静态图实现对抗训练（FGM），梯度运算过程出现[operator < elementwise_add > error]错误，交换加法算符位置可解决

[xwen99]

• 问题：
  • 静态图实现对抗训练（FGM），梯度运算过程出现[operator < elementwise_add > error]错误
  • 交换加法算符位置可解决
• 版本、环境信息：
  • PaddlePaddle版本：1.8.1
  • PaddleHub版本：1.7.1
  • GPU：v100
  • 系统环境：AI Studio 环境，Python 3.7
• 训练信息
  • 单机单卡

• 复现信息：

  • 相关代码：

    
    def _scale_l2(x, norm_length):
    # shape(x) = (batch, num_timesteps, d)
    # Divide x by max(abs(x)) for a numerically stable L2 norm.
    # 2norm(x) = a * 2norm(x/a)
    # Scale over the full sequence, dims (1, 2)
    alpha = fluid.layers.reduce_max(fluid.layers.abs(x), dim=[1, 2], keep_dim=True) + 1e-12
    l2_norm = alpha * fluid.layers.sqrt(
        fluid.layers.reduce_sum(fluid.layers.pow(x / alpha, 2), dim=[1, 2], keep_dim=True) + 1e-6)
    x_unit = x / l2_norm
    return norm_length * x_unit

  def adversarial_loss(self, loss): grad = fluid.backward.append_backward(loss, [self.feature.name])[0][1] perturb = _scale_l2(grad, 1.0) adv_loss = self.cl_loss_from_embedding(self.feature + perturb) # 报错在这里

  如果交换位置，变成 perturb + self.feature 就不报错

  return adv_loss

对抗损失 adv_loss 在 build_env 时被加到正常的损失上：

    def _build_env(self):
        """
        building the program and strategy for specific running phase.
        """
        if self.env.is_inititalized:
            return

        self._build_env_start_event()
        self.env.is_inititalized = True
        self.env.main_program = clone_program(
            self._base_main_program, for_test=False)

        self.env.startup_program = fluid.Program()
        with fluid.program_guard(self.env.main_program,
                                 self._base_startup_program):
            with fluid.unique_name.guard(self.env.UNG):
                self.env.outputs = self._build_net()
                if self.is_train_phase or self.is_test_phase:
                    self.env.labels = self._add_label()
                    self.env.loss = self._add_loss()
                    self.env.metrics = self._add_metrics()

        if self.is_predict_phase or self.is_test_phase:
            self.env.main_program = clone_program(
                self.env.main_program, for_test=True)
            hub.common.paddle_helper.set_op_attr(
                self.env.main_program, is_test=True)

        if self.is_train_phase:
            with fluid.program_guard(self.env.main_program,
                                    self._base_startup_program):
                with fluid.unique_name.guard(self.env.UNG):
                    self.env.adv_loss = self.adversarial_loss(self.loss)
                    self.env.loss += self.env.adv_loss # 这里
        ...

• 问题描述：
  • 报错信息：

    
    Traceback (most recent call last):
    File "reading_comprehension.py", line 104, in <module>
    reading_comprehension_task.finetune_and_eval()
    File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 868, in finetune_and_eval
    return self.finetune(do_eval=True)
    File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 883, in finetune
    self.init_if_necessary()
    File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 367, in init_if_necessary
    if not self.load_checkpoint():
    File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 834, in load_checkpoint
    main_program=self.main_program)
    File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 508, in main_program
    self._build_env()
    File "/home/aistudio/work/DuReader-Robust-With-Paddlehub/my_reading_comprehension_task.py", line 563, in _build_env
    self.env.adv_loss = self.adversarial_loss(self.loss)
    File "/home/aistudio/work/DuReader-Robust-With-Paddlehub/my_reading_comprehension_task.py", line 520, in adversarial_loss
    adv_loss = self.cl_loss_from_embedding(self.feature + perturb)
    File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py", line 242, in __impl__
    attrs={'axis': axis})
    File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py", line 2610, in append_op
    attrs=kwargs.get("attrs", None))
    File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py", line 1971, in __init__
    self.desc.infer_shape(self.block.desc)
    paddle.fluid.core_avx.EnforceNotMet: 

---

C++ Call Stacks (More useful to developers):

0 std::string paddle::platform::GetTraceBackString<std::string const&>(std::string const&, char const, int) 1 paddle::platform::EnforceNotMet::EnforceNotMet(std::string const&, char const, int) 2 paddle::operators::ElementwiseOp::InferShape(paddle::framework::InferShapeContext*) const 3 paddle::framework::OpDesc::InferShape(paddle::framework::BlockDesc const&) const

---

Python Call Stacks (More useful to users):

File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py", line 2610, in append_op attrs=kwargs.get("attrs", None)) File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py", line 242, in impl attrs={'axis': axis}) File "/home/aistudio/work/DuReader-Robust-With-Paddlehub/my_reading_comprehension_task.py", line 520, in adversarial_loss adv_loss = self.cl_loss_from_embedding(self.feature + perturb) File "/home/aistudio/work/DuReader-Robust-With-Paddlehub/my_reading_comprehension_task.py", line 563, in _build_env self.env.adv_loss = self.adversarial_loss(self.loss) File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 508, in main_program self._build_env() File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 834, in load_checkpoint main_program=self.main_program) File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 367, in init_if_necessary if not self.load_checkpoint(): File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 883, in finetune self.init_if_necessary() File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlehub/finetune/task/base_task.py", line 868, in finetune_and_eval return self.finetune(do_eval=True) File "reading_comprehension.py", line 104, in reading_comprehension_task.finetune_and_eval()

---

Error Message Summary:

NotFoundError: No Input(Y) found for ElementwiseOp operator. [Hint: Expected ctx->HasInput("Y") == true, but received ctx->HasInput("Y"):0 != true:1.] at (/paddle/paddle/fluid/operators/elementwise/elementwise_op.h:42) [operator < elementwise_add > error]

[xwen99]

复现环境：https://aistudio.baidu.com/aistudio/projectdetail/445781

[jerrywgz]

可以先尝试升级至paddle2.0.0.a0，目前1.8版本中已知elementwise的操作可能会出现问题，在新版本已经修复

[paddle-bot-old[bot]]

Since you haven\'t replied for more than a year, we have closed this issue/pr. If the problem is not solved or there is a follow-up one, please reopen it at any time and we will continue to follow up. 由于您超过一年未回复，我们将关闭这个issue/pr。 若问题未解决或有后续问题，请随时重新打开，我们会继续跟进。