1. The bgr turn rgb
The default is read into opencv bgr order, to be processed into rgb sequence.
Python treatment:
img = img[:, :, (2, 1, 0)]
C ++ handling:
cv::cvtColor(testimg, testimg, CV_BGR2RGB);
2. The image matrix transformation
1) opencv read image matrix format is: (height, width, channels), the channel last is a three-dimensional matrix, i.e. (height, width, channel).
In depth study, because of the different channels convolution, it will take another way: channel first, namely (channels, height, width). To meet this requirement, you can do so:
print(img.shape)
img = img.transpose(2,0,1)
print(img.shape)
Output:
（480,640,3）
（3,480,640）
2) learning in the depth of the structures CNN, often as the corresponding image data processing, such as expansion to the image dimensions, such as extended (batch_size, channels, height, width). For this requirement, you can do so:
img = np.expand_dims(img, axis=0)
print(img.shape)
Output:
（1,3,480,640）
C ++ to the corresponding code:
auto img_tensor = torch::from_blob(testimg.data, { 1,480,640, 3 }, torch::kFloat32);
img_tensor = img_tensor.permute({ 0,3,1,2 });
3. c ++ model forward return value
The return type for the torch :: jit :: IValue
torch::jit::IValue result = module->forward(inputs);
If only one return value can be transferred directly tensor:
auto outputs = module->forward(inputs).toTensor();
Note that if there are multiple return values, you need to turn tuple:
auto outputs = module->forward(inputs).toTuple();
torch::Tensor out1 = outputs->elements()[0].toTensor();
torch::Tensor out2 = outputs→elements()[1].toTensor();
4. Use GPU
The model and inputs are put on the gpu:
std::shared_ptr<torch::jit::script::Module> module = torch::jit::load(argv[2]);
//put to cuda
module->to(at::kCUDA);
// Note that the tensor into the gpu, rather than vector <torch :: jit :: IValue>
std::vector<torch::jit::IValue> inputs;
image_tensor.to(at::kCUDA)
inputs.push_back(image_tensor)
You can specify GPU id: to (torch :: Device (torch :: kCUDA, id))
