Can't set num_steps to None in Trainer component about tfx HOT 5 CLOSED

That's a bug. I'll work on a fix (second in my queue).

from tfx.

Hi Loic -- what are you using as the stop condition for your estimator? I'm curious how a "runs forever" Trainer configuration behaves in airflow/kubeflow. If you're raising an error after n passes through the data, the error should abort the remainder of the pipeline execution as well.

from tfx.

Well I'm not raising an error, I'm simply defining the num_epoch in my inputs with the Estimator API (I can add precisions tomorrow), and the training stops exactly after num_epoch epochs.

from tfx.

To confirm, your TrainSpec looks something like

  train_spec = tf.estimator.TrainSpec( 
      train_input_fn,
      # max_steps=hparams.train_steps,
      max_steps=None,
      num_epochs=1)

?

from tfx.

The configuration is done in the input_fn, not in the TrainSpec:
train_input_fn = lambda: utils.input_fn(hparams.train_files, hparams.transform_output, batch_size=train_batch_size, num_epochs=1)

And this call to read_batch_features in my utils.input_fn:
tf.contrib.learn.io.read_batch_features(filenames, batch_size, transformed_feature_spec, reader=_gzip_reader_fn, num_epochs=num_epochs)

A deprecated function, but it is used in the Taxi example.

from tfx.