DataFeeder constructor 버그

[devming]

System 정보

• centos 7.0
• python 3.4
• tensorflow 1.2.1

문제 사항

tensorflow.contrib.learn.DNNRegressor.fit(x_train_dict , y_train,steps=1000)

위와 같은 코드를 수행했을 때, (여기서 x_train_dict는 dict 타입이고, y_train는 array타입입니다.)
프로그램이 다음과 같은 Exception을 발생시켰습니다.

File "/home/star/yuce.ddxq.mobi/zhuge/management/commands/forecast_product_sale.py", line 148, in tflearn_dnn_train2
    regressor.fit(x_train_dict, y_train, steps=10000, batch_size=10)
  File "/usr/lib/python3.4/site-packages/tensorflow/python/util/deprecation.py", line 289, in new_func
    return func(*args, **kwargs)
  File "/usr/lib/python3.4/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 439, in fit
    SKCompat(self).fit(x, y, batch_size, steps, max_steps, monitors)
  File "/usr/lib/python3.4/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1340, in fit
    epochs=None)
  File "/usr/lib/python3.4/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 137, in _get_input_fn
    epochs=epochs)
  File "/usr/lib/python3.4/site-packages/tensorflow/contrib/learn/python/learn/learn_io/data_feeder.py", line 152, in setup_train_data_feeder
    x, y, n_classes, batch_size, shuffle=shuffle, epochs=epochs)
  File "/usr/lib/python3.4/site-packages/tensorflow/contrib/learn/python/learn/learn_io/data_feeder.py", line 326, in __init__
    dict([(k, check_array(v, v.dtype)) for k, v in list(y.items())]) if x_is_dict else check_array(y, y.dtype)
AttributeError: 'numpy.ndarray' object has no attribute 'items'

관련된 코드는 다음과 같습니다.

   x_is_dict, y_is_dict = isinstance(x, dict), y is not None and isinstance(
        y, dict)
    if isinstance(y, list):
      y = np.array(y)

    self._x = dict([(k, check_array(v, v.dtype)) for k, v in list(x.items())
                   ]) if x_is_dict else check_array(x, x.dtype)
    self._y = None if y is None else \
      dict([(k, check_array(v, v.dtype)) for k, v in list(y.items())]) if x_is_dict else check_array(y, y.dtype)

코드 마지막줄에 잘못된 것으로 보입니다. 여기서 x_is_dict 대신 y_is_dict를 사용해야하는 것 아닌가요?

바뀐 코드는 다음과 같습니다. :

dict([(k, check_array(v, v.dtype)) for k, v in list(y.items())]) if y_is_dict else check_array(y, y.dtype)

이렇게 하니 동작했습니다.