A Question About Weight Format

[zyt1024]

Hello! In the implementation of pointwise_conv_2d_cl in hls4ml, for a 1x1 convolutional kernel, is the weight format (Cout, Cin) or( Cin, Cout) in storage? I think is (Cout,Cin), is right?

[vloncar]

Depends on the Strategy, in case of latency strategy for 1x1 kernel it is (Cin, Cout) but for resource it is (Cout, Cin). More generally, for a kernel_height x kernel_width Conv2D the layout of weights is (kernel_height, kernel_width, Cin, Cout) for latency strategy, and (Cout, kernel_height, kernel_width, Cin) for resource strategy.

[zyt1024]

ok,Thanks,I understand.