Use CNTK Pre-trained Image Models in FastAI

mehrangh66 · June 24, 2020, 10:45pm

Is there anyway to use the CNTK Pre-trained Image Models in fast ai?

Microsoft/CNTK/blob/master/PretrainedModels/Image.md#resnet

# CNTK Pre-trained Image Models

This page contains pre-trained image models either converted from other toolkits or trained from scratch with CNTK. The list of available models includes:

* [AlexNet](#alexnet)
* [GoogLeNet](#googlenet)
* [ResNet](#resnet)
* [VGG](#vgg)

## AlexNet

### CNTK Pre-trained
#### AlexNet for ImageNet 1K
|CNTK model download path | https://www.cntk.ai/Models/CNTK_Pretrained/AlexNet_ImageNet_CNTK.model
|:---------|:---
|Training script | [AlexNet_ImageNet.cntk](../Examples/Image/Classification/AlexNet/BrainScript/AlexNet_ImageNet.cntk)
|Single crop top 1 / top 5 error | 40.106% / 17.746%

### Caffe-Converted
#### AlexNet for ImageNet 1K

This file has been truncated. show original

bwarner · June 25, 2020, 4:28am

Fastai is a PyTorch library, so you can use any compatible PyTorch model in Fastai.

You can directly use them using Learner or define a cut and split for cnn_learner. You will also need to set the pillow loading mode as CNTK, as Fastai defaults is to load as RGB.