AI-Model-Zoo icon indicating copy to clipboard operation
AI-Model-Zoo copied to clipboard

Published 20 hours ago • verified publisher icon Xilinx

Quantization of tf_mobilenetv1_1.0_imagenet_224_224_1.14G_1.1 with the DNNDK v3.1 package

Open Arroas opened this issue 4 years ago • 0 comments

Hello: I'm trying to port the tf_mobilenetv1_1.0 model to a ZedBoard. I'm using the DNNDK v3.1 package (Zynq-7000 family chips are not supported by Vitis-AI), and I'm having trouble with the quantization step. After quantization, there is a high decrease in accuracy of the model:

Accuracy Frozen Graph Quantized Graph
Top1 0.642000013589859 0.01399999968707561
Top5 0.851999986171723 0.06199999935925007

I believe the problem is that I'm using the wrong image pre-processing when calibrating (resize to 224x224, bgr2rgb and normalize between 0-1).

What image pre-processing has been used when training the tf_mobilenetv1_1.0 model?

Arroas avatar Jun 02 '20 11:06 Arroas