Question

TensorFlow的C ++接口似乎没有重塑方法。有没有人知道如何转换，例如[A,B,C,D]进入[A*B,C,D]？看起来这样做的唯一方法就是使用Eigen？但是，那里的文档非常简洁，代码是模板地狱，不容易解析。

Answer 1

检查重构张量是否具有相同数量的源张量元素的解决方案：

// Extracted image features from MobileNet_224
tensorflow::Tensor image_features(tensorflow::DT_FLOAT,
                                  tensorflow::TensorShape({1, 14, 14, 512}));

tensorflow::Tensor image_features_reshaped(tensorflow::DT_FLOAT,
                                           tensorflow::TensorShape({1, 196, 512}));

// Reshape tensor from [1, 14, 14, 512] to [1, 196, 512]
if(!image_features_reshaped.CopyFrom(image_features, tensorflow::TensorShape({1, 196, 512})))
{
  LOG(ERROR) << "Unsuccessfully reshaped image features tensor [" << image_features.DebugString() << "] to [1, 196, 512]";
  return false;
}

LOG(INFO) << "Reshaped features tensor: " << image_features_reshaped.DebugString();

Answer 2

这应该有效：

Tensor my_tensor; // [A, B, C, D]
Tensor reshaped_tensor = my_tensor.shaped<float, 3>({A*B, C, D});  //[A*B, C, D]

用C ++重塑张量

2 个答案: