DA-ActNN-YOLOV5: Hybrid YOLO v5 Model with Data Augmentation and Activation of Compression Mechanism for Potato Disease Identification

Algorithm 1

Algorithm description of ActNN compression process. The training process uses minibatch random gradient descent training.

	Input: training data: X
(1)	Convert compressible model: Replace the model module with ActNN
(2)	Initialize Data Loader: Create mini-batch data loaders based on X
(3)	Initialize Model: Load model parameters or initialize model parameters, Let W be the model parameters
(4)	for number of training iterations do
(5)	for (data, target) of mini-batch data loader do
(6)	Neural network models complete a computational process based on data
	Compress the activation (CA) value parameter according to the data in forwarding propagation
	Calculate the gradient based on the activation parameter in backpropagation decompression again (DCA)
	Calculate the gradient that satisfies (4)
(7)	Update the model parameter W based on data
(8)	end for
(9)	end for
	Output: Model and model parameters W