I’m thinking what feature is in image recognition
Machine is checking rgb gradient among pixels in convolution neural network of PyTorch ?
There is method such as SIFT（Scale-Invariant Feature Transform) and HOG（Histograms of Oriented Gradients)
Which one is used in PyTorch ?
No, while the “RGB” representation could be valid for the input to a CNN, the intermediate activations are created by the conv layers and have multiple channels which cannot be mapped to the colorspace again.
Neither, as the conv layers are using a trainable conv kernel and perform a convolution (or rather a cross-correlation) using the input activation.
I would recommend to take a look at e.g. CS231n as it explains the concepts of CNNs pretty well.