jefferyzhang 发表于 2019-11-18 17:55
参看demo,这里填的是你要送进去量化的图片的相对路径
- RKNN output shape(reshape): (1 1 1 512)
- D Real output shape: (1, 1, 1, 512)
- D Process InceptionResnetV1/Bottleneck/BatchNorm/cond/FusedBatchNorm_1_3 ...
- D RKNN output shape(batchnormalize): (1 1 1 512)
- D Real output shape: (1, 1, 1, 512)
- D Process InceptionResnetV1/Bottleneck/BatchNorm/Reshape_1_2 ...
- D RKNN output shape(reshape): (1 512)
- D Real output shape: (1, 512)
- D Process attach_InceptionResnetV1/Bottleneck/BatchNorm/Reshape_1/out0_0 ...
- D RKNN output shape(output): (1 512)
- D Real output shape: (1, 512)
- I Build free_grah complete.
- I Generated network graph with 1 outputs.
- I @attach_InceptionResnetV1/Bottleneck/BatchNorm/Reshape_1/out0_0:out0: (1, 512)
- D Init coefficients ...
- I Start tensor porvider ...
- E File Not Found: .\accury_test\test
- I Runing 1 epochs, algorithm: normal
- I iterations: 0
- W No more data in provider.
- [TRAINER]Quantization fail.
- I Clean.
- E Quantization data is None, quantize may failed, please check log!
- done
- E RKNN model data is None, please load model first!
- (rknn_win_env) E:\usb_test\example\yolov3>call:deactivate
- 请按任意键继续. . .
复制代码
jefferyzhang 发表于 2019-11-19 16:22
E File Not Found: .\accury_test\test
你datasets.txt里填的是什么内容?
jefferyzhang 发表于 2019-11-20 09:06
路径是要写到图片的,多少图片就要多少行
liuwenhua 发表于 2019-11-20 10:02
根据你说的测试可以了,但是不明白做这一步的用途是什么?现在我在测试facenet模型转换过程时,不支持多 ...
jefferyzhang 发表于 2019-11-20 10:34
用途就是量化。 量化是一门技术,你可以自行百度下,一两句话说不清楚。
TF也支持训练中量化,我们rknn也 ...
liuwenhua 发表于 2019-11-20 16:38
多谢,我是用rknn 1.2.x版本可以正常推理,但遇到一个问题,没有进行量化得模型,推理结果正常,量化后得 ...
- spend_time 0.4040415287017822
- dist 0.9188065
- 0001.jpg and 000_0.jpg distance is 0.9188065
- dist 0.6297745
- 0001.jpg and 0014.jpg distance is 0.6297745
- dist 1.329717
- 0001.jpg and 0031.jpg distance is 1.329717
- dist 1.0085028
- 0001.jpg and 003_2.jpg distance is 1.0085028
- dist 0.9817231
- 000_0.jpg and 0014.jpg distance is 0.9817231
- dist 1.2583163
- 000_0.jpg and 0031.jpg distance is 1.2583163
- dist 0.81444544
- 000_0.jpg and 003_2.jpg distance is 0.81444544
- dist 1.2606885
- 0014.jpg and 0031.jpg distance is 1.2606885
- dist 1.0857416
- 0014.jpg and 003_2.jpg distance is 1.0857416
- dist 1.328266
- 0031.jpg and 003_2.jpg distance is 1.328266
复制代码
- spend_time 0.06249642372131348
- dist 0.0
- 0001.jpg and 000_0.jpg distance is 0.0
- dist 0.0
- 0001.jpg and 0014.jpg distance is 0.0
- dist 0.0
- 0001.jpg and 0031.jpg distance is 0.0
- dist 0.0
- 0001.jpg and 003_2.jpg distance is 0.0
- dist 0.0
- 000_0.jpg and 0014.jpg distance is 0.0
- dist 0.0
- 000_0.jpg and 0031.jpg distance is 0.0
- dist 0.0
- 000_0.jpg and 003_2.jpg distance is 0.0
- dist 0.0
- 0014.jpg and 0031.jpg distance is 0.0
- dist 0.0
- 0014.jpg and 003_2.jpg distance is 0.0
- dist 0.0
- 0031.jpg and 003_2.jpg distance is 0.0
复制代码
jefferyzhang 发表于 2019-11-20 21:16
这个不至于吧,量化图片要给足够多,100起步,上不封顶。不然量化出得精度不够稳定,很容易爆了。
然后版 ...
欢迎光临 Toybrick (https://t.rock-chips.com/) | Powered by Discuz! X3.3 |