Merge remote-tracking branch 'remotes/origin/master' into BP2 (b5095e80) · Commits · 钟慕尧 / deepchem

deepchem/models/tensorgraph/layers.py

+32 −4

Original line number	Diff line number	Diff line
		@@ -141,17 +141,45 @@ def convert_to_layers(in_layers):


		class Conv1D(Layer):
		"""A 1D convolution on the input.

		def __init__(self, width, out_channels, **kwargs):
		This layer expects its input to be a three dimensional tensor of shape (batch size, width, # channels).
		"""

		def __init__(self,
		width,
		out_channels,
		stride=1,
		padding='SAME',
		activation_fn=tf.nn.relu,
		**kwargs):
		"""Create a Conv1D layer.

		Parameters
		----------
		width: int
		the width of the convolutional kernel
		out_channels: int
		the number of outputs produced by the convolutional kernel
		stride: int
		the stride between applications of the convolutional kernel
		padding: str
		the padding method to use, either 'SAME' or 'VALID'
		activation_fn: object
		the Tensorflow activation function to apply to the output
		"""
		self.width = width
		self.out_channels = out_channels
		self.stride = stride
		self.padding = padding
		self.activation_fn = activation_fn
		self.out_tensor = None
		super(Conv1D, self).__init__(**kwargs)

		def create_tensor(self, in_layers=None, set_tensors=True, **kwargs):
		inputs = self._get_input_tensors(in_layers)
		if len(inputs) != 1:
		raise ValueError("Only One Parent to conv1D over")
		raise ValueError("Conv1D layer must have exactly one parent")
		parent = inputs[0]
		if len(parent.get_shape()) != 3:
		raise ValueError("Parent tensor must be (batch, width, channel)")
		@@ -160,9 +188,9 @@ class Conv1D(Layer):
		f = tf.Variable(
		tf.random_normal([self.width, parent_channel_size, self.out_channels]))
		b = tf.Variable(tf.random_normal([self.out_channels]))
		t = tf.nn.conv1d(parent, f, stride=1, padding="SAME")
		t = tf.nn.conv1d(parent, f, stride=self.stride, padding=self.padding)
		t = tf.nn.bias_add(t, b)
		out_tensor = tf.nn.relu(t)
		out_tensor = self.activation_fn(t)
		if set_tensors:
		self._record_variable_scope(self.name)
		self.out_tensor = out_tensor

deepchem/models/tests/example_DTNN.mat

+3.34 KiB (13.4 KiB)

File changed.

No diff preview for this file type.

View original file

View changed file

deepchem/models/tests/test_overfit.py

+99 −2

Original line number	Diff line number	Diff line
		@@ -940,7 +940,6 @@ class TestOverfit(test_util.TensorFlowTestCase):
		np.random.seed(123)
		tf.set_random_seed(123)

		# Load mini log-solubility dataset.
		input_file = os.path.join(self.current_dir, "example_DTNN.mat")
		dataset = scipy.io.loadmat(input_file)
		X = dataset['X']
		@@ -983,7 +982,6 @@ class TestOverfit(test_util.TensorFlowTestCase):
		np.random.seed(123)
		tf.set_random_seed(123)

		# Load mini log-solubility dataset.
		input_file = os.path.join(self.current_dir, "example_DTNN.mat")
		dataset = scipy.io.loadmat(input_file)
		X = dataset['X']
		@@ -1012,6 +1010,105 @@ class TestOverfit(test_util.TensorFlowTestCase):

		assert scores[regression_metric.name] > .9

		def test_ANI_multitask_regression_overfit(self):
		"""Test ANI-1 regression overfits tiny data."""
		input_file = os.path.join(self.current_dir, "example_DTNN.mat")
		np.random.seed(123)
		tf.set_random_seed(123)
		dataset = scipy.io.loadmat(input_file)
		X = np.concatenate([np.expand_dims(dataset['Z'], 2), dataset['R']], axis=2)
		X = X[:, :13, :]
		y = dataset['T']
		w = np.ones_like(y)
		dataset = dc.data.DiskDataset.from_numpy(X, y, w, ids=None)
		regression_metric = dc.metrics.Metric(
		dc.metrics.pearson_r2_score, mode="regression")
		n_tasks = y.shape[1]
		batch_size = 10

		transformers = [
		dc.trans.NormalizationTransformer(transform_y=True, dataset=dataset),
		dc.trans.ANITransformer(
		max_atoms=13,
		atom_cases=[1, 6, 7, 8],
		radial_cutoff=8.,
		angular_cutoff=5.,
		radial_length=8,
		angular_length=4)
		]

		for transformer in transformers:
		dataset = transformer.transform(dataset)

		n_feat = transformers[-1].get_num_feats() - 1
		model = dc.models.ANIRegression(
		n_tasks,
		13,
		n_feat,
		atom_number_cases=[1, 6, 7, 8],
		batch_size=batch_size,
		learning_rate=0.001,
		use_queue=False,
		mode="regression")

		# Fit trained model
		model.fit(dataset, nb_epoch=50)

		# Eval model on train
		scores = model.evaluate(dataset, [regression_metric], transformers[0:1])

		assert scores[regression_metric.name] > .8

		def test_BP_symmetry_function_overfit(self):
		"""Test ANI-1 regression overfits tiny data."""
		input_file = os.path.join(self.current_dir, "example_DTNN.mat")
		np.random.seed(123)
		tf.set_random_seed(123)
		dataset = scipy.io.loadmat(input_file)
		X = np.concatenate([np.expand_dims(dataset['Z'], 2), dataset['R']], axis=2)
		X = X[:, :13, :]
		y = dataset['T']
		w = np.ones_like(y)
		dataset = dc.data.DiskDataset.from_numpy(X, y, w, ids=None)
		regression_metric = dc.metrics.Metric(
		dc.metrics.pearson_r2_score, mode="regression")
		n_tasks = y.shape[1]
		batch_size = 10

		transformers = [
		dc.trans.NormalizationTransformer(transform_y=True, dataset=dataset),
		dc.trans.ANITransformer(
		max_atoms=13,
		atom_cases=[1, 6, 7, 8],
		atomic_number_differentiated=False,
		radial_cutoff=8.,
		angular_cutoff=5.,
		radial_length=8,
		angular_length=4)
		]

		for transformer in transformers:
		dataset = transformer.transform(dataset)

		n_feat = transformers[-1].get_num_feats() - 1
		model = dc.models.ANIRegression(
		n_tasks,
		13,
		n_feat,
		atom_number_cases=[1, 6, 7, 8],
		batch_size=batch_size,
		learning_rate=0.001,
		use_queue=False,
		mode="regression")

		# Fit trained model
		model.fit(dataset, nb_epoch=50)

		# Eval model on train
		scores = model.evaluate(dataset, [regression_metric], transformers[0:1])

		assert scores[regression_metric.name] > .8

		def test_DAG_singletask_regression_overfit(self):
		"""Test DAG regressor multitask overfits tiny data."""
		np.random.seed(123)

deepchem/trans/transformers.py

+3 −2

Original line number	Diff line number	Diff line
		@@ -965,11 +965,11 @@ class ANITransformer(Transformer):
		def transform_array(self, X, y, w):
		if self.transform_X:
		n_samples = X.shape[0]
		batches = np.linspace(0, n_samples, int(n_samples / 100)).astype(int)
		batches = np.linspace(0, n_samples, int(n_samples / 100) + 1).astype(int)
		X_out = []
		sess = tf.Session(graph=self.compute_graph)
		num_transformed = 0
		for i, start in enumerate(batches[:-1]):
		for i, start in enumerate(batches):
		if start == batches[-1]:
		X_batch = X[start:]
		else:
		@@ -978,6 +978,7 @@ class ANITransformer(Transformer):
		X_out.append(output)
		num_transformed = num_transformed + X_batch.shape[0]
		print('%i samples transformed' % num_transformed)

		X_new = np.concatenate(X_out, axis=0)
		assert X_new.shape[0] == X.shape[0]
		return (X_new, y, w)

Admin message