tf.summary.histogram fails with TypeError: __init__() got an unexpected keyword argument 'values'

[Daniel451]

• TensorBoard version (from pip package, also printed out when running tensorboard)
  -- TensorBoard 1.14.0a20190301
• TensorFlow version if different from TensorBoard
  -- TensorFlow-gpu 2.0-alpha
• OS Platform and version (e.g., Linux Ubuntu 16.04)
  -- Linux, Arch, Kernel 4.19
• Python version (e.g. 2.7, 3.5)
  -- Python 3.6.7

I have built a GAN with TF 2.0-alpha using tf.keras. I am using TensorBoard via train_writer = tf.summary.create_file_writer("log-train"). Adding scalars with tf.summary.scalar("loss g", loss_g, step=batch_id) and storing predicted images with tf.summary.image("g pred", prediction[0:3], step=batch_id) works without any problems.

However, when I try to add the generator's weights to TensorBoard with tf.summary.histogram it fails.

for layer in G_model.layers:
    for weight in layer.weights:
        tf.summary.histogram(weight.name, weight, step=batch_id)

G_model is a tf.keras model created with the functional API.

The exception seems to be TensorBoard-specific:

Traceback (most recent call last):
  File "train.py", line 84, in <module>
    tf.summary.histogram(weight.name, weight, step=batch_id)
  File "(...)/python3.6/site-packages/tensorboard/plugins/histogram/summary_v2.py", line 73, in histogram
    tensor = _buckets(data, bucket_count=buckets)
  File "(...)/python3.6/site-packages/tensorboard/plugins/histogram/summary_v2.py", line 91, in _buckets
    with tf.name_scope('buckets', values=[data, bucket_count]):
TypeError: __init__() got an unexpected keyword argument 'values'

If I simply delete , values=[data, bucket_count] in summary_v2.py (line 91) it works, weights are showing up correctly in TensorBoard then.

Thus, original code of summary_v2.py in line 91:

with tf.name_scope('buckets', values=[data, bucket_count]):

throws the exception posted above, while this code:

with tf.name_scope('buckets'):

is working properly.

[wchargin]

Hi @Daniel451! Thanks for the report; I can reproduce this.

This should be fixed by #1933, which did not make it into the
1.14.0a20190301 version of TensorBoard bundled with the TF 2.0 alpha,
but did make it into the next day’s 1.14.0a20190302.

We can push a patch release 1.14.0a20190301.post001 that would; cc
@nfelt @stephanwlee as to whether we want to do this, noting that we
also considered cherry-picking #1945.

[nfelt]

@Daniel451 just to be clear, for now a workaround is to pip install --upgrade tb-nightly which will get you a version of the nightly TensorBoard build that includes the fix for this in #1933.

[Harshini-Gadige]

Closing this issue as this should be resolved with tb-nightly.