training.py 文件源码

def clip_gradients_by_norm(grads_and_vars, add_to_summary=True):
    if add_to_summary:
        for grad, var in grads_and_vars:
            if grad is not None:
                variable_summaries(grad, 'grad/{}'.format(var.name[:-2]))

    # Clip by norm. Grad can be null when not training some modules.
    with tf.name_scope('clip_gradients_by_norm'):
        grads_and_vars = [
            (
                tf.check_numerics(
                    tf.clip_by_norm(gv[0], 10.),
                    'Invalid gradient'
                ), gv[1]
            )
            if gv[0] is not None else gv
            for gv in grads_and_vars
        ]

    if add_to_summary:
        for grad, var in grads_and_vars:
            if grad is not None:
                variable_summaries(
                    grad, 'clipped_grad/{}'.format(var.name[:-2]))

    return grads_and_vars