def _mnist_batch_train(model, batch):
optimizer = tf.train.GradientDescentOptimizer(0.01)
model_vars = tf_computation_utils.get_variables('v', _mnist_model_type)
assign_vars_op = tf_computation_utils.assign(model_vars, model)
with tf.control_dependencies([assign_vars_op]):
train_op = optimizer.minimize(_mnist_batch_loss(model_vars, batch))
with tf.control_dependencies([train_op]):
return tf_computation_utils.identity(model_vars)
def test_get_variables_with_named_tuple_type(self):
x = tf_computation_utils.get_variables('foo',
[('x', tf.int32),
('y', tf.string), tf.bool])
self.assertIsInstance(x, anonymous_tuple.AnonymousTuple)
self.assertEqual(
str(x), '<x=<tf.Variable \'foo/x:0\' shape=() dtype=int32>,'
'y=<tf.Variable \'foo/y:0\' shape=() dtype=string>,'
'<tf.Variable \'foo/2:0\' shape=() dtype=bool>>')
def test_get_variables_with_named_tuple_type(self):
x = tf_computation_utils.get_variables('foo', [('x', tf.int32),
('y', tf.string), tf.bool])
self.assertIsInstance(x, anonymous_tuple.AnonymousTuple)
self.assertLen(x, 3)
self.assertEqual(dir(x), ['x', 'y'])
self._assertMatchesVariable(x[0], 'foo/x:0', (), tf.int32)
self._assertMatchesVariable(x[1], 'foo/y:0', (), tf.string)
self._assertMatchesVariable(x[2], 'foo/2:0', (), tf.bool)
def test_get_variables_with_tensor_type(self):
x = tf_computation_utils.get_variables('foo', tf.int32)
self.assertIsInstance(x, tf.Variable)
self.assertIs(x.dtype.base_dtype, tf.int32)
self.assertEqual(x.shape, tf.TensorShape([]))
self.assertEqual(str(x.name), 'foo:0')