Fix for #7225: CIFAR-10 eval fails with error TypeError: Input 'predictions' of 'InTopKV2' Op has type float16 that contributed by mattmann.

[chrismattmann]

No description provided.

[chrismattmann]

@sherrym and @shlens can you take a look? Thank you!

[tfboyd]

LGTM.

I did not test this; but I am going to trust the community. It seems silly or me to block a PR, that seems legit and looks to address the issue in the github issue.

Thank you for the contribution.

[puneetjindal]

@tfboyd I am getting the same error when i am trying to train with TPU v2-8 with my GCE VM 16.04 LTS. My error and code respectively is as follows:

I0310 18:53:33.341825 139964930656000 resnet50.py:195] Compiling model.
Traceback (most recent call last):
File "/home/puneet_jindal/.local/lib/python3.5/site-packages/tensorflow_core/python/framework/op_def_library.py", line 468, in _a
pply_op_helper
preferred_dtype=default_dtype)
File "/home/puneet_jindal/.local/lib/python3.5/site-packages/tensorflow_core/python/framework/ops.py", line 1290, in convert_to_t
ensor
(dtype.name, value.dtype.name, value))
ValueError: Tensor conversion requested dtype float32 for Tensor with dtype int32: <tf.Tensor 'metrics/sparse_top_k_categorical_acc
uracy/Cast:0' shape=(None,) dtype=int32>

`def sparse_top_k_categorical_accuracy(y_true, y_pred, k=5):
"""TPU version of sparse_top_k_categorical_accuracy."""
y_pred_rank = tensorflow.convert_to_tensor(y_pred).get_shape().ndims
y_true_rank = tensorflow.convert_to_tensor(y_true).get_shape().ndims

if ((y_true_rank is not None) and
(y_pred_rank is not None) and
(len(tensorflow.keras.backend.int_shape(y_true)) == len(tensorflow.keras.backend.int_shape(y_pred)))):
y_true = tensorflow.squeeze(y_true, [-1])

y_true = tensorflow.cast(y_true, 'int32')
#y_pred = tensorflow.cast(y_pred, 'float32')
return tensorflow.nn.in_top_k(y_pred, y_true, k)
`
This code is taken from https://github.com/tensorflow/tpu/blob/master/models/experimental/resnet50_keras/resnet50.py