Correct MultipleQuantizeSquash #40717
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
@wozna Hi, Does all daily CI models could all pass? Or the coverage CI should test all the models? |
| @@ -45,6 +45,7 @@ void LogCannotQuantizeOp(Node* op, const char* details = nullptr) { | |||
| << " (type: " << op->Op()->Type() << ", id: " << op->id() << ")."; | |||
| if (details) msg_ss << " " << details; | |||
| PrettyLogDetail(msg_ss.str().c_str()); | |||
| op->Op()->SetAttr("mkldnn_data_type", std::string("float32")); | |||
There was a problem hiding this comment.
Later on, could we refactor these log functions? It looks VLOG(4) is enough for these log functions?
There was a problem hiding this comment.
Does function below void LogQuantizationDisabled(Node* op) maybe also need to set "mkldnn_data_type", std::string("float32") ?
|
Hi, @jczaja please review. This PR need one more approval to merge. Thanks |
|
你的PR已合入Paddle库，请关注后续测试结果。 |
* Correct MultipleQuantizeSquash * Correct logging
PR types
Bug fixes
PR changes
Others
Describe
This fixed the problem with running the yolov3_darknet model.
This PR corrects MultipleQuantizeSquash in
cpu_quantize_squash_pass.cc.It turned out that this mechanism didn't support the situation when under one input name are many inputs (eg. concat, sum).
Before new inputs for the next operators were created, now we just updating the next operator input.
Moreover, I added setting
mkldnn_data_typeto float32 if the operator is not quantised, because before even if the operator wasn't quantized this attribute was set to int8.