network_dense.py 文件源码

def _apply_prune_on_grads(self,
                              grads_and_vars: list,
                              threshold: float):

        # we need to make gradients correspondent
        # to the pruned weights to be zero

        grads_and_vars_sparse = []

        for grad, var in grads_and_vars:
            if 'weights' in var.name:
                small_weights = tf.greater(threshold, tf.abs(var))
                mask = tf.cast(tf.logical_not(small_weights), tf.float32)
                grad = grad * mask

            grads_and_vars_sparse.append((grad, var))

        return grads_and_vars_sparse