Created by: luwanglin

环境：aistudio 硬件信息 当前环境高级版 切换环境 CPU 8 RAM 32GB GPU v100 显存 16GB 磁盘 100GB 环境配置 Python版本 python3.7 框架版本 PaddlePaddle 1.7.1

1. 模型实现如下，自己实现的LRN算子

LRN 的实现

class LocalResponseNorm(fluid.dygraph.Layer):
    __constants__ = ['size', 'alpha', 'beta', 'k']

    def __init__(self, size, alpha=1e-4, beta=0.75, k=1.):
        super(LocalResponseNorm, self).__init__()
        self.size = size
        self.alpha = alpha
        self.beta = beta
        self.k = k

    def forward(self, input):
        dim = len(input.shape)
        if dim < 3:
            raise ValueError('Expected 3D or higher dimensionality \
                         input (got {} dimensions)'.format(dim))
        div = fluid.layers.unsqueeze(input * input, axes=1)
        if dim == 3:
            div = fluid.layers.pad(div, (0, 0, 0, 0, self.size // 2, (self.size - 1) // 2, 0, 0, 0, 0))
            div = fluid.layers.pool2d(div, pool_size=(self.size, 1), pool_stride=1, pool_type='avg')
            div_shape = div.shape
            div_shape.pop(1)
            div = fluid.layers.reshape(div, shape=div_shape)
        else:
            sizes = input.shape
            div = fluid.layers.reshape(div, (sizes[0], 1, sizes[1], sizes[2], -1))
            div = fluid.layers.pad(div, (0, 0, 0, 0, self.size // 2, (self.size - 1) // 2, 0, 0, 0, 0))
            div = fluid.layers.pool3d(div, (self.size, 1, 1), pool_stride=1, pool_type='avg')
            div_shape = div.shape
            div_shape.pop(1)
            div = fluid.layers.reshape(div, shape=div_shape)
            div = fluid.layers.reshape(div, sizes)
        div = fluid.layers.pow((div * self.alpha + self.k), factor=self.beta)
        # print('final', div.shape)
        return input / div


class LBNet_highwayThree(fluid.dygraph.Layer):  # SHT网络
    def __init__(self, nc=2):
        super(LBNet_highwayThree, self).__init__()

        self.convolutions1 = fluid.dygraph.Sequential(
            nn.Conv2D(nc, 16, filter_size=7, stride=1),
            nn.BatchNorm(16, act='relu'),

            nn.Conv2D(16, 16, filter_size=1, stride=1),
            nn.BatchNorm(16, act='relu'),

            nn.Conv2D(16, 16, filter_size=3, stride=1, padding=1),
            nn.BatchNorm(16, act='relu'),
            nn.Conv2D(16, 16, filter_size=1, stride=1)
        )

        self.high_way = fluid.dygraph.Sequential(
            nn.BatchNorm(2, act='relu'),
            nn.Conv2D(2, 16, filter_size=7, stride=1),
        )

        self.convolutions2 = fluid.dygraph.Sequential(
            nn.BatchNorm(16, act='relu'),
            LocalResponseNorm(5, 0.0001, 0.75, 2),
            nn.Pool2D(pool_size=2, pool_stride=2),
            nn.Conv2D(16, 64, filter_size=7, stride=1),
            nn.BatchNorm(64, act='relu'),
            LocalResponseNorm(5, 0.0001, 0.75, 2),
            nn.Pool2D(pool_size=2, pool_stride=2),

            nn.Conv2D(64, 256, filter_size=7, stride=1)
        )
        self.mlp = fluid.dygraph.Sequential(
            nn.Linear(21 * 21 * 256, 1)
        )

    def forward(self, x):
        x1 = self.convolutions1(x)
        x2 = self.high_way(x)
        x = x1 + x2
        x = self.convolutions2(x)
        x = fluid.layers.reshape(x, (-1, 21 * 21 * 256))
        x = fluid.layers.relu(x)
        x = fluid.layers.dropout(x, dropout_prob=0.5)
        x = self.mlp(x)
        return x

2.复现信息，前向传播和反向传播测试如下：

if __name__ == '__main__':
    test_data = np.ones((2, 2, 126, 126)).astype('float32')
    label = np.array([[0], [1]]).astype('float32')
    with fluid.dygraph.guard():
        lb = LBNet_highwayThree()
        # lrn = LocalResponseNorm(5, 0.0001, 0.75, 2)
        x = fluid.dygraph.base.to_variable(test_data)
        label = fluid.dygraph.base.to_variable(label)

        output = lb(x)
        print(output)
        loss = fluid.layers.sigmoid_cross_entropy_with_logits(output, label)
        print('loss:', loss)
        loss.backward()

3. 问题描述 前向传播能够正常进行，但是反向传播计算梯度时报错

Traceback (most recent call last):
  File "work/SHT/model.py", line 104, in <module>
    loss.backward()
  File "</opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/decorator.py:decorator-gen-60>", line 2, in backward
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/wrapped_decorator.py", line 25, in __impl__
    return wrapped_func(*args, **kwargs)
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py", line 207, in __impl__
    return func(*args, **kwargs)
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/varbase_patch_methods.py", line 116, in backward
    self._run_backward(backward_strategy, framework._dygraph_tracer())
paddle.fluid.core_avx.EnforceNotMet:


Error Message Summary:
----------------------
Error:  at (/paddle/paddle/fluid/operators/batch_norm_op.cc:445)