search

onnx / models

A collection of pre-trained, state-of-the-art models in the ONNX format
http://onnx.ai/models/
Apache License 2.0
7.96k stars 1.41k forks source link

Pre-processing steps for most models missing #172

Open dibya001 opened 5 years ago

dibya001 commented 5 years ago

There are no pre-processing steps mentioned for models like alexnet, googlenet, incpetion v1 and inception v2, etc. I could find the preprocessing steps for Squeezenet, vgg, mobilenet, resnet models.

Can someone let me know what are the preprocessing steps followed for these models? Can the preprocessing steps applied to the caffe model be applied for corresponding ONNX models?

snnn commented 5 years ago

The model you mentioned were converted from caffe2. @bddppq

dibya001 commented 5 years ago

Ok so I guess I can apply the same preprocessing steps for those as caffe2.

For models like squeezenet whose preprocessing steps are mentioned,, when I am applying those preprocessing steps and running them with caffe2 backend(instead of mxnet backend as shown) I am getting pretty low accuracy. Any possible reason for it?