Weights Not Trainable with `tf.contrib.layers.recompute_grad`

When I use tf.contrib.layers.recompute_grad on a layer, the trainable weights are eliminated. How do I save the weights that are recalculated on the backwards pass at each epoch?