EfficientNet
A PyTorch implementation of EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks.
[arxiv] [Official TF Repo]
Implement based on Official TF Repo. Only opened EfficientNet is included.
This repo not contains baseline network search(Mnas-Net) and compound coefficient search methods.
Some details(HyperParams, transform, EMA ...) are different with Original repo.
Pretrained network
This is not end-to-end imagenet trainning weight. Using the Official TF Pretrained weight.
Please check the conversion/readme.md and pretrained_example.ipynb
How to use:
python3 main.py -h
usage: main.py [-h] --save_dir SAVE_DIR [--root ROOT] [--gpus GPUS]
[--num_workers NUM_WORKERS] [--model {b0}] [--epoch EPOCH]
[--batch_size BATCH_SIZE] [--test]
[--dropout_rate DROPOUT_RATE]
[--dropconnect_rate DROPCONNECT_RATE] [--optim {adam,rmsprop}]
[--lr LR] [--beta [BETA [BETA ...]]] [--momentum MOMENTUM]
[--eps EPS] [--decay DECAY]
Pytorch EfficientNet
optional arguments:
-h, --help show this help message and exit
--save_dir SAVE_DIR Directory name to save the model
--root ROOT The Directory of data path.
--gpus GPUS Select GPU Numbers | 0,1,2,3 |
--num_workers NUM_WORKERS
Select CPU Number workers
--model {b0} The type of Efficient net.
--epoch EPOCH The number of epochs
--batch_size BATCH_SIZE
The size of batch
--test Only Test
--dropout_rate DROPOUT_RATE
--dropconnect_rate DROPCONNECT_RATE
--optim {adam,rmsprop}
--lr LR Base learning rate when train batch size is 256.
--beta [BETA [BETA ...]]
--momentum MOMENTUM
--eps EPS
--decay DECAY
TODO
- Hyper Parameter / Imagenet Transformation Check
- Implementation of Resolution Change
- Validation on Imagenet Dataset
- Clean up logging