# FcaNet

**Repository Path**: songjh623/FcaNet

## Basic Information

- **Project Name**: FcaNet
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2024-04-13
- **Last Updated**: 2024-04-13

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# FcaNet: Frequency Channel Attention Networks
PyTorch implementation of the paper "[FcaNet: Frequency Channel Attention Networks](https://arxiv.org/abs/2012.11879)".

![alt text](images/main.png "vis")

# Simplest usage

Models pretrained on ImageNet can be simply accessed by (without any configuration or installation):
```
model = torch.hub.load('cfzd/FcaNet', 'fca34' ,pretrained=True)
model = torch.hub.load('cfzd/FcaNet', 'fca50' ,pretrained=True)
model = torch.hub.load('cfzd/FcaNet', 'fca101' ,pretrained=True)
model = torch.hub.load('cfzd/FcaNet', 'fca152' ,pretrained=True)
```

# Install
Please see [INSTALL.md](./INSTALL.md)

# Models
## Classification models on ImageNet

Due to the conversion between FP16 training and the provided FP32 models, the evaluation results are slightly different(max -0.06%/+0.05%) compared with the reported results.
|  Model | Reported | Evaluation Results | Link |   
|:--------:|:------------:|:----------------:|:-------------------:|
| FcaNet34 |     75.07    |       75.02      |          [GoogleDrive](https://drive.google.com/file/d/1zTbhu1ynShAG9qH_HKJfOQeWBWVD9dCW/view?usp=sharing)/[BaiduDrive(code:m7v8)](https://pan.baidu.com/s/1ylq_tJaRlCSHhDcmVinnfA) |
|  FcaNet50  |     78.52     |       78.57       |          [GoogleDrive](https://drive.google.com/file/d/1O7qtqHSXdwRzUv_m6wrqB4TVF-rqJInB/view?usp=sharing)/[BaiduDrive(code:mgkk)](https://pan.baidu.com/s/1yrZYotYYGCFU7as_6h8GZw) |
|  FcaNet101  |    79.64    |       79.63       |          [GoogleDrive](https://drive.google.com/file/d/11CydGJ_dzjdOn9Z3E00enc125SpFJD8k/view?usp=sharing)/[BaiduDrive(code:8t0j)](https://pan.baidu.com/s/1SPE7_uxbKLz9s1k97fRl5Q) |
|  FcaNet152  |     80.08     |       80.02       |          [GoogleDrive](https://drive.google.com/file/d/1rtNAmePV4dIpTEkLCT08p6yiRyH0FC3B/view?usp=sharing)/[BaiduDrive(code:5yeq)](https://pan.baidu.com/s/1rLtkHIDqks4jSIq0Nqt-fQ) |


## Detection and instance segmentation models on COCO

|  Model | Backbone | AP | AP50 | AP75 |Link |   
|:--------:|:------------:|:----------------:|:-------------------:|:--:|:--:|
| Faster RCNN | FcaNet50 | 39.0 | 61.1 | 42.3| [GoogleDrive](https://drive.google.com/file/d/1_LG_hqoD7v-vmWkw9_dq5Yis8IAF8cBe/view?usp=sharing)/[BaiduDrive(code:q15c)](https://pan.baidu.com/s/1nIEWG7uFFxDy82XHnTW5JA) |
| Faster RCNN | FcaNet101 | 41.2 | 63.3 | 44.6 | [GoogleDrive](https://drive.google.com/file/d/1ObDEKQZbRrs_OsQBpBYa9WusKT9XkQso/view?usp=sharing)/[BaiduDrive(code:pgnx)](https://pan.baidu.com/s/15BRDqH3VzCDfZVPmSC0D7w) |
| Mask RCNN | Fca50 det </br>  Fca50 seg | 40.3</br> 36.2 | 62.0 </br> 58.6 | 44.1 </br> 38.1 | [GoogleDrive](https://drive.google.com/file/d/18Kh3upRLauUXLmmOPSH9kCZe6aXoIlRm/view?usp=sharing)/[BaiduDrive(code:d9rn)](https://pan.baidu.com/s/1m40e9Bcxcf04lxq3LlFTbw) |

# Training
Please see `launch_training_classification.sh` and `launch_training_detection.sh` for training on ImageNet and COCO, respectively.

# Testing
Please see `launch_eval_classification.sh` and `launch_eval_detection.sh` for testing on ImageNet and COCO, respectively.

# FAQ
Since the paper is uploaded to arxiv, many academic peers ask us: the proposed DCT basis can be viewed as a simple tensor, then how about learning the tensor directly? Why use DCT instead of learnable tensor? Learnable tensor can be better than DCT.

Our concrete answer is: **the proposed DCT is better than the learnable way, although it is counter-intuitive.**

| Method                                      |      ImageNet Top-1 Acc   | Link |
|---------------------------------------------|:-------------:|:----------:|
| Learnable tensor, random initialization | 77.914 |        [GoogleDrive](https://drive.google.com/file/d/1KsiSuyHNtIxBEabdaDT_0Lt5foSNBgtb/view?usp=sharing)/[BaiduDrive(code:p2hl)](https://pan.baidu.com/s/1IBavAQ-GnvElqJBY6oblWg)     |
| Learnable tensor, DCT initialization    |     78.352     |      [GoogleDrive](https://drive.google.com/file/d/1CwpW0lRwU7uXEf8ldm4hJUs-9O33MKfI/view?usp=sharing)/[BaiduDrive(code:txje)](https://pan.baidu.com/s/1CC2j9tjySH-GGZ9oM7cpjA)       |
| Fixed tensor, random initialization     | 77.742 |        [GoogleDrive](https://drive.google.com/file/d/1jRXX4QqYsU8wyOHWW34qVF0DuLZ3R884/view?usp=sharing)/[BaiduDrive(code:g5t9)](https://pan.baidu.com/s/1HpJyi9h6L0_rvEg218JK9g)     |
| Fixed tensor, DCT initialization (Ours) |     78.574     |     [GoogleDrive](https://drive.google.com/file/d/1O7qtqHSXdwRzUv_m6wrqB4TVF-rqJInB/view?usp=sharing)/[BaiduDrive(code:mgkk)](https://pan.baidu.com/s/1yrZYotYYGCFU7as_6h8GZw)        |


To verify this results, one can select the cooresponding types of tensor in the L73-L83 in `model/layer.py`, uncomment it and train the whole network.


# TODO
- [x] Object detection models
- [x] Instance segmentation models
- [x] Fix the incorrect results of detection models
- [ ] Make the switching between configs more easier