dzy
dzy
“分不太清楚，哪个是bnn代码哪个是xnor代码”，融在一起了，都在 util_wbwtab.py(https://github.com/666DZY666/Model-Compression-Deploy/blob/master/compression/quantization/WbWtAb/util_wbwtab.py) 里。 “xnort论文提及是引入特征因子”，缩放因子在这里: https://github.com/666DZY666/Model-Compression-Deploy/blob/3959f194033a520d40fca4c2758874681981ea3c/compression/quantization/WbWtAb/util_wbwtab.py#L107 “取消A的二值，先只做W的二值”，--W 2 --A 32，这样设置。
1、wbwtab和wqaq 分别是 低位(三/二值)和高位量化，可参考 代码结构 示意图； 2、iao，https://arxiv.org/abs/1712.05877
已修复。
这里的压缩率目前是手动算的，BN参数这里做了融合，所以不需要计算，只算卷积参数就行。量化模型的大小是（模型参数量 * 每个参数bits / 8）Bytes，再用这个和之前的未量化模型比较得出压缩率
这里，https://github.com/666DZY666/micronet/commit/45136477a61528f171351f70763d5334cb31aebe#diff-9673a485be40841c87237bb5f5dc0b0e718116160d9898089e9cefb662679524R24
可以参考add量化自己加一下concat量化。concat量化后应该不会对网络性能有太大影响。
> AIMET currently supports OPSET 12 and QNN supports OPSET 11. The minimum version of onnx required for the GridSample is opset16，qnn supports. But how does aimet export the GridSample?
> @666DZY666 the references are LSQ (https://arxiv.org/abs/1902.08153) and LSQ+ (https://arxiv.org/abs/2004.09576 but AIMET optimizes min/max instead of scale offset so our implementation is not exactly the same. Why does aimet optimize...
When will torch1.13 be supported?
> Hi @666DZY666 it should be supported hopefully soon, but we do not have an ETA yet. Thanks, looking forward to support, 1.9.1 is prone to incompatibility with task code.