ModelFeast
ModelFeast copied to clipboard
Published
20 hours ago •
daili0015
daili0015
Pytorch model zoo for human, include all kinds of 2D CNN, 3D CNN, and CRNN
ModelFeast
Please star ModelFeast if it helps you. This is very important to me! Thanks very much !
ModelFeast is more than model-zoo! It is:
- A gather of the most popular 2D, 3D CNN models
- A tool to make deep learn much more simply and flexibly
- A pytorch project template
Avalible models
2D CNN
- Xception
- InceptionV3
- InceptionResnetV2
- SqueezeNet1_0, SqueezeNet1_1
- VGG11, VGG13, VGG16, VGG19
- ResNet18, ResNet34, ResNet50, ResNet101, ResNet152
- ResNext101_32x4d, ResNext101_64x4d
- DenseNet121, DenseNet169, DenseNet201, DenseNet161
Pretrained models ( trained on ImageNet ) for 2D CNN is now avalible on Baiduyun(fst6) and Google Drive
3D CNN
- resnet18v2_3d, resnet34v2_3d, resnet50v2_3d, resnet101v2_3d, resnet152v2_3d, resnet200v2_3d
- resnext50_3d, resnext101_3d, resnext152_3d
- densenet121_3d, densenet169_3d, densenet201_3d, densenet264_3d
- resnet10_3d, resnet18_3d, resnet34_3d, resnet101_3d, resnet152_3d, resnet200_3d
- wideresnet50_3d
- i3d50, i3d101, i3d152
CNN-RNN
This part is still on progress. Not avalible to train now, but model architecture can been seen here.
Get started
Determine what you need and read corresponding tutorials
- I want to train a model as simple as possible
- I just need the codes of CNNs
- I need a standard pytorch project template
Or you can use modelfeast simply via pip !
pip3 install modelfeast
Features
The features are more than you could think of:
- Train and save model within 3 lines !
- Multi GPU support !
- Include the most popular 2D CNN, 3D CNN, and CRNN models !
- Allow any input image size (pytorch official model zoo limit your input size harshly) !
- Help you sweep all kinds of classification competitions.
Reference
https://github.com/lanpa/tensorboardX
https://github.com/pytorch/vision/tree/master/torchvision/models
https://github.com/kenshohara/3D-ResNets-PyTorch
https://github.com/victoresque/pytorch-template
daili0015