Merge branch 'torchmodel' of https://github.com/peastman/deepchem into torchmodel (7db1a9a0) · Commits · 钟慕尧 / deepchem

deepchem/models/losses.py

+36 −8

Original line number	Diff line number	Diff line
		@@ -32,7 +32,7 @@ class L1Loss(Loss):

		def _compute_tf_loss(self, output, labels):
		import tensorflow as tf
		output, labels = _make_shapes_consistent(output, labels)
		output, labels = _make_tf_shapes_consistent(output, labels)
		output, labels = _ensure_float(output, labels)
		return tf.abs(output - labels)

		@@ -46,7 +46,7 @@ class L2Loss(Loss):

		def _compute_tf_loss(self, output, labels):
		import tensorflow as tf
		output, labels = _make_shapes_consistent(output, labels)
		output, labels = _make_tf_shapes_consistent(output, labels)
		output, labels = _ensure_float(output, labels)
		return tf.square(output - labels)

		@@ -64,13 +64,14 @@ class HingeLoss(Loss):

		def _compute_tf_loss(self, output, labels):
		import tensorflow as tf
		output, labels = _make_shapes_consistent(output, labels)
		output, labels = _make_tf_shapes_consistent(output, labels)
		return tf.keras.losses.hinge(labels, output)

		def _create_pytorch_loss(self):
		import torch

		def loss(output, labels):
		output, labels = _make_pytorch_shapes_consistent(output, labels)
		return torch.mean(torch.clamp(1 - labels * output, min=0), dim=-1)

		return loss
		@@ -85,7 +86,7 @@ class BinaryCrossEntropy(Loss):

		def _compute_tf_loss(self, output, labels):
		import tensorflow as tf
		output, labels = _make_shapes_consistent(output, labels)
		output, labels = _make_tf_shapes_consistent(output, labels)
		output, labels = _ensure_float(output, labels)
		return tf.keras.losses.binary_crossentropy(labels, output)

		@@ -94,6 +95,7 @@ class BinaryCrossEntropy(Loss):
		bce = torch.nn.BCELoss(reduction='none')

		def loss(output, labels):
		output, labels = _make_pytorch_shapes_consistent(output, labels)
		return torch.mean(bce(output, labels), dim=-1)

		return loss
		@@ -109,7 +111,7 @@ class CategoricalCrossEntropy(Loss):

		def _compute_tf_loss(self, output, labels):
		import tensorflow as tf
		output, labels = _make_shapes_consistent(output, labels)
		output, labels = _make_tf_shapes_consistent(output, labels)
		output, labels = _ensure_float(output, labels)
		return tf.keras.losses.categorical_crossentropy(labels, output)

		@@ -117,6 +119,7 @@ class CategoricalCrossEntropy(Loss):
		import torch

		def loss(output, labels):
		output, labels = _make_pytorch_shapes_consistent(output, labels)
		return -torch.sum(labels * torch.log(output), dim=-1)

		return loss
		@@ -132,7 +135,7 @@ class SigmoidCrossEntropy(Loss):

		def _compute_tf_loss(self, output, labels):
		import tensorflow as tf
		output, labels = _make_shapes_consistent(output, labels)
		output, labels = _make_tf_shapes_consistent(output, labels)
		output, labels = _ensure_float(output, labels)
		return tf.nn.sigmoid_cross_entropy_with_logits(labels, output)

		@@ -141,6 +144,7 @@ class SigmoidCrossEntropy(Loss):
		bce = torch.nn.BCEWithLogitsLoss(reduction='none')

		def loss(output, labels):
		output, labels = _make_pytorch_shapes_consistent(output, labels)
		return bce(output, labels)

		return loss
		@@ -157,7 +161,7 @@ class SoftmaxCrossEntropy(Loss):

		def _compute_tf_loss(self, output, labels):
		import tensorflow as tf
		output, labels = _make_shapes_consistent(output, labels)
		output, labels = _make_tf_shapes_consistent(output, labels)
		output, labels = _ensure_float(output, labels)
		return tf.nn.softmax_cross_entropy_with_logits(labels, output)

		@@ -166,6 +170,7 @@ class SoftmaxCrossEntropy(Loss):
		ls = torch.nn.LogSoftmax(dim=1)

		def loss(output, labels):
		output, labels = _make_pytorch_shapes_consistent(output, labels)
		return -torch.sum(labels * ls(output), dim=-1)

		return loss
		@@ -190,7 +195,7 @@ class SparseSoftmaxCrossEntropy(Loss):
		return torch.nn.CrossEntropyLoss(reduction='none')


		def _make_shapes_consistent(output, labels):
		def _make_tf_shapes_consistent(output, labels):
		"""Try to make inputs have the same shape by adding dimensions of size 1."""
		import tensorflow as tf
		shape1 = output.shape
		@@ -215,6 +220,29 @@ def _make_shapes_consistent(output, labels):
		(str(shape1), str(shape2)))


		def _make_pytorch_shapes_consistent(output, labels):
		"""Try to make inputs have the same shape by adding dimensions of size 1."""
		import torch
		shape1 = output.shape
		shape2 = labels.shape
		len1 = len(shape1)
		len2 = len(shape2)
		if len1 == len2:
		return (output, labels)
		shape1 = tuple(shape1)
		shape2 = tuple(shape2)
		if len1 > len2 and all(i == 1 for i in shape1[len2:]):
		for i in range(len1 - len2):
		labels = torch.unsqueeze(labels, -1)
		return (output, labels)
		if len2 > len1 and all(i == 1 for i in shape2[len1:]):
		for i in range(len2 - len1):
		output = torch.unsqueeze(output, -1)
		return (output, labels)
		raise ValueError("Incompatible shapes for outputs and labels: %s versus %s" %
		(str(shape1), str(shape2)))


		def _ensure_float(output, labels):
		"""Make sure the outputs and labels are both floating point types."""
		import tensorflow as tf

deepchem/models/tests/test_torch_model.py

+2 −2

Original line number	Diff line number	Diff line
		@@ -36,7 +36,7 @@ def test_overfit_subclass_model():
		x = layer(x)
		if i < len(self.layers) - 1:
		x = F.relu(x)
		return F.sigmoid(x), x
		return torch.sigmoid(x), x

		pytorch_model = ExampleModel([10, 1])
		model = dc.models.TorchModel(
		@@ -318,7 +318,7 @@ def test_tensorboard():
		y = [[0.0, 1.0] for x in range(n_data_points)]
		dataset = dc.data.NumpyDataset(X, y)
		pytorch_model = torch.nn.Sequential(
		torch.nn.Linear(n_features, 2), torch.nn.Softmax())
		torch.nn.Linear(n_features, 2), torch.nn.Softmax(dim=1))
		model = dc.models.TorchModel(
		pytorch_model,
		dc.models.losses.CategoricalCrossEntropy(),

deepchem/models/torch_model.py

+22 −8

Original line number	Diff line number	Diff line
		@@ -121,6 +121,7 @@ class TorchModel(Model):
		tensorboard: bool = False,
		wandb: bool = False,
		log_frequency: int = 100,
		device: Optional[torch.device] = None,
		**kwargs) -> None:
		"""Create a new TorchModel.

		@@ -156,6 +157,9 @@ class TorchModel(Model):
		a global step corresponds to one batch of training. If you'd
		like a printout every 10 batch steps, you'd set
		`log_frequency=10` for example.
		device: torch.device
		the device on which to run computations. If None, a device is
		chosen automatically.
		"""
		super(TorchModel, self).__init__(
		model_instance=model, model_dir=model_dir, **kwargs)
		@@ -171,6 +175,16 @@ class TorchModel(Model):
		self.optimizer = optimizer
		self.tensorboard = tensorboard

		# Select a device.

		if device is None:
		if torch.cuda.is_available():
		device = torch.device('cuda')
		else:
		device = torch.device('cpu')
		self.device = device
		model.to(device)

		# W&B logging
		if wandb and not is_wandb_available():
		logger.warning(
		@@ -522,7 +536,7 @@ class TorchModel(Model):
		output_values = self.model(inputs)
		if isinstance(output_values, torch.Tensor):
		output_values = [output_values]
		output_values = [t.detach().numpy() for t in output_values]
		output_values = [t.detach().cpu().numpy() for t in output_values]

		# Apply tranformers and record results.
		if uncertainty:
		@@ -810,7 +824,7 @@ class TorchModel(Model):
		output_shape = tuple(output.shape[1:])
		output = output.reshape([-1])
		result = []
		grad_output = torch.zeros(output.shape[0])
		grad_output = torch.zeros(output.shape[0], device=self.device)
		for i in range(output.shape[0]):
		grad_output.zero_()
		grad_output[i] = 1
		@@ -819,7 +833,7 @@ class TorchModel(Model):
		X.grad.zero_()
		final_result.append(
		torch.reshape(torch.stack(result),
		output_shape + input_shape).numpy())
		output_shape + input_shape).cpu().numpy())
		if len(final_result) == 1:
		return final_result[0]
		return final_result
		@@ -834,13 +848,13 @@ class TorchModel(Model):
		labels = [
		x.astype(np.float32) if x.dtype == np.float64 else x for x in labels
		]
		labels = [torch.as_tensor(x) for x in labels]
		labels = [torch.as_tensor(x, device=self.device) for x in labels]
		if weights is not None:
		weights = [
		x.astype(np.float32) if x.dtype == np.float64 else x for x in weights
		]
		weights = [torch.as_tensor(x) for x in weights]
		inputs = [torch.as_tensor(x) for x in inputs]
		weights = [torch.as_tensor(x, device=self.device) for x in weights]
		inputs = [torch.as_tensor(x, device=self.device) for x in inputs]

		return (inputs, labels, weights)

		@@ -1024,7 +1038,7 @@ class TorchModel(Model):
		source_vars = list(source_model.model.parameters())

		for source_var in source_vars:
		value_map[source_var] = source_var.detach().numpy()
		value_map[source_var] = source_var.detach().cpu().numpy()

		return value_map

		@@ -1094,7 +1108,7 @@ class TorchModel(Model):

		for source_var, dest_var in assignment_map.items():
		assert source_var.shape == dest_var.shape
		dest_var.data = torch.as_tensor(value_map[source_var])
		dest_var.data = torch.as_tensor(value_map[source_var], device=self.device)


		class _StandardLoss(object):

Admin message