def test_train(self):
model = NumberRecognizeNN(Resource.INPUT_SIZE, Resource.OUTPUT_SIZE)
r = Resource(self.TEST_DIR)
trainer = Trainer(model, r)
dp = DataProcessor()
data, target = r.load_training_data()
print("Test Train the model")
trainer.train(data, target, epoch=5)
def test_model_api(self):
model = NumberRecognizeNN(Resource.INPUT_SIZE, Resource.OUTPUT_SIZE)
r = Resource(self.TEST_DIR)
trainer = Trainer(model, r)
dp = DataProcessor()
data, target = r.load_training_data()
api_test_size = 200
print("Train the model for API Test.")
trainer.train(data[:-api_test_size], target[:-api_test_size], epoch=5)
model_api = ModelAPI(r)
predicted = model_api.predict(data[-api_test_size:])
teacher = target[-api_test_size:]
score = accuracy_score(teacher, predicted)
print("Model API score is {}".format(score))
def post(self):
resp = {"result": str(-1)}
data = self.get_arguments("data[]")
r = Resource()
if not os.path.isdir(r.model_path):
from ml.model import NumberRecognizeNN
from ml.trainer import Trainer
model = NumberRecognizeNN(r.INPUT_SIZE, r.OUTPUT_SIZE)
trainer = Trainer(model, r)
x, y = r.load_training_data()
trainer.train(x, y)
api = ModelAPI(r)
if len(data) > 0:
_data = [float(d) for d in data]
predicted = api.predict(_data)
resp["result"] = str(predicted[0])
self.write(resp)
def train(data_file, batch_size, epoch, test_size):
r = Resource()
dp = DataProcessor()
model = NumberRecognizeNN(Resource.INPUT_SIZE, Resource.OUTPUT_SIZE)
try:
dp.means, dp.stds = r.load_normalization_params()
r.load_model(model)
print("load the model")
except Exception as ex:
print("trained model does not exist.")
x = None
y = None
if data_file:
x, y = r.load_data(data_file)
else:
x, y = r.load_training_data()
trainer = Trainer(model, r)
print("begin training")
trainer.train(x, y, batch_size=batch_size, epoch=epoch, test_size=test_size)
def run(logdir, model_name, opt_name, loss_name):
logger.info(f'model_name: {model_name}')
logger.info(f'opt_name: {opt_name}')
model = get_model(model_name)(
input_shape=(9, 9, 43),
classes=MOVE_DIRECTION_LABEL_NUM,
activation_name=FLAGS.activation_name)
logger.info(model.summary())
if FLAGS.saved_model_path:
model = tf.keras.models.load_model(FLAGS.saved_model_path)
optimizer = get_optimizer(opt_name)
loss = get_loss(loss_name)
trainer = Trainer(model, logdir, optimizer, loss,
factor=FLAGS.reduce_factor,
patience=FLAGS.reduce_patience,
min_lr=FLAGS.reduce_factor)
with trainer.summary_writer.as_default():
tf.summary.text('parameters', FLAGS.flags_into_string(), step=0)
dataset = SFEN(FLAGS.train_csv, FLAGS.valid_csv,
FLAGS.epochs, FLAGS.batch_size)
# 3143460
if FLAGS.logging_steps:
logging_steps = FLAGS.logging_steps
else:
logging_steps = dataset.train_size // FLAGS.batch_size
logger.info(f'logging_steps: {logging_steps}')
trainer.train(dataset, logging_steps, FLAGS.learning_rate)
trainer.save()
def post(self):
# レスポンス用マップを作成
resp = {"result": str(-1)}
# 送られてきたデータを受け取る
data = self.get_arguments("data[]")
r = Resource()
# path が実在するディレクトリの場合
if not os.path.isdir(r.model_path):
# model.pyのインポート
from ml.model import NumberRecognizeNN
# trainer.pyのインポート
from ml.trainer import Trainer
# モデル作成クラス (Chainer)をnew
model = NumberRecognizeNN(r.INPUT_SIZE, r.OUTPUT_SIZE)
# トレーニングクラスをnew
trainer = Trainer(model, r)
# トレーニングデータを取得
x, y = r.load_training_data()
# トレーニング実施
trainer.train(x, y)
# モデルをAPIに保存
api = ModelAPI(r)
if len(data) > 0:
_data = [float(d) for d in data]
# 送られてきたデータから予測値を算出
predicted = api.predict(_data)
# レスポンスにデータを格納
resp["result"] = str(predicted[0])
# データを送信
self.write(resp)