Created by: Gaffey

• 版本、环境信息：    1）PaddlePaddle版本：1.5post97    3）GPU：cuda 9.0 cudnn 7.1

• 训练信息    1）单机/单卡

• 复现信息： 一段可复现的简单代码

import os

from PIL import Image
import paddle
from paddle import fluid
from paddle.fluid.layer_helper import LayerHelper

import numpy as np

class CIFAR(fluid.dygraph.Layer):
    def __init__(self, name_scope):
        super(CIFAR, self).__init__(name_scope)
        self._conv1 = fluid.dygraph.Conv2D(self.full_name(), 64, 3, 1, 1, act=None)
        self._conv2 = fluid.dygraph.Conv2D(self.full_name(), 64, 3, 1, 1, act=None)
        self.global_pooling = fluid.dygraph.Pool2D(self.full_name(), 32, "avg", 32, 0, True)
        #scale = (2.0 / (512**2*10))**0.5
        self._fc = fluid.dygraph.FC(self.full_name(),
                                    10,
                                    param_attr=fluid.param_attr.ParamAttr(
                                        initializer=fluid.initializer.NormalInitializer(
                                            loc=0.0, scale=0.01)),
                                    act="softmax")

    def forward(self, inputs):
        x = self._conv1(inputs)
        x1 = self._conv2(inputs)
        x = self.global_pooling(x)
        x = self._fc(x)
        return x

def train(train_reader, test_reader, model):
    optimizer = fluid.optimizer.MomentumOptimizer(learning_rate=0.1, momentum=0.9, regularization=fluid.regularizer.L2DecayRegularizer(5e-4))
    for epoch in range(100):
        acc_list = []
        model.train()
        for batch_id, data in enumerate(train_reader()):
            dy_x_data = np.array([x[0].reshape(3, 32, 32)
                                  for x in data]).astype('float32')
            y_data = np.array(
                    [x[1] for x in data]).astype('int64').reshape(-1, 1)

            img = fluid.dygraph.to_variable(dy_x_data)
            label = fluid.dygraph.to_variable(y_data)
            label.stop_gradient = True
            prediction = model.forward(img)
            loss = fluid.layers.cross_entropy(prediction, label)
            avg_loss = fluid.layers.mean(loss)

            avg_loss.backward()

            optimizer.minimize(avg_loss)
            # save checkpoint
            model.clear_gradients()

def main():
    with fluid.dygraph.guard():
        cifar = CIFAR("cifar10")
        test_reader = paddle.batch(
                        paddle.dataset.cifar.test10(), batch_size=128, drop_last=True)

        train_reader = paddle.batch(
                        paddle.dataset.cifar.train10(),
                        batch_size=128,
                        drop_last=True)
        train(train_reader, test_reader, cifar)

if __name__ == "__main__":
    main()

• 问题描述： 该代码会报错

  File "simple_test.py", line 68, in <module>
    main()
  File "simple_test.py", line 65, in main
    train(train_reader, test_reader, cifar)
  File "simple_test.py", line 51, in train
    optimizer.minimize(avg_loss)
  File "<decorator-gen-20>", line 2, in minimize
  File "/ssd1/wenshuo/anaconda2/lib/python2.7/site-packages/paddle/fluid/wrapped_decorator.py", line 25, in __impl__
    return wrapped_func(*args, **kwargs)
  File "/ssd1/wenshuo/anaconda2/lib/python2.7/site-packages/paddle/fluid/dygraph/base.py", line 88, in __impl__
    return func(*args, **kwargs)
  File "/ssd1/wenshuo/anaconda2/lib/python2.7/site-packages/paddle/fluid/optimizer.py", line 600, in minimize
    loss, startup_program=startup_program, params_grads=params_grads)
  File "/ssd1/wenshuo/anaconda2/lib/python2.7/site-packages/paddle/fluid/optimizer.py", line 556, in apply_optimize
    self.regularization)
  File "/ssd1/wenshuo/anaconda2/lib/python2.7/site-packages/paddle/fluid/regularizer.py", line 79, in append_regularization_ops
    outputs={"Out": new_grad})
  File "/ssd1/wenshuo/anaconda2/lib/python2.7/site-packages/paddle/fluid/framework.py", line 1739, in append_op
    kwargs.get("stop_gradient", False))
  File "/ssd1/wenshuo/anaconda2/lib/python2.7/site-packages/paddle/fluid/dygraph/tracer.py", line 59, in trace_op
    framework._current_expected_place(), stop_gradient)
paddle.fluid.core_avx.EnforceNotMet: holder_ should not be null
Tensor not initialized yet when Tensor::place() is called. at [/paddle/paddle/fluid/framework/tensor.h:133]

如果注释掉x1则错误消失。猜测原因是x1实际上没有连接到最终的loss上，所以没有进行反向。在minimize中没有判断grad是否为空，所以出现了未初始化的错误。建议修改错误提示或者在minimize时判断该op是否需要update。在进行网络结构的动态调整时可能会出现类似的需求。