disable_autofuse

Applicability

Product	Supported
Atlas A3 training products/Atlas A3 inference products	☓
Atlas A2 training products/Atlas A2 inference products	√
Atlas 200I/500 A2 inference products	☓
Atlas inference products	☓
Atlas training products	☓

Description

When a TensorFlow network runs on the CANN platform, multiple operators are automatically fused based on the graph structure to obtain better performance. If you do not want to perform automatic fusion on some operators, you can use this API to mark them.

Prototype

def disable_autofuse()

Parameters

None

Returns

None

Restrictions

The disable_autofuse API needs to be called using the with statement. Only operators in the corresponding scope are not automatically fused.

Example

import tensorflow as tf
from npu_bridge.npu_init import *

a = tf.placeholder(tf.float32, (5, 1))
b = tf.constant([1, 2, 3, 4, 5], dtype=tf.float32, shape=(5, 1))
add_0 = tf.add(a, b)
mul_0 = tf.multiply(add_0, b)

with disable_scope():
    # Operators in this scope are not automatically fused.
    abs_0 = tf.abs(mul_0)
    div_0 = tf.divide(a, abs_0)

with tf.Session() as sess:
    result = sess.run(div_0, feed_dict={a: [[1], [2], [3], [4], [5]]})
    print(result)

Parent topic: npu_bridge.estimator.npu.npu_scope