Merge pull request #2538 from peastman/uncertainty (cbab9678) · Commits · 钟慕尧 / deepchem

deepchem/models/graph_models.py

+39 −6

Original line number	Diff line number	Diff line
		@@ -692,8 +692,19 @@ class DAGModel(KerasModel):
		output_types = ['prediction', 'variance', 'loss', 'loss']

		def loss(outputs, labels, weights):
		diff = labels[0] - outputs[0]
		return tf.reduce_mean(diff * diff / tf.exp(outputs[1]) + outputs[1])
		output, labels = dc.models.losses._make_tf_shapes_consistent(
		outputs[0], labels[0])
		output, labels = dc.models.losses._ensure_float(output, labels)
		losses = tf.square(output - labels) / tf.exp(outputs[1]) + outputs[1]
		w = weights[0]
		if len(w.shape) < len(losses.shape):
		if tf.is_tensor(w):
		shape = tuple(w.shape.as_list())
		else:
		shape = w.shape
		shape = tuple(-1 if x is None else x for x in shape)
		w = tf.reshape(w, shape + (1,) * (len(losses.shape) - len(w.shape)))
		return tf.reduce_mean(losses * w) + sum(self.model.losses)
		else:
		outputs = [output]
		output_types = ['prediction']
		@@ -960,8 +971,19 @@ class GraphConvModel(KerasModel):
		output_types = ['prediction', 'variance', 'loss', 'loss', 'embedding']

		def loss(outputs, labels, weights):
		diff = labels[0] - outputs[0]
		return tf.reduce_mean(diff * diff / tf.exp(outputs[1]) + outputs[1])
		output, labels = dc.models.losses._make_tf_shapes_consistent(
		outputs[0], labels[0])
		output, labels = dc.models.losses._ensure_float(output, labels)
		losses = tf.square(output - labels) / tf.exp(outputs[1]) + outputs[1]
		w = weights[0]
		if len(w.shape) < len(losses.shape):
		if tf.is_tensor(w):
		shape = tuple(w.shape.as_list())
		else:
		shape = w.shape
		shape = tuple(-1 if x is None else x for x in shape)
		w = tf.reshape(w, shape + (1,) * (len(losses.shape) - len(w.shape)))
		return tf.reduce_mean(losses * w) + sum(self.model.losses)
		else:
		output_types = ['prediction', 'embedding']
		loss = L2Loss()
		@@ -1101,8 +1123,19 @@ class MPNNModel(KerasModel):
		output_types = ['prediction', 'variance', 'loss', 'loss']

		def loss(outputs, labels, weights):
		diff = labels[0] - outputs[0]
		return tf.reduce_mean(diff * diff / tf.exp(outputs[1]) + outputs[1])
		output, labels = dc.models.losses._make_tf_shapes_consistent(
		outputs[0], labels[0])
		output, labels = dc.models.losses._ensure_float(output, labels)
		losses = tf.square(output - labels) / tf.exp(outputs[1]) + outputs[1]
		w = weights[0]
		if len(w.shape) < len(losses.shape):
		if tf.is_tensor(w):
		shape = tuple(w.shape.as_list())
		else:
		shape = w.shape
		shape = tuple(-1 if x is None else x for x in shape)
		w = tf.reshape(w, shape + (1,) * (len(losses.shape) - len(w.shape)))
		return tf.reduce_mean(losses * w) + sum(self.model.losses)
		else:
		outputs = [output]
		output_types = ['prediction']

deepchem/models/tests/test_overfit.py

+11 −0

Original line number	Diff line number	Diff line
		@@ -769,6 +769,17 @@ def test_multitask_regressor_uncertainty():
		assert noise < np.mean(std) < 1.0


		def test_multitask_regressor_delaney_uncertainty():
		"""Test computing uncertainty on a larger dataset."""
		tasks, datasets, transformers = dc.molnet.load_delaney('ECFP')
		train_dataset, valid_dataset, test_dataset = datasets
		model = dc.models.MultitaskRegressor(len(tasks), 1024, uncertainty=True)
		model.fit(train_dataset, nb_epoch=20)
		metric = dc.metrics.Metric(dc.metrics.pearsonr)
		scores = model.evaluate(test_dataset, [metric], transformers)
		assert scores['pearsonr'] > 0.5


		@pytest.mark.slow
		def test_DAG_singletask_regression_overfit():
		"""Test DAG regressor multitask overfits tiny data."""

Admin message