output_shape doesn't make sense, the output remains no structure

[CaptainDuke]

As mentioned before, I tried to read hdf5 data using tf-io, but the output_shape doesn't make sense. ([1,8]) expected but returns ( ).

For comparing, I worte dataset2 form tf.data.Dataset.from_tensor_slices(). Also I did some tf.reshape ops, where dataset read via tf-io failed.

I thinks it's a kind of serious problem since we can not continue tf.layers.conv2d or other ops without output_shapes.

Hope this issue would be fixed. Thanks.

[CaptainDuke]

Code above
issue_test.py.zip

[CaptainDuke]

temporary solution:
add dataset.batch() and then tf.reshape

dataset = hdf5_io.HDF5Dataset(
       [filename], columns, output_types, output_shapes)
dataset = dataset.batch(10)
iterator = dataset.make_one_shot_iterator()
get_next = iterator.get_next()

input_x = tf.reshape(get_next, output_shape)
sess = tf.Session()
vv = sess.run(tf.shape(input_x))
print(vv)

Besides, if we want to use dataset.shuffle and dataset.prefetch then, dataset.make_one_shot_iterator() works.

[yongtang]

@CaptainDuke in 2.0 the output shape and dtype are implicitly probed. I believe the issue has been resolved. Will close the issue for now but feel free to reopen if issue still persists.