校准执行过程中提示“IfmrQuantCalibration with offset scale is illegal"或“ IfmrQuantCalibration without offset scale is illegal”

问题描述

在调用Caffe框架执行中间校准模型推理过程中，由于输入数据范围不合法，导致量化算法计算得到的scale不合理，从而校准过程失败，终止Caffe校准流程。

可能原因

数据范围 [-inf , +inf]：

因为AMCT的量化算法需要强制过零点，所以计算出的scale也就是inf/255=inf，该情况下量化因子后续无法承载，因此量化算法会提示错误信息，不支持该数据范围。

场景1

I0513 16:12:37.508546 80719 ifmr_layer.cu:58] Doing layer: "Convolution1" calibration, already store 1/1 data.
I0513 16:12:37.508558 80719 ifmr_layer.cu:65] Start to do ifmr quant.
[ERROR][ProcessScale][48] Not support scale is +inf.
[ERROR][IfmrQuantCalibration][338] IfmrQuantCalibration with offset scale is illegal.
[ERROR][IfmrQuantInternel][462] IfmrQuantInternel calibration failed.
F0513 16:12:37.535028 80719 ifmr_layer.cu:70] Check failed: ret == 0 (-65519 vs. 0) Do IFMR calibration failed
*** Check failure stack trace: ***
Aborted

场景2

I0513 16:14:48.173506 81053 ifmr_layer.cu:58] Doing layer: "Convolution1" calibration, already store 1/1 data.
I0513 16:14:48.173519 81053 ifmr_layer.cu:65] Start to do ifmr quant.
[ERROR][ProcessScale][48] Not support scale is +inf.
[ERROR][IfmrQuantCalibration][359] IfmrQuantCalibration without offset scale is illegal.
[ERROR][IfmrQuantInternel][462] IfmrQuantInternel calibration failed.
F0513 16:14:48.199769 81053 ifmr_layer.cu:70] Check failed: ret == 0 (-65519 vs. 0) Do IFMR calibration failed
*** Check failure stack trace: ***
Aborted

数据范围 $\text{[math]}$ ：（其中EPSILON包括DBL_EPSILON double类型，FLT_EPSILON float类型，当前使用的是FLT_EPSILON类型）

AMCT量化支持计算得到的最大 $\text{[math]}$ ，因为在昇腾AI处理器量化动作做的是乘法计算： $\text{[math]}$ ，如果scale大于 $\text{[math]}$ ， $\text{[math]}$ 会小于FLT_EPSILON，此时量化后结果就不可信。因此AMCT量化算法仅支持原始数据范围在 $\text{[math]}$ 内进行量化，否则会提示不支持并提示错误信息。

2021-05-11 09:25:42,222 - INFO - [AMCT]:[WeightsCalibrationPass]: Do layer 'Convolution1[arq_quantize]' weights calibration success!
2021-05-11 09:25:42,222 - INFO - [AMCT]:[Optimizer]: Do <class 'amct_caffe.optimizer.insert_ifmr_layer.InsertIFMRLayerPass'>
2021-05-11 09:25:42,223 - INFO - [AMCT]:[Optimizer]: Do <class 'amct_caffe.optimizer.lstm_calibration_replace.LSTMCalibrationReplacePass'>
2021-05-11 09:25:42,223 - INFO - [AMCT]:[Optimizer]: Do <class 'amct_caffe.optimizer.insert_searchn_layer.InsertSearchNLayerPass'>
2021-05-11 09:25:42,223 - INFO - [AMCT]:[Graph]: Doing whole model dump...
2021-05-11 09:25:42,225 - INFO - [AMCT]:[Utils]: The weights_file is saved in xx/data/AMCT15_CAFFE_GPU_GPU_FAQ_LEVEL1_TC_001/AMCT15_CAFFE_GPU_GPU_FAQ_LEVEL1_TC_001_tmp.caffemodel
2021-05-11 09:25:42,226 - INFO - [AMCT]:[Utils]: The model_file is saved in xx/data/AMCT15_CAFFE_GPU_GPU_FAQ_LEVEL1_TC_001/AMCT15_CAFFE_GPU_GPU_FAQ_LEVEL1_TC_001_tmp.prototxt
[ERROR][ProcessScale][52] Not support scale greater than 1 / FLT_EPSILON.
[ERROR][IfmrQuantCalibration][338] IfmrQuantCalibration with offset scale is illegal.
[ERROR][IfmrQuantInternel][462] IfmrQuantInternel calibration failed.
WARNING: Logging before InitGoogleLogging() is written to STDERR
F0511 09:25:42.292863 80432 ifmr_layer.cu:70] Check failed: ret == 0 (-65519 vs. 0) Do IFMR calibration failed
*** Check failure stack trace: ***
Aborted

处理建议

根据提示信息，跳过日志中的量化层，例如上述提示信息中的Convolution1层。

父主题： FAQ