Remove doubled softmax and handle keras edge case. (9d385e0f) · Commits · 钟慕尧 / deepchem

deepchem/models/init.py

+9 −1

Original line number	Diff line number	Diff line
		@@ -373,11 +373,19 @@ class Model(object):
		n_tasks = self.get_num_tasks()
		for (X_batch, y_batch, w_batch, ids_batch) in dataset.iterbatches(
		batch_size, deterministic=True):
		################################################################# DEBUG
		n_samples = len(X_batch)
		y_pred_batch = self.predict_proba_on_batch(X_batch, pad_batch=pad_batches)
		################################################################# DEBUG
		print("X_batch.shape")
		print(X_batch.shape)
		print("y_pred_batch.shape")
		print(y_pred_batch.shape)
		################################################################# DEBUG
		y_pred_batch = y_pred_batch[:n_samples]
		################################################################# DEBUG
		print("y_pred_batch.shape, (n_samples, n_tasks, n_classes)")
		print(y_pred_batch.shape, (n_samples, n_tasks, n_classes))
		################################################################# DEBUG
		y_pred_batch = np.reshape(y_pred_batch, (n_samples, n_tasks, n_classes))
		y_pred_batch = undo_transforms(y_pred_batch, transformers)
		y_preds.append(y_pred_batch)

+17 −4

Original line number	Diff line number	Diff line
		@@ -64,10 +64,15 @@ class KerasModel(Model):
		X: np.ndarray
		Features
		pad_batch: bool, optional
		Ignored for Keras Models. Only used for Tensorflow models
		with rigid batch-size requirements.
		Used for Tensorflow models with rigid batch-size requirements.
		"""
		return self.model_instance.predict_on_batch(X)
		n_samples = len(X)
		n_tasks = self.get_num_tasks()
		if pad_batch:
		X = pad_features(self.batch_size, X)
		y_pred = self.model_instance.predict_on_batch(X)
		y_pred = np.reshape(y_pred, (n_samples, n_tasks))
		return y_pred

		# TODO(rbharath): The methods below aren't extensible and depend on
		# implementation details of fcnet. Better way to expose this information?
		@@ -92,4 +97,12 @@ class KerasModel(Model):
		n_classes: int
		Number of classifier classes
		"""
		return self.model_instance.predict_proba_on_batch(X, n_classes)
		n_samples = len(X)
		n_tasks = self.get_num_tasks()

		if pad_batch:
		X = pad_features(self.batch_size, X)
		y_pred_proba = self.model_instance.predict_proba_on_batch(X,
		n_classes)
		y_pred_proba = np.reshape(y_pred_proba, (n_samples, n_tasks, n_classes))
		return y_pred_proba

+9 −7

Original line number	Diff line number	Diff line
		@@ -541,7 +541,6 @@ class TensorflowClassifier(TensorflowGraphModel):
		with self.eval_graph.graph.as_default():
		# run eval data through the model
		n_tasks = self.n_tasks
		outputs = []
		with self._get_shared_session(train=False).as_default():
		feed_dict = self.construct_feed_dict(X)
		data = self._get_shared_session(train=False).run(
		@@ -556,13 +555,16 @@ class TensorflowClassifier(TensorflowGraphModel):
		raise ValueError(
		'Unrecognized rank combination for output: %s ' %
		(batch_outputs.shape,))
		outputs.append(batch_outputs)

		# TODO(rbharath): This is a bug! We're actually applying softmax twice.
		# I believe this is harmless since softmax of softmax doesn't change
		# properties, but I need to check this...
		# We apply softmax to predictions to get class probabilities.
		outputs = softmax(np.squeeze(np.hstack(outputs)))
		#outputs = softmax(np.squeeze(batch_outputs))
		#outputs = softmax(batch_outputs)

		# Note that softmax is already applied in construct_grpah
		outputs = batch_outputs

		return np.copy(outputs)