Question

我正在用C ++为tensrflow编写新的操作是Linal.cc：

#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/shape_inference.h"
#include<vector>
using namespace tensorflow;

REGISTER_OP("Linal")
    //.Attr("preserve_index: int")
    .Input("x: float")
    .Input("a: float")
    .Output("y: float");

#include "tensorflow/core/framework/op_kernel.h"

using namespace tensorflow;

class Linl: public OpKernel {
 public:
  explicit Linal(OpKernelConstruction* context) : OpKernel(context) {}

  void Compute(OpKernelContext* context) override {
    //Get x
    const Tensor& input_tensor = context->input(0);
    //size_t size_of_x = input_tensor.shape().dim_size(1);
    auto input = input_tensor.flat<float>();
    size_t size_of_x = input.size();

    std::vector<float> x;
    for(int i = 0; i < size_of_x; ++i){
        x.push_back(input(i));
    }


    //Get a
    const Tensor& a_tensor = context->input(1);

    auto a_vector = a_tensor.flat<float>();
    float a = a_vector(0);


    // Create an output tensor
    Tensor* output_tensor = NULL;

    OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(),
                                                     &output_tensor));

    auto output_flat = output_tensor->flat<float>();

    for(int i = 0; i < size_of_x; ++i){
        output_flat(i) = a * x[i];
    }
  }
};

REGISTER_KERNEL_BUILDER(Name("Linal").Device(DEVICE_CPU), Linal);

我想检查自己的渐变是否像张量流渐变一样工作。它是testing.py

import numpy as np
import tensorflow as tf
import modul_gradient
from tensorflow.python.framework import ops

@tf.RegisterGradient("Linal")
def modgrad(op, grad):
    x = op.inputs[0] # the first argument (normally you need those to calculate the gradient, like the gradient of x^2 is 2x. )
    a = op.inputs[1] # the second argument
    return grad * a, x * grad


Linal_module = tf.load_op_library('./LUT.so')

with tf.Session() as sess:
    #get x and a
    x = tf.constant([0.2, 1.2, 6, 7, 8])
    a = tf.constant([3.2])

    #describe the calculation graph in tf
    product = LUT_module.lut(x, a)
    product_tf = tf.multiply(x, a)


    #calculate gradient for our op and default gradient
    gr = tf.gradients(product, [x, a])
    gr_tf = tf.gradients(product_tf, [x, a])

    tf.initialize_all_variables().run()

    #this radients should be equal
    print(gr[1].eval())
    print(gr_tf[1].eval())

但是我得到了ValueError：运算输入和计算的输入梯度之间的形状不兼容。正向操作：LUT。输入索引：1.原始输入形状：（1，）。计算出的输入梯度形状：（5，）

请你帮我一下。我真的不明白怎么了

ValueError：运算输入和计算的输入梯度之间的形状不兼容

0 个答案: