MindSpore Lite优化模型(训练后量化开启全量化提示“Cant find input:input”

最新推荐文章于 2022-07-20 16:24:58 发布

小乐快乐

最新推荐文章于 2022-07-20 16:24:58 发布

阅读量475

点赞数

文章标签： python 深度学习计算机视觉

本文链接：https://blog.csdn.net/weixin_45666880/article/details/125265796

版权

问题描述：

使用MindSpore Lite模型转换工具，配置全量化（FULL_QUANT）之后，会报错。

配置信息

converter_lite --fmk=ONNX --modelFile=RetinaFace.onnx --outputFile=RetinaFace_full_quant --inputDataFormat=NCHW --fp16=off --configFile=full_quant_params.ini

full_quant_params.ini配置如下

[common_quant_param]

quant_type=FULL_QUANT

[data_preprocess_param]

calibrate_path=input_name_1:./calibrate_input

calibrate_size=30

input_type=IMAGE

normalize_mean=[104, 117, 123]

resize_width=320

resize_height=320

resize_method=LINEAR

错误信息如下

WARNING: Logging before InitGoogleLogging() is written to STDERR

[WARNING] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\quantizer\quantize_util.cc:131] CanOpFullQuantized] node:Slice_647 type:StridedSlice is not support quantization.

[WARNING] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\quantizer\quantize_util.cc:131] CanOpFullQuantized] node:Resize_652 type:Resize is not support quantization.

[WARNING] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\quantizer\quantize_util.cc:131] CanOpFullQuantized] node:Slice_666 type:StridedSlice is not support quantization.

[WARNING] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\quantizer\quantize_util.cc:131] CanOpFullQuantized] node:Resize_671 type:Resize is not support quantization.

[WARNING] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\quantizer\quantize_util.cc:131] CanOpFullQuantized] node:Softmax_823 type:Softmax is not support quantization.

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\preprocess\image_preprocess.cc:167] PreProcess] Cant find input:input

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\preprocess\image_preprocess.cc:129] PreProcess] data_buffer is nullptr or size == 0

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\src\common\tensor_util.cc:244] CheckTensorsInvalid] Graph input tensor data is nullptr input

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\src\lite_session.cc:679] RunGraph] CheckInputs failed.

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\quantizer\full_quant_quantizer.cc:990] DoInference] run model failed!

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\anf_transform.cc:391] DoSingleGraphQuantize] DoQuantization failed -1

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\anf_transform.cc:405] DoQuantize] Do Quantize failed.

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\anf_transform.cc:499] TransformFuncGraph] Do Quantize failed.

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\anf_transform.cc:538] Transform] optimizer failed.

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\converter.cc:110] Convert] Transform anf graph return nullptr

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\tools\converter\converter.cc:170] RunConverter] CONVERT RESULT FAILED:-2 NULL pointer returned.

CONVERT RESULT FAILED:-2 NULL pointer returned.

[ERROR] LITE(190516,1,?):2022-3-14 18:7:50 [mindspore\lite\src\lite_session.cc:794] ~LiteSession] Not support multi-threading

请问这里的“Cant find input:input”是找不到模型的输入吗？还是什么呢，希望知晓的朋友给予解答。

解答：

配置文件calibrate_path应该改为模型的输入的名字input