Tensorflow ValueError: Batch length of predictions should be same

Question

I'm trying to visualize the output of a convolutional autoencoder using TensorFlow Estimator API

input_fn = tf.estimator.inputs.numpy_input_fn(
    x={"x": train_data},
    y=None,
    batch_size=8,
    num_epochs=None,
    shuffle=True)

autoencoder = tf.estimator.Estimator(model_fn=autoencoder_model_fn, model_dir=model_dir)
tensors_to_log = {"loss": "loss"}
logging_hook = tf.train.LoggingTensorHook(tensors=tensors_to_log, every_n_iter=1000)

autoencoder.train(
    input_fn=input_fn,
    steps=50000,
    hooks=[logging_hook])

input_fn_predict = tf.estimator.inputs.numpy_input_fn(
    x={"x": train_data},
    y=None,
    batch_size=1,
    num_epochs=None,
    shuffle=False)
predictions = autoencoder.predict(input_fn=input_fn_predict)
predictions = [p['decoded_image'] for p in predictions]
print predictions[0].shape

I get the following error :

Traceback (most recent call last):
  File "AutoEncoder.py", line 164, in <module>
    main()
  File "AutoEncoder.py", line 157, in main
    predictions = [p['decoded_image'] for p in predictions]
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/estimator/estimator.py", line 425, in predict
    for i in range(self._extract_batch_length(preds_evaluated)):
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/estimator/estimator.py", line 592, in _extract_batch_length
    'different batch length then others.' % key)
ValueError: Batch length of predictions should be same. features has different batch length then others.

Can anybody see what I did wrong? As I understand it, my batch size during prediction is a constant, equal to 1... Thanks in advance!

Romain · Accepted Answer · 2018-04-26 09:20:34Z

2

I succeed in dealing this issue. I just updated tensorflow (I had the 1.6 version and you need the 1.7 to have different batch size)

To see your tensorflow version you just need in your batch :

>python
>>>import tensorflow as tf
>>>print(tf.__version__)

Then, when you have the 1.7 version (or more), you use inside predict the argument :

yield_single_examples=False

(It is by default on True, and you will have the same mistake).

In your code, it will be :

predictions = autoencoder.predict(input_fn=input_fn_predict, yield_single_examples=False)

answered Apr 26, 2018 at 9:20

Romain

412 bronze badges

Sign up to request clarification or add additional context in comments.

Collectives™ on Stack Overflow

Tensorflow ValueError: Batch length of predictions should be same

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related