CMT: Convolutional Neural Networks Meet Vision Transformers
A PyTorch implementation of CMT based on paper CMT: Convolutional Neural Networks Meet Vision Transformers.
Network Overview
CMT variants structure overview
Number of Parameters
| Model | # Params of the paper | # Parames of this implement | MACs(G) |
|---|---|---|---|
| CMT-Ti | 9.49 M | 10.32 M | 1.21 |
| CMT-XS | 15.24 M | 16.40 M | 2.04 |
| CMT-S | 25.14 M | 27.38 M | 3.88 |
| CMT-B | 45.72 M | 47.06 M | 6.83 |
Usage
Training
python main.py
optional arguments:
-h, --help show this help message and exit
--gpu_device GPU_DEVICE
Select specific GPU to run the model
--batch-size N Input batch size for training (default: 64)
--epochs N Number of epochs to train (default: 20)
--num-class N Number of classes to classify (default: 10)
--lr LR Learning rate (default: 0.01)
--weight-decay WD Weight decay (default: 1e-5)
--model-path PATH Path to save the model
Experiments Results (Cifar-10)
| Model | Dataset | Learning Rate | LR Scheduler | Optimizer | Weight decay | Acc@1 | Acc@5 |
|---|---|---|---|---|---|---|---|
| CMT-Ti | Cifar10 | 6e-5 | Cosine LR | AdamW | 1e-5 | 88.16% | 99.49% |
TODOs
- Train on the cifar-10 dataset (Due to the computation limit).
Citation
@misc{guo2021cmt,
title={CMT: Convolutional Neural Networks Meet Vision Transformers},
author={Jianyuan Guo and Kai Han and Han Wu and Chang Xu and Yehui Tang and Chunjing Xu and Yunhe Wang},
year={2021},
eprint={2107.06263},
archivePrefix={arXiv},
primaryClass={cs.CV}}
Author
Hong-Jia Chen