Commit 8d7f3991 authored Mar 31, 2016 by Steven Kearnes

Fix regressor cost

The regression loss function is wrong in tfvs. `tf.nn.l2_loss` returns a scalar (the sum of the per-example weights), which obviously will frequently get the weighted losses wrong. This PR calculates per-example losses, which are then weighted appropriately. (Sorry, this should have been fixed when tfvs was released.)

parent 3c8e5b6c

deepchem/models/tensorflow_models/init.py

+1 −1

Original line number	Diff line number	Diff line
		@@ -655,7 +655,7 @@ class TensorflowRegressor(TensorflowGraph):
		A tensor with shape batch_size containing the weighted cost for each
		example.
		"""
		return tf.mul(tf.nn.l2_loss(output - labels), weights)
		return tf.mul(0.5 * tf.square(output - labels), weights)

		def example_counts(self, y_true):
		"""Get counts of examples in each class.

Admin message