tf.Estimator.train throws as_list() is not defined on an unknown TensorShape

Question

I created a custom input_func and converted a keras model into tf.Estimator for training. However, it keeps throwing me error.

• Here is my model summary. I have attempted to set the Input layer with batch_shape=(16, 320, 320, 3) for testing but the problem still persits

  inputs  = Input(batch_shape=(16, 320, 320, 3), name='input_images')
  outputs = yolov2.predict(intputs)
  model   = Model(inputs, outputs)
  
  model.compile(optimizer= tf.keras.optimizers.Adam(lr=learning_rate),
                loss     = compute_loss)
  

• I used tf.keras.estimator.model_to_estimator for conversion. I also create a input_fn for training:

  def input_fun(images, labels, batch_size, shuffle=True):
       dataset = create_tfdataset(images, labels)
       dataset = dataset.shuffle().batch(batch_size)
       iterator = dataset.make_one_shot_iterator()
       images, labels = iterator.next()
  
       return {'input_images': images}, labels
  
  estimator = tf.keras.estimator.model_to_estimator(keras_model=model)
  estimator.train(input_fn = lambda: input_fn(images, labels, 32),
                  max_steps = 1000)
  

• And it throws me this error

  input_tensor = Input(tensor=x, name='input_wrapper_for_' + name)
  ...
  File "/home/dat/anaconda3/envs/webapp/lib/python2.7/site-packages/tensorflow/python/layers/base.py", line 1309, in __init__
  self._batch_input_shape = tuple(input_tensor.get_shape().as_list())
  
  "as_list() is not defined on an unknown TensorShape.")
   ValueError: as_list() is not defined on an unknown TensorShape.
  

Answer 1

I had the same problem. In input_fun, if you look at images in line "return {'input_images': images}, labels", you'll see that your tensor has no shape. You have to call set_shape for each image. Look at https://github.com/tensorflow/models/blob/master/official/resnet/imagenet_main.py, they call vgg_preprocessing.preprocess_image to set the shape