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.tarThis 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.tarPrint 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 nameCreate 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 diagsSome samples of the same are as follows
