Reproduce well known models and datasets

lgarithm / crystalnet

crystalnet -- a mini core AI library (being refactored, see https://github.com/lgarithm/stdnn-ops)

MIT License

16 stars 3 forks source link

Reproduce well known models and datasets #47

Open lgarithm opened 6 years ago

lgarithm commented 6 years ago

Datasets

[X] mnist
[ ] fashion mnist
[ ] cifar10
[ ] cifar100
[ ] tiny imagenet
[ ] imagenet (ILSVRC)
[ ] MPII/COCO (#55)
[ ] pascal VOC (#54)
[ ] openpose

Models

trivial models
- [X] slp
- [X] mlp
- [X] cnn
Alexnet
VGG
- [X] 16
- [ ] 19
Resnet (n) (#53)
Inception
YOLO
- [X] v2
- [ ] v3
[ ] SqueezeNet
[ ] MobileNet

luomai commented 6 years ago

At this stage, I would rather concentrate on finalizing the API and quickly reaches the milestone 0.1. I think AlexNet and VGG16 should be good enough.

We should also start reducing redundant code and make the library simple to read and learn. This is the key motivation as well as the major reason for people to get interested in this project.

luomai commented 6 years ago

A killer example of crystalnet is the white-box implementation of YOLO. People have found many useful implementations of other models in other frameworks already.

lgarithm commented 6 years ago

VGG16 is now exactly reproduced, but 1.5 times slower (single image inference) than TF. Alexnet is not exactly reproduced, due to the original model has a split and merge structure.