Added Huber loss and relevant torch+TF tests (e7d390ef) · Commits · 钟慕尧 / deepchem

deepchem/models/losses.py

+15 −0

Original line number	Diff line number	Diff line
		@@ -40,6 +40,21 @@ class L1Loss(Loss):
		import torch
		return torch.nn.L1Loss(reduction='none')

		class HuberLoss(Loss):
		"""Smoothened version of L1 loss.
		Less sensitive to small errors than L1, becomes linear in error for larger errors.
		When delta = 1.0 (default), this loss is equivalent to SmoothL1Loss.
		"""

		def _compute_tf_loss(self, output, labels):
		import tensorflow as tf
		output, labels = _make_tf_shapes_consistent(output, labels)
		return tf.keras.losses.Huber(reduction="none")(labels, output)


		def _create_pytorch_loss(self):
		import torch
		return torch.nn.SmoothL1Loss(reduction='none')

		class L2Loss(Loss):
		"""The squared difference between the true and predicted values."""

deepchem/models/tests/test_gcn.py

+37 −43

Original line number	Diff line number	Diff line
		@@ -22,13 +22,12 @@ except:
		def test_gcn_regression():
		# load datasets
		featurizer = MolGraphConvFeaturizer()
		tasks, dataset, transformers, metric = get_dataset(
		'regression', featurizer=featurizer)
		tasks, dataset, transformers, metric = get_dataset('regression',
		featurizer=featurizer)

		# initialize models
		n_tasks = len(tasks)
		model = GCNModel(
		mode='regression',
		model = GCNModel(mode='regression',
		n_tasks=n_tasks,
		number_atom_features=30,
		batch_size=10,
		@@ -44,8 +43,7 @@ def test_gcn_regression():

		tasks, all_dataset, transformers = load_delaney(featurizer=featurizer)
		train_set, _, _ = all_dataset
		model = dc.models.GCNModel(
		n_tasks=len(tasks),
		model = dc.models.GCNModel(n_tasks=len(tasks),
		graph_conv_layers=[2],
		residual=False,
		predictor_hidden_feats=2)
		@@ -57,13 +55,12 @@ def test_gcn_regression():
		def test_gcn_classification():
		# load datasets
		featurizer = MolGraphConvFeaturizer()
		tasks, dataset, transformers, metric = get_dataset(
		'classification', featurizer=featurizer)
		tasks, dataset, transformers, metric = get_dataset('classification',
		featurizer=featurizer)

		# initialize models
		n_tasks = len(tasks)
		model = GCNModel(
		mode='classification',
		model = GCNModel(mode='classification',
		n_tasks=n_tasks,
		number_atom_features=30,
		batch_size=10,
		@@ -80,8 +77,7 @@ def test_gcn_classification():
		tasks, all_dataset, transformers = load_bace_classification(
		featurizer=featurizer)
		train_set, _, _ = all_dataset
		model = dc.models.GCNModel(
		mode='classification',
		model = dc.models.GCNModel(mode='classification',
		n_tasks=len(tasks),
		graph_conv_layers=[2],
		residual=False,
		@@ -94,14 +90,13 @@ def test_gcn_classification():
		def test_gcn_reload():
		# load datasets
		featurizer = MolGraphConvFeaturizer()
		tasks, dataset, transformers, metric = get_dataset(
		'classification', featurizer=featurizer)
		tasks, dataset, transformers, metric = get_dataset('classification',
		featurizer=featurizer)

		# initialize models
		n_tasks = len(tasks)
		model_dir = tempfile.mkdtemp()
		model = GCNModel(
		mode='classification',
		model = GCNModel(mode='classification',
		n_tasks=n_tasks,
		number_atom_features=30,
		model_dir=model_dir,
		@@ -112,8 +107,7 @@ def test_gcn_reload():
		scores = model.evaluate(dataset, [metric], transformers)
		assert scores['mean-roc_auc_score'] >= 0.85

		reloaded_model = GCNModel(
		mode='classification',
		reloaded_model = GCNModel(mode='classification',
		n_tasks=n_tasks,
		number_atom_features=30,
		model_dir=model_dir,

deepchem/models/tests/test_losses.py

+20 −0

Original line number	Diff line number	Diff line
		@@ -38,6 +38,26 @@ class TestLosses(unittest.TestCase):
		expected = [[0.1, 0.2], [0.6, 0.6]]
		assert np.allclose(expected, result)

		@unittest.skipIf(not has_tensorflow, 'TensorFlow is not installed')
		def test_huber_loss_tf(self):
		"""Test HuberLoss."""
		loss = losses.HuberLoss()
		outputs = tf.constant([[0.1, 0.8], [0.4, 0.6]])
		labels = tf.constant([[1.0, -1.0], [-1.0, 1.0]])
		result = np.mean(loss._compute_tf_loss(outputs, labels).numpy())
		expected = 0.67125
		assert np.allclose(expected, result)

		@unittest.skipIf(not has_pytorch, 'PyTorch is not installed')
		def test_huber_loss_pytorch(self):
		"""Test HuberLoss."""
		loss = losses.HuberLoss()
		outputs = torch.tensor([[0.1, 0.8], [0.4, 0.6]])
		labels = torch.tensor([[1.0, -1.0], [-1.0, 1.0]])
		result = np.mean(loss._create_pytorch_loss()(outputs, labels).numpy())
		expected = 0.67125
		assert np.allclose(expected, result)

		@unittest.skipIf(not has_tensorflow, 'TensorFlow is not installed')
		def test_l2_loss_tf(self):
		"""Test L2Loss."""

Admin message