提交 810a33c8 编写于 作者: Y Youwei Song 提交者: Jiabin Yang

fix dygraph partitial backward problem (#19625)

* fix dygraph partitial backward problem, test=develop

* add unittest, fix ClearGradient. test=develop

* add filter and error in python side, test=develop

* rebase develop, test=develop

* bug fix for list equals in py3.5, test=develop

* bug fix for list equals, test=develop
上级 ec9bc1bd
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import print_function
import unittest
import paddle.fluid as fluid
import numpy as np
class TestImperativePartitialBackward(unittest.TestCase):
def test_partitial_backward(self):
with fluid.dygraph.guard():
x = np.random.randn(2, 4, 5).astype("float32")
x = fluid.dygraph.to_variable(x)
fc1 = fluid.dygraph.FC("fc1", 10, num_flatten_dims=2)
fc2 = fluid.dygraph.FC("fc2", 10, num_flatten_dims=2)
y = fc1(x[:, :2])
z = fc2(x[:, 2:])
loss = fluid.layers.reduce_mean(y)
loss.backward()
for param in fc1.parameters():
self.assertIsNotNone(param._ivar._grad_ivar())
for param in fc2.parameters():
self.assertIsNone(param._ivar._grad_ivar())
optimizer = fluid.optimizer.AdamOptimizer()
_, params_grads = optimizer.minimize(loss)
self.assertListEqual(
sorted([p.name for p in fc1.parameters()]),
sorted([p_g[0].name for p_g in params_grads]))
fc1.clear_gradients()
fc2.clear_gradients()
if __name__ == '__main__':
unittest.main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册