Merge pull request #970 from peastman/lstm (b1383f67) · Commits · 钟慕尧 / deepchem

deepchem/models/tensorgraph/layers.py

+71 −1

Original line number	Diff line number	Diff line
		@@ -820,7 +820,7 @@ class GRU(Layer):
		"""A Gated Recurrent Unit.

		This layer expects its input to be of shape (batch_size, sequence_length, ...).
		It consists of a set of independent sequence (one for each element in the batch),
		It consists of a set of independent sequences (one for each element in the batch),
		that are each propagated independently through the GRU.
		"""

		@@ -879,6 +879,76 @@ class GRU(Layer):
		self.out_tensor, self.rnn_initial_states, self.rnn_final_states, self.rnn_zero_states = tensor


		class LSTM(Layer):
		"""A Long Short Term Memory.

		This layer expects its input to be of shape (batch_size, sequence_length, ...).
		It consists of a set of independent sequences (one for each element in the batch),
		that are each propagated independently through the LSTM.
		"""

		def __init__(self, n_hidden, batch_size, **kwargs):
		"""Create a Long Short Term Memory.

		Parameters
		----------
		n_hidden: int
		the size of the LSTM's hidden state, which also determines the size of its output
		batch_size: int
		the batch size that will be used with this layer
		"""
		self.n_hidden = n_hidden
		self.batch_size = batch_size
		super(LSTM, self).__init__(**kwargs)
		try:
		parent_shape = self.in_layers[0].shape
		self._shape = (batch_size, parent_shape[1], n_hidden)
		except:
		pass

		def create_tensor(self, in_layers=None, set_tensors=True, **kwargs):
		inputs = self._get_input_tensors(in_layers)
		if len(inputs) != 1:
		raise ValueError("Must have one parent")
		parent_tensor = inputs[0]
		lstm_cell = tf.contrib.rnn.LSTMCell(self.n_hidden)
		zero_state = lstm_cell.zero_state(self.batch_size, tf.float32)
		if set_tensors:
		initial_state = tf.contrib.rnn.LSTMStateTuple(
		tf.placeholder(tf.float32, zero_state.c.get_shape()),
		tf.placeholder(tf.float32, zero_state.h.get_shape()))
		else:
		initial_state = zero_state
		out_tensor, final_state = tf.nn.dynamic_rnn(
		lstm_cell, parent_tensor, initial_state=initial_state, scope=self.name)
		if set_tensors:
		self._record_variable_scope(self.name)
		self.out_tensor = out_tensor
		self.rnn_initial_states.append(initial_state.c)
		self.rnn_initial_states.append(initial_state.h)
		self.rnn_final_states.append(final_state.c)
		self.rnn_final_states.append(final_state.h)
		self.rnn_zero_states.append(
		np.zeros(zero_state.c.get_shape(), np.float32))
		self.rnn_zero_states.append(
		np.zeros(zero_state.h.get_shape(), np.float32))
		return out_tensor

		def none_tensors(self):
		saved_tensors = [
		self.out_tensor, self.rnn_initial_states, self.rnn_final_states,
		self.rnn_zero_states
		]
		self.out_tensor = None
		self.rnn_initial_states = []
		self.rnn_final_states = []
		self.rnn_zero_states = []
		return saved_tensors

		def set_tensors(self, tensor):
		self.out_tensor, self.rnn_initial_states, self.rnn_final_states, self.rnn_zero_states = tensor


		class TimeSeriesDense(Layer):

		def __init__(self, out_channels, **kwargs):

deepchem/models/tensorgraph/tests/test_layers.py

+16 −0

Original line number	Diff line number	Diff line
		@@ -30,6 +30,7 @@ from deepchem.models.tensorgraph.layers import InputFifoQueue
		from deepchem.models.tensorgraph.layers import InteratomicL2Distances
		from deepchem.models.tensorgraph.layers import IterRefLSTMEmbedding
		from deepchem.models.tensorgraph.layers import L2Loss
		from deepchem.models.tensorgraph.layers import LSTM
		from deepchem.models.tensorgraph.layers import LSTMStep
		from deepchem.models.tensorgraph.layers import Log
		from deepchem.models.tensorgraph.layers import Multiply
		@@ -172,6 +173,21 @@ class TestLayers(test_util.TensorFlowTestCase):
		out_tensor = out_tensor.eval()
		assert out_tensor.shape == (batch_size, n_steps, n_hidden)

		def test_lstm(self):
		"""Test that LSTM can be invoked."""
		batch_size = 10
		n_hidden = 7
		in_channels = 4
		n_repeat = 2
		n_steps = 6
		in_tensor = np.random.rand(batch_size, n_steps, in_channels)
		with self.test_session() as sess:
		in_tensor = tf.convert_to_tensor(in_tensor, dtype=tf.float32)
		out_tensor = LSTM(n_hidden, batch_size)(in_tensor)
		sess.run(tf.global_variables_initializer())
		out_tensor = out_tensor.eval()
		assert out_tensor.shape == (batch_size, n_steps, n_hidden)

		def test_time_series_dense(self):
		"""Test that TimeSeriesDense can be invoked."""
		batch_size = 10

deepchem/models/tensorgraph/tests/test_layers_pickle.py

+11 −1

Original line number	Diff line number	Diff line
		@@ -5,7 +5,7 @@ from deepchem.models import TensorGraph
		from deepchem.models.tensorgraph.graph_layers import WeaveLayer, WeaveGather, DTNNEmbedding, DTNNGather, DTNNStep, \
		DTNNExtract, DAGLayer, DAGGather, MessagePassing, SetGather
		from deepchem.models.tensorgraph.layers import Feature, Conv1D, Dense, Flatten, Reshape, Squeeze, Transpose, \
		CombineMeanStd, Repeat, Gather, GRU, L2Loss, Concat, SoftMax, \
		CombineMeanStd, Repeat, Gather, GRU, LSTM, L2Loss, Concat, SoftMax, \
		Constant, Variable, StopGradient, Add, Multiply, Log, Exp, InteratomicL2Distances, \
		SoftMaxCrossEntropy, ReduceMean, ToFloat, ReduceSquareDifference, Conv2D, MaxPool2D, ReduceSum, GraphConv, GraphPool, \
		GraphGather, BatchNorm, WeightedError, \
		@@ -114,6 +114,16 @@ def test_GRU_pickle():
		tg.save()


		def test_LSTM_pickle():
		tg = TensorGraph()
		feature = Feature(shape=(tg.batch_size, 10, 10))
		layer = LSTM(n_hidden=10, batch_size=tg.batch_size, in_layers=feature)
		tg.add_output(layer)
		tg.set_loss(layer)
		tg.build()
		tg.save()


		def test_L2loss_pickle():
		tg = TensorGraph()
		feature = Feature(shape=(tg.batch_size, 1))

Admin message