siddsax / visualizemsdnet

Visualizing the Multi-scale dense network with Grad-CAM

Grad-CAM Visualizations for MSD-Net

This repository contains the code for visualizing the model proposed in the paper Multi-Scale Dense Networks for Resource Efficient Image Classification link. The base code here for that paper is borrowed from the wonderfully replicated work of @avirambh from here. Also for Grad-CAM, the basic code is taken from here. Here is a more generalized version that can be used for any custom network.

Also checkout my two part blog on this. Part-1, Part-2

Download the trained MSD-Net models for Cifar-10 from here

First we can test the model for accuracies the pre-trained models as

python main.py --model msdnet -b 20  -j 2 cifar10 --growth 6-12-24 --gpu 0 --resume --evaluate-from anytime_cifar_10.pth.tar

This nearly replicates the original paper which reports about 94.5% when budget is averaged.

Classifier No.	Accuracy
0:	90.800
1:	92.330
2:	93.180
3:	93.760
4:	94.140
5:	94.130
6:	94.270
7:	94.220
8:	94.210
9:	94.170
Mean	93.521

This table also shows that the accuracy generally increases as number of blocks increase.

Print whether the classifiers are predicting each point correctly or not

python heatmapping.py --model msdnet -b 20  -j 2 cifar10  --growth 6-12-24 --gpu 0 --resume --evaluate-from anytime_cifar_10.pth.tar

Print whether the classifiers after each blocks are predicting correctly or not for a particular class

python heatmapping.py --model msdnet -b 20  -j 2 cifar10  --growth 6-12-24 --gpu 0 --resume --evaluate-from anytime_cifar_10.pth.tar --cl=dog # here dog is the class name

Create the heatmap of a particular classifier and image

python heatmapping.py --model msdnet -b 20  -j 2 cifar10  --growth 6-12-24 --gpu 0 --resume --evaluate-from anytime_cifar_10.pth.tar --imgNo=100 --maxC=5 # this means img 100 predicted with classifier 5 will be generated in folder diags

Some samples of the same are as follows

visualizemsdnet's Issues

链接问题

你好，这个博客和预训练权重两个地址都无法打开

报错

语句： model = getattr(models, args.model)(args)
报以下错误：
AttributeError: module 'torchvision.models' has no attribute 'msdnet'
请问作者碰到了吗？

Two links you provided in README.md are not available now. Can you provide these resources again?

I find 2 links in README.md page. But they two are not useful. Details are following:

• Part-1 Not Availabel
• Part-2 Not Availabel

Hope you can provide thses files again. Thank you very much.