Beginning to convert layers to support eager mode (8a8a1348) · Commits · 钟慕尧 / deepchem

deepchem/models/tensorgraph/layers.py

+33 −15

Original line number	Diff line number	Diff line
		@@ -8,6 +8,7 @@ import tensorflow as tf
		import numpy as np

		from deepchem.models.tensorgraph import model_ops, initializations, regularizers, activations
		import tensorflow.contrib.eager as tfe
		import math


		@@ -25,7 +26,6 @@ class Layer(object):
		in_layers = [in_layers]
		self.in_layers = in_layers
		self.op_type = "gpu"
		self.variable_scope = ''
		self.variable_values = None
		self.out_tensor = None
		self.rnn_initial_states = []
		@@ -33,6 +33,10 @@ class Layer(object):
		self.rnn_zero_states = []
		self.tensorboard = False
		self.tb_input = None
		if tfe.in_eager_mode():
		self.variables = []
		else:
		self.variable_scope = ''

		def _get_layer_number(self):
		class_name = self.__class__.__name__
		@@ -83,8 +87,8 @@ class Layer(object):
		return self.clone(in_layers)
		raise ValueError('%s does not implement shared()' % self.__class__.__name__)

		def __call__(self, *in_layers):
		return self.create_tensor(in_layers=in_layers, set_tensors=False)
		def __call__(self, *in_layers, training=False):
		return self.create_tensor(in_layers=in_layers, set_tensors=False, training=training)

		@property
		def shape(self):
		@@ -106,6 +110,8 @@ class Layer(object):
		if True, try to reshape the inputs to all have the same shape
		"""
		if in_layers is None:
		if tfe.in_eager_mode():
		raise ValueError('in_layers must be specified in eager mode')
		in_layers = self.in_layers
		if not isinstance(in_layers, Sequence):
		in_layers = [in_layers]
		@@ -241,6 +247,8 @@ class Layer(object):
		This means the newly created layers will share variables with the original
		ones.
		"""
		if tfe.in_eager_mode():
		raise ValueError('copy() is not supported in eager mode')
		if self in replacements:
		return replacements[self]
		copied_inputs = [
		@@ -328,7 +336,7 @@ class TensorWrapper(Layer):

		def create_tensor(self, in_layers=None, **kwargs):
		"""Take no actions."""
		pass
		return self.out_tensor


		def convert_to_layers(in_layers):
		@@ -466,17 +474,12 @@ class Conv1D(Layer):
		self.kernel_constraint = kernel_constraint
		self.bias_constraint = bias_constraint
		super(Conv1D, self).__init__(in_layers, **kwargs)
		if tfe.in_eager_mode():
		self._layer = self._build_layer()
		self.variables = self._layer.variables

		def create_tensor(self, in_layers=None, set_tensors=True, **kwargs):
		inputs = self._get_input_tensors(in_layers)
		if len(inputs) != 1:
		raise ValueError("Conv1D layer must have exactly one parent")
		parent = inputs[0]
		if len(parent.get_shape()) == 2:
		parent = tf.expand_dims(parent, 2)
		elif len(parent.get_shape()) != 3:
		raise ValueError("Parent tensor must be (batch, width, channel)")
		out_tensor = tf.keras.layers.Conv1D(
		def _build_layer(self):
		return tf.keras.layers.Conv1D(
		filters=self.filters,
		kernel_size=self.kernel_size,
		strides=self.strides,
		@@ -490,7 +493,22 @@ class Conv1D(Layer):
		bias_regularizer=self.bias_regularizer,
		activity_regularizer=self.activity_regularizer,
		kernel_constraint=self.kernel_constraint,
		bias_constraint=self.bias_constraint)(parent)
		bias_constraint=self.bias_constraint)

		def create_tensor(self, in_layers=None, set_tensors=True, **kwargs):
		inputs = self._get_input_tensors(in_layers)
		if len(inputs) != 1:
		raise ValueError("Conv1D layer must have exactly one parent")
		parent = inputs[0]
		if len(parent.get_shape()) == 2:
		parent = tf.expand_dims(parent, 2)
		elif len(parent.get_shape()) != 3:
		raise ValueError("Parent tensor must be (batch, width, channel)")
		if tfe.in_eager_mode():
		layer = self._layer
		else:
		layer = self._build_layer()
		out_tensor = layer(parent)
		if set_tensors:
		self._record_variable_scope(self.name)
		self.out_tensor = out_tensor

deepchem/models/tensorgraph/tests/test_layers_eager.py

0 → 100644

+74 −0

Original line number	Diff line number	Diff line
		import numpy as np
		import tensorflow as tf
		import deepchem.models.tensorgraph.layers as layers
		import tensorflow.contrib.eager as tfe
		from tensorflow.python.eager import context
		from tensorflow.python.framework import test_util

		class TestLayersEager(test_util.TensorFlowTestCase):
		"""
		Test that layers function in eager mode.
		"""

		def test_conv_1d(self):
		"""Test invoking Conv1D in eager mode."""
		with context.eager_mode():
		with tfe.IsolateTest():
		width = 5
		in_channels = 2
		filters = 3
		kernel_size = 2
		batch_size = 10
		input = np.random.rand(batch_size, width, in_channels).astype(np.float32)
		layer = layers.Conv1D(filters, kernel_size)
		result = layer(input)
		self.assertEqual(result.shape[0], batch_size)
		self.assertEqual(result.shape[2], filters)

		# Creating a second layer should produce different results, since it has
		# different random weights.

		layer2 = layers.Conv1D(filters, kernel_size)
		result2 = layer2(input)
		assert not np.allclose(result, result2)

		# But evaluating the first layer again should produce the same result as before.

		result3 = layer(input)
		assert np.allclose(result, result3)


		def test_add(self):
		"""Test invoking Add in eager mode."""
		with context.eager_mode():
		with tfe.IsolateTest():
		result = layers.Add()([1,2], [3,4])
		assert np.array_equal(result, [4,6])

		def test_multiply(self):
		"""Test invoking Multiply in eager mode."""
		with context.eager_mode():
		with tfe.IsolateTest():
		result = layers.Multiply()([1,2], [3,4])
		assert np.array_equal(result, [3,8])

		def test_divide(self):
		"""Test invoking Divide in eager mode."""
		with context.eager_mode():
		with tfe.IsolateTest():
		result = layers.Divide()([1,2], [2,5])
		assert np.array_equal(result, [0.5,0.4])

		def test_log(self):
		"""Test invoking Log in eager mode."""
		with context.eager_mode():
		with tfe.IsolateTest():
		result = layers.Log()(2.5)
		assert np.allclose(result, np.log(2.5))

		def test_exp(self):
		"""Test invoking Exp in eager mode."""
		with context.eager_mode():
		with tfe.IsolateTest():
		result = layers.Exp()(2.5)
		assert np.allclose(result, np.exp(2.5))

Admin message