FPGA-Based Real-Time Embedded Fish Embryo Detection System

Algorithm 1

Quantization algorithm for each layer.

	Full precision inputs, weights and bias, outputs:
	Bit width: for inputs and weights, b = 4, for bias, b = 8
	Output: Y, predict number: Label
(1)	Calculate q(x), q(w), q(bias) according to equation (6)
(2)	Y = Forward(q(x), q(w), q(bias))
(3)	Calculate losses and regularize all learnable parameters: L = loss (, Label)
(4)	According to Equation (8), calculate gradient:
(5)	Use the Adam optimizer [31] to update a portion of the full precision parameters according to Equation (9); W = Backward (W, Adam, )
(6)	Returning to 1.