Redesigned Gather layer (b4303810) · Commits · 钟慕尧 / deepchem

deepchem/models/tensorgraph/layers.py

+36 −18

Original line number	Diff line number	Diff line
		@@ -301,7 +301,7 @@ class Dense(Layer):
		self.time_series = time_series
		try:
		parent_shape = self.in_layers[0].shape
		self._shape = tuple(parent_shape[:-1] + [out_channels])
		self._shape = tuple(parent_shape[:-1]) + (out_channels,)
		except:
		pass
		self._reuse = False
		@@ -518,15 +518,22 @@ class Repeat(Layer):


		class Gather(Layer):
		"""Gather slices from the input."""
		"""Gather elements or slices from the input."""

		def __init__(self, in_layers=None, indices=None, axis=0, **kwargs):
		def __init__(self, in_layers=None, indices=None, **kwargs):
		"""Create a Gather layer.

		The slice indices may be specified in two ways. If the indices are constants,
		you can pass them to this constructor as a list or array. Alternatively, the
		indices can be calculated by another layer. In that case, pass None for indices,
		and instead provide them as the second input layer.
		The indices can be viewed as a list of element identifiers, where each
		identifier is itself a length N array specifying the indices of the
		first N dimensions of the input tensor. Those elements (or slices, depending
		on the shape of the input) are then stacked together. For example,
		indices=[[0],[2],[4]] will produce [input[0], input[2], input[4]], while
		indices=[[1,2]] will produce [input[1,2]].

		The indices may be specified in two ways. If they are constants, you can pass
		them to this constructor as a list or array. Alternatively, the indices can
		be calculated by another layer. In that case, pass None for indices, and
		instead provide them as the second input layer.

		Parameters
		----------
		@@ -537,19 +544,16 @@ class Gather(Layer):
		indices: array
		the slice indices (if they are constants) or None (if the indices are provided by
		an input)
		axis: int
		the axis along which to select slices. The indices refer to positions along this axis.
		"""
		self.indices = indices
		self.axis = axis
		super(Gather, self).__init__(in_layers, **kwargs)
		try:
		s = list(self.in_layers[0].shape)
		s = tuple(self.in_layers[0].shape)
		if indices is None:
		s[axis] = None
		s2 = self.in_layers[1].shape
		self._shape = (s2[0],) + s[s2[-1]:]
		else:
		s[axis] = len(indices)
		self._shape = tuple(s)
		self._shape = (len(indices),) + s[np.array(indices).shape[-1]:]
		except:
		pass

		@@ -564,7 +568,7 @@ class Gather(Layer):
		raise ValueError("Must have one parent")
		indices = self.indices
		parent_tensor = inputs[0]
		out_tensor = tf.gather(parent_tensor, indices, axis=self.axis)
		out_tensor = tf.gather_nd(parent_tensor, indices)
		if set_tensors:
		self.out_tensor = out_tensor
		return out_tensor
		@@ -860,6 +864,14 @@ class Variable(Layer):
		return out_tensor


		def _max_dimension(x, y):
		if x is None:
		return y
		if y is None:
		return x
		return max(x, y)


		class Add(Layer):
		"""Compute the (optionally weighted) sum of the input layers."""

		@@ -881,7 +893,7 @@ class Add(Layer):
		shape2, shape1 = shape1, shape2
		offset = len(shape1) - len(shape2)
		for i in range(len(shape2)):
		shape1[i + offset] = max(shape1[i + offset], shape2[i])
		shape1[i + offset] = _max_dimension(shape1[i + offset], shape2[i])
		self._shape = tuple(shape1)
		except:
		pass
		@@ -916,7 +928,7 @@ class Multiply(Layer):
		shape2, shape1 = shape1, shape2
		offset = len(shape1) - len(shape2)
		for i in range(len(shape2)):
		shape1[i + offset] = max(shape1[i + offset], shape2[i])
		shape1[i + offset] = _max_dimension(shape1[i + offset], shape2[i])
		self._shape = tuple(shape1)
		except:
		pass
		@@ -937,7 +949,7 @@ class Log(Layer):
		def __init__(self, in_layers=None, **kwargs):
		super(Log, self).__init__(in_layers, **kwargs)
		try:
		self._shape = in_layers[0].shape
		self._shape = self.in_layers[0].shape
		except:
		pass

		@@ -1026,6 +1038,8 @@ class SoftMaxCrossEntropy(Layer):
		class ReduceMean(Layer):

		def __init__(self, in_layers=None, axis=None, **kwargs):
		if axis is not None and not isinstance(axis, Sequence):
		axis = [axis]
		self.axis = axis
		super(ReduceMean, self).__init__(in_layers, **kwargs)
		if axis is None:
		@@ -1074,6 +1088,8 @@ class ToFloat(Layer):
		class ReduceSum(Layer):

		def __init__(self, in_layers=None, axis=None, **kwargs):
		if axis is not None and not isinstance(axis, Sequence):
		axis = [axis]
		self.axis = axis
		super(ReduceSum, self).__init__(in_layers, **kwargs)
		if axis is None:
		@@ -1103,6 +1119,8 @@ class ReduceSum(Layer):
		class ReduceSquareDifference(Layer):

		def __init__(self, in_layers=None, axis=None, **kwargs):
		if axis is not None and not isinstance(axis, Sequence):
		axis = [axis]
		self.axis = axis
		super(ReduceSquareDifference, self).__init__(in_layers, **kwargs)
		if axis is None:

deepchem/models/tensorgraph/tests/test_layers.py

+4 −4

Original line number	Diff line number	Diff line
		@@ -153,10 +153,10 @@ class TestLayers(test_util.TensorFlowTestCase):
		"""Test that Gather can be invoked."""
		in_tensor = np.random.uniform(size=(5, 4)).astype(np.float32)
		with self.test_session() as sess:
		out_tensor = Gather(indices=[2, 3], axis=0)(in_tensor).eval()
		assert np.array_equal(in_tensor[[2, 3]], out_tensor)
		out_tensor = Gather(axis=1)(in_tensor, np.array([1, 3])).eval()
		assert np.array_equal(in_tensor[:, [1, 3]], out_tensor)
		out_tensor = Gather(indices=[[2], [3]])(in_tensor).eval()
		assert np.array_equal([in_tensor[2], in_tensor[3]], out_tensor)
		out_tensor = Gather()(in_tensor, np.array([[1, 1], [0, 3]])).eval()
		assert np.array_equal([in_tensor[1, 1], in_tensor[0, 3]], out_tensor)

		def test_gru(self):
		"""Test that GRU can be invoked."""

deepchem/models/tensorgraph/tests/test_layers_pickle.py

+1 −1

Original line number	Diff line number	Diff line
		@@ -94,7 +94,7 @@ def test_Repeat_pickle():
		def test_Gather_pickle():
		tg = TensorGraph()
		feature = Feature(shape=(tg.batch_size, 1))
		layer = Gather(indices=[0, 2, 3], in_layers=feature)
		layer = Gather(indices=[[0], [2], [3]], in_layers=feature)
		tg.add_output(layer)
		tg.set_loss(layer)
		tg.build()

Admin message