Created by: lyp2github

为使您的问题得到快速解决，在建立Issue前，请您先通过如下方式搜索是否有相似问题:【搜索issue关键字】【使用labels筛选】【官方文档】

建立issue时，为快速解决问题，请您根据使用情况给出如下信息：

• 标题：dygraph模式下，fluid.gradients 报错
• 版本、环境信息：    1）PaddlePaddle版本：1.6.1.post97 GPU版本    3）GPU：Tesla P40 cuDNN Version: 7.0,    4）系统环境：CentOS release 6.3 (Final)，Python3.6, 2018 x86_64 x86_64 x86_64 GNU/Linux
• 模型信息 #!/usr/bin/env python #-- coding=utf8 -- import paddle.fluid as fluid import numpy as np

if name == 'main': with fluid.dygraph.guard(): rng = np.random.RandomState(0) inp_ = rng.uniform(-1, 1, [3, 2]).astype('float32') w1_ = rng.uniform(-1, 1, [2, 5]).astype('float32') w2_ = rng.uniform(-1, 1, [5, 4]).astype('float32') inp = fluid.dygraph.base.to_variable(inp_) w1 = fluid.dygraph.base.to_variable(w1_) w2 = fluid.dygraph.base.to_variable(w2_) inp.stop_gradient = False w1.stop_gradient = False w2.stop_gradient = False y = fluid.layers.mul(fluid.layers.mul(inp, w1), w2) x = [w1, w2] dfdx_f1 =fluid.gradients(y, x) print(dfdx_f1)

####################### Traceback (most recent call last): File "test_torch_dygraph.py", line 51, in test_mul_grad_paddle_vs_torch() File "test_torch_dygraph.py", line 37, in test_mul_grad_paddle_vs_torch dfdx_f1 =fluid.gradients(f, x) File "/home/users/liuyaping/local/python36/lib/python3.6/site-packages/paddle/fluid/backward.py", line 1350, in gradients outs = calc_gradient(targets, inputs, target_gradients, no_grad_set) File "/home/users/liuyaping/local/python36/lib/python3.6/site-packages/paddle/fluid/backward.py", line 1299, in calc_gradient prog._sync_with_cpp() File "/home/users/liuyaping/local/python36/lib/python3.6/site-packages/paddle/fluid/framework.py", line 4345, in _sync_with_cpp block._sync_with_cpp() File "/home/users/liuyaping/local/python36/lib/python3.6/site-packages/paddle/fluid/framework.py", line 2579, in _sync_with_cpp op = Operator(self, op_desc) File "/home/users/liuyaping/local/python36/lib/python3.6/site-packages/paddle/fluid/framework.py", line 1721, in init "type to initialized an Operator can not be None.") ValueError: type to initialized an Operator can not be None.