How to use a decaying learning rate with an estimator in tensorflow?

Question

I am trying to use a LinearClassifier with a GradientDescentOptimizer with a decaying learning rate.

My code:

def main():
# load data
    features = np.load('data/feature_data.npz')
    tx = features['arr_0']
    y = features['arr_1']

## Prepare logistic regression
    n_point, n_feat = tx.shape

# Input functions
    def get_input_fn_from_numpy(tx, y, num_epochs=None, shuffle=True):
    # Preprocess data
        return tf.estimator.inputs.numpy_input_fn(
        x={"x":tx},
        y=y,
        num_epochs=num_epochs,
        shuffle=shuffle,
        batch_size=128
        )

    cols_label = "x"
    feature_cols = [tf.contrib.layers.real_valued_column(cols_label)]

    my_input_fn_train = get_input_fn_from_numpy(tx, y)

    model_dir = 'data/tmp/' + datetime.datetime.now().strftime("%m-%d_%H:%M:%S")
    global_step = tf.Variable(0, trainable=False)
    learning_rate=tf.train.exponential_decay(0.001*np.ones((20,1), dtype=np.float32), global_step, 10000, 0.95, staircase=False)
    regressor = tf.contrib.learn.LinearClassifier(feature_columns=feature_cols,
                                              model_dir=model_dir,
                                                  optimizer=tf.train.GradientDescentOptimizer(learning_rate=learning_rate))

    regressor.fit(input_fn=get_input_fn_from_numpy(tx_train, y_train), steps=100000)
    results = regressor.evaluate(input_fn=my_input_fn_test)

I get the error:

  File "training.py", line 71, in <module>
main()
  File "training.py", line 63, in main
regressor.fit(input_fn=get_input_fn_from_numpy(tx_train, y_train), steps=100000)
  File "/lib/python3.6/site-packages/tensorflow/python/util/deprecation.py", line 296, in new_func
return func(*args, **kwargs)
  File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 458, in fit
loss = self._train_model(input_fn=input_fn, hooks=hooks)
  File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 958, in _train_model
model_fn_ops = self._get_train_ops(features, labels)
 File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1165, in _get_train_ops
return self._call_model_fn(features, labels, model_fn_lib.ModeKeys.TRAIN)
  File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 1136, in _call_model_fn
model_fn_results = self._model_fn(features, labels, **kwargs)
  File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/linear.py", line 186, in _linear_model_fn
logits=logits)
  File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/head.py", line 854, in create_model_fn_ops
enable_centered_bias=self._enable_centered_bias)
  File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/head.py", line 649, in _create_model_fn_ops
batch_size, loss_fn, weight_tensor)
  File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/head.py", line 1911, in _train_op
train_op = train_op_fn(loss)
  File "/lib/python3.6/site-packages/tensorflow/contrib/learn/python/learn/estimators/linear.py", line 179, in _train_op_fn
zip(grads, my_vars), global_step=global_step))
  File "/lib/python3.6/site-packages/tensorflow/python/training/optimizer.py", line 456, in apply_gradients
update_ops.append(processor.update_op(self, grad))
  File "/lib/python3.6/site-packages/tensorflow/python/training/optimizer.py", line 97, in update_op
return optimizer._apply_dense(g, self._v)  # pylint: disable=protected-access
  File "/lib/python3.6/site-packages/tensorflow/python/training/gradient_descent.py", line 50, in _apply_dense
use_locking=self._use_locking).op
  File "/lib/python3.6/site-packages/tensorflow/python/training/gen_training_ops.py", line 370, in apply_gradient_descent
name=name)
  File "/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 330, in apply_op
g = ops._get_graph_from_inputs(_Flatten(keywords.values()))
  File "/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 4262, in _get_graph_from_inputs
_assert_same_graph(original_graph_element, graph_element)
  File "/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 4201, in _assert_same_graph
"%s must be from the same graph as %s." % (item, original_item))
ValueError: Tensor("ExponentialDecay:0", shape=(20, 1), dtype=float32) must be from the same graph as Tensor("linear/x/weight/part_0:0", shape=(20, 1), dtype=float32_ref).

I am using tensorflow 1.3. It works if i replace the learning rate by a constant, say 0.01. I have used a decaying learning rate in the past in combination with minimize operation but was trying to use it within LinearClassifier. I see that something seems inconsistent in the fact that I don't link the global step to the step in the fit, but was wondering how this can work. I suppose I could use a placeholder as suggested here but I don't see why I should code the update rule myself if i don't need to.

Any suggestions on how to solve this ?

Answer 1

Have you tried to get the global_step by calling tf.train.get_global_step()? This should return the global_step used by your LinearClassifier model.

Instead of

global_step = tf.Variable(0, trainable=False)

use

global_step = tf.train.get_global_step()

This worked for me using my own Estimator class, where I use the tf.train.MomentumOptimizer to minimize the tf.nn.sparse_softmax_cross_entropy_with_logits.