tensorflow: Cannot train canned estimators in multiple estimator.train() calls when using tf.keras.optimizers or tf.optimizers

@JoshEZiegler Right. My use case was a bit different and I missed that. But since we can pass any callable, then maybe something like this will work for you:

from functools import partial
AdamWithParams = partial(Adam, learning_rate = 0.1)

...

optimizer = AdamWithParams

More on what partial is you can find in the docs. In my code it worked as expected.

Instead of partial, use lambda to create optimizer object each time. It’s way more easier in this way.

linear_regressor = tf.estimator.LinearRegressor(
    feature_columns=feature_columns,
    optimizer=lambda:tf.keras.optimizers.SGD(learning_rate=0.0000001, clipnorm=5.0),
)

@yhliang2018 This issue still shows up with tf-nightly-2.2.0.dev20200306. See this colab notebook.

Was there a specific version where you believe it should work?

@awolant Nice! Thanks, that sounds like the perfect workaround.

I’ll leave this issue open because I don’t believe that is the intended way to use optimizers with Estimators. At the very least not the way it is done in the TF docs.

I tried running the linked code and I confirm I also see an error:

RuntimeError: Cannot set iterations to a new Variable after the Optimizer weights have been created

With some investigation, it looks like running estimator.train() multiple times is ok with the default optimizer, or by specifying the optimizer with a string. I believe that this is because the estimator.train() actually creates a new instance of the string-specified optimizer with each call, but retains the optimizer object if one was specified in the train call (see below).

A possible workaround for the above error could be to modify this function to return a fresh optimizer with the same parameters as the opt instance specified. However, I’m not sure if this is easily achievable or if there are any use cases that this would break…

From /tensorflow/estimator/blob/master/tensorflow_estimator/python/estimator/canned/optimizers.py

def get_optimizer_instance(opt, learning_rate=None):

if isinstance(opt, six.string_types):

    if opt in six.iterkeys(_OPTIMIZER_CLS_NAMES):
      if not learning_rate:
        raise ValueError('learning_rate must be specified when opt is string.')
      return _OPTIMIZER_CLS_NAMES[opt](learning_rate=learning_rate)
    raise ValueError(
        'Unsupported optimizer name: {}. Supported names are: {}'.format(
            opt, tuple(sorted(six.iterkeys(_OPTIMIZER_CLS_NAMES)))))
  if callable(opt):
    opt = opt()
  if not isinstance(opt, optimizer_lib.Optimizer):
    raise ValueError(
        'The given object is not an Optimizer instance. Given: {}'.format(opt))
  return opt