现在采用asymmetric_quantized-u8/dynamic_fixed_point-8, 精度损失7%,无法满足产品需求。所以尝试混合量化,大概知道是哪个layer精度损失大,但是按照手册上操作后,会有这样的提示“W Hybrid model reset tensor "@conv2_4:weight" dtype to "float32"
W Hybrid model reset tensor "@conv2_4:bias" dtype to "float32"”作者: jefferyzhang 时间: 2019-10-18 12:04