_dict
training_set = dataset.DataGenerator()
coord_batch, dist_batch, input_batch = training_set.next_batch(self.batch_size, self.max_length, self.input_dimension)
feed = {self.input_coordinates: coord_batch, self.input_description: input_batch}
# Actor Forward pass
seq_input, permutation, seq_proba = sess.run([self.input_coordinates,self.positions,self.proba],feed_dict=feed)
# Critic Forward pass
b_s = sess.run(self.prediction_c,feed_dict=feed)
# Environment response
trip, circuit_length, reward = sess.run([self.trip,self.distances,self.reward], feed_dict=feed)
# Train step
if self.step==0:
loss1, train_step1 = sess.run([self.loss1,self.train_step1],feed_dict=feed)
else:
loss1, train_step1, loss2, train_step2= sess.run([self.loss1,self.train_step1,self.loss2,self.train_step2],feed_dict=feed)
self.step+=1
if self.step%100==0:
self.ptr.temperature*=self.temperature_decay
return seq_input, permutation, seq_proba, b_s, trip, circuit_length, reward
def prepare_db(opt):
print("Use %s dataset" % (opt.dataset))
if opt.dataset == 'cold':
dir_name = "/train_std.txt"
train_df = pd.read_csv(os.getcwd() + dir_name, header=None)
train_df = train_df.iloc[1:, :]
train_dataset = dataset.DataGenerator(train_df)
dir_name = "/val_std.txt"
eval_df = pd.read_csv(os.getcwd() + dir_name, header=None)
eval_df = eval_df.iloc[1:, :]
print(len(eval_df))
eval_dataset = dataset.DataGenerator(eval_df)
dir_name = "/test_std.txt"
test_df = pd.read_csv(os.getcwd() + dir_name, header=None)
test_df = test_df.iloc[1:, :]
print(len(test_df))
test_dataset = dataset.DataGenerator(test_df)
return {'train': train_dataset, 'eval': eval_dataset,'test':test_dataset}
else:
raise NotImplementedError
logger.info("load weight:%s", args.weights_path)
model.load_weights(args.weights_path, by_name=True)
logger.info(model.summary())
# ネットワーク構成を画像として保存
utils.plot_model(model, './model.png', True, True)
# レイヤの重み更新有無を確認
for i, layer in enumerate(model.layers):
if layer.__class__.__name__ == 'TimeDistributed':
name = layer.layer.name
trainable = layer.layer.trainable
else:
name = layer.name
trainable = layer.trainable
logger.info('%s %s:%s', i, name, trainable)
# 学習、検証データ生成器の準備
gen = dataset.DataGenerator(config)
train_data_generator = gen.generate(os.path.join(args.data_dir, "train"),
train_generator, train_discriminator)
val_data_generator = gen.generate(os.path.join(args.data_dir, "val"),
train_generator, train_discriminator)
model_file_path = './nnmodel/glcic-stage{}-{}'.format(
args.stage, '{epoch:02d}-{val_loss:.2f}.h5')
callbacks = [
keras.callbacks.TerminateOnNaN(),
keras.callbacks.TensorBoard(log_dir='./tb_log',
histogram_freq=0,
write_graph=True,
write_images=False),
# keras.callbacks.ReduceLROnPlateau(monitor='val_loss',
# verbose=1,
# factor=0.7,
os.path.join(main_res_path, "models", "best.hdf5"))[0]
model = keras.models.load_model(best_model_file)
sess = keras.backend.get_session()
save_images_path = ''
if save_images:
save_images_path = os.path.join(best_model_res_path, 'sample_images')
os.mkdir(save_images_path)
transfer = True if model_type == 'vgg_16_transfer' else False
num_channels = 3 if transfer else 1
input_shape = (256, 256, num_channels)
val_g = ds.DataGenerator(data_partition_path=os.path.join(
data_partition_dir, 'val.npy'),
data_source_dir=data_source_dir,
dim=input_shape,
save_images=save_images,
save_images_path=save_images_path)
val_preds = model.predict_generator(val_g)
np.savetxt(os.path.join(best_model_res_path, 'val_preds.csv'), val_preds)
# note shuffling is disabled, batch size is 1 so that data yielded from this generator
# follows the order of the data in test.npy
test_g = ds.DataGenerator(data_partition_path=os.path.join(
data_partition_dir, 'test.npy'),
data_source_dir=data_source_dir,
dim=input_shape,
save_images=save_images,
save_images_path=save_images_path)
model.compile(loss='binary_crossentropy',
optimizer=optimizers.SGD(lr=lr,
decay=decay,
momentum=0.9,
nesterov=False),
metrics=['accuracy'])
print("compiled model")
sys.stdout.flush()
# prepare the data generators
train_g = ds.DataGenerator(data_partition_path=os.path.join(
data_partition_dir, 'train.npy'),
data_source_dir=data_source_dir,
batch_size=batch_size,
dim=input_shape,
shuffle=True,
augmentation_flag=enable_augmentations,
save_images=save_images,
save_images_path=save_images_path)
val_g = ds.DataGenerator(data_partition_path=os.path.join(
data_partition_dir, 'val.npy'),
data_source_dir=data_source_dir,
dim=input_shape,
save_images=save_images,
save_images_path=save_images_path)
# training callbacks
csv_logger = CSVLogger(os.path.join(final_model_res_path, 'training.log'))
# tensorboard_callback=keras.callbacks.TensorBoard()