Created by: wozna

This squash is based on PR https://github.com/PaddlePaddle/Paddle/pull/18754 This squash change chain with conv op. and dequantize op. to conv with forced float output. Thanks to this you will avoid additional conversion to int8.

In mobilenet ssd there are 12 pattern like this and the use of this squash improves accuracy (on clx machine): - Before squash: acc_fp32 = 0.7733, acc_int8=0.77 - After that squash: acc_fp32 = 0.7733, acc_int8=0.7719

Unfortunately, after a few tests using mkl-dnn and int8 on clx machine, fps decreased: - Before squash: avg fps = 151.634 - After squash: avg fps = 146.8749 Is's almost 3%. That could be some kind of option if accuracy is more important than performance.