File size: 14,634 Bytes
d4ed221
 
92e3dd8
 
 
 
 
 
 
 
 
 
 
d4ed221
92e3dd8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15520aa
 
92e3dd8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5a566f3
92e3dd8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5a566f3
 
 
 
 
 
 
92e3dd8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
---
license: apache-2.0
datasets:
- imagenet-1k
- coco
metrics:
  - mean_average_precision
pipeline_tag: object-detection
tags:
- vision
language:
- en
library_name: pytorch
---

# TransNeXt

Official Model release
for ["TransNeXt: Robust Foveal Visual Perception for Vision Transformers"](https://arxiv.org/pdf/2311.17132.pdf) [CVPR 2024]
.
## Model Details
- **Code:** https://github.com/DaiShiResearch/TransNeXt
- **Paper:** [TransNeXt: Robust Foveal Visual Perception for Vision Transformers](https://arxiv.org/abs/2311.17132)
- **Author:**  [Dai Shi](https://github.com/DaiShiResearch) 
- **Email:** daishiresearch@gmail.com
  
## Methods
#### Pixel-focused attention (Left) & aggregated attention (Right):

![pixel-focused_attention](https://github.com/DaiShiResearch/TransNeXt/blob/main/figures/pixel-focused_attention.jpg?raw=true "pixel-focused_attention")

#### Convolutional GLU (First on the right):

![Convolutional GLU](https://github.com/DaiShiResearch/TransNeXt/blob/main/figures/feedforward_variants.jpg?raw=true "Convolutional GLU")

## Results

#### Image Classification, Detection and Segmentation:

![experiment_figure](https://github.com/DaiShiResearch/TransNeXt/blob/main/figures/experiment_figure.jpg?raw=true "experiment_figure")

#### Attention Visualization:

![foveal_peripheral_vision](https://github.com/DaiShiResearch/TransNeXt/blob/main/figures/foveal_peripheral_vision.jpg?raw=true "foveal_peripheral_vision")

## Model Zoo

### Image Classification

***Classification code & weights & configs & training logs are >>>[here](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/ )<<<.***

**ImageNet-1K 224x224 pre-trained models:**

| Model | #Params | #FLOPs |IN-1K | IN-A | IN-C&#8595; |IN-R|Sketch|IN-V2|Download |Config| Log |
|:---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:|
| TransNeXt-Micro|12.8M|2.7G| 82.5 | 29.9 | 50.8|45.8|33.0|72.6|[model](https://huggingface.co/DaiShiResearch/transnext-micro-224-1k/resolve/main/transnext_micro_224_1k.pth?download=true) |[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_micro.py)|[log](https://huggingface.co/DaiShiResearch/transnext-micro-224-1k/raw/main/transnext_micro_224_1k.txt) |
| TransNeXt-Tiny |28.2M|5.7G| 84.0| 39.9| 46.5|49.6|37.6|73.8|[model](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_tiny.py)|[log](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/raw/main/transnext_tiny_224_1k.txt)|
| TransNeXt-Small |49.7M|10.3G| 84.7| 47.1| 43.9|52.5| 39.7|74.8 |[model](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_small.py)|[log](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/raw/main/transnext_small_224_1k.txt)|
| TransNeXt-Base |89.7M|18.4G| 84.8| 50.6|43.5|53.9|41.4|75.1| [model](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_base.py)|[log](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/raw/main/transnext_base_224_1k.txt)|

**ImageNet-1K 384x384 fine-tuned models:**

| Model | #Params | #FLOPs |IN-1K | IN-A |IN-R|Sketch|IN-V2| Download |Config| 
|:---:|:---:|:---:|:---:| :---:|:---:|:---:| :---:|:---:|:---:|
| TransNeXt-Small |49.7M|32.1G| 86.0| 58.3|56.4|43.2|76.8| [model](https://huggingface.co/DaiShiResearch/transnext-small-384-1k-ft-1k/resolve/main/transnext_small_384_1k_ft_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/finetune/transnext_small_384_ft.py)|
| TransNeXt-Base |89.7M|56.3G| 86.2| 61.6|57.7|44.7|77.0| [model](https://huggingface.co/DaiShiResearch/transnext-base-384-1k-ft-1k/resolve/main/transnext_base_384_1k_ft_1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/finetune/transnext_base_384_ft.py)|

**ImageNet-1K 256x256 pre-trained model fully utilizing aggregated attention at all stages:**

*(See Table.9 in Appendix D.6 for details)*

| Model |Token mixer| #Params | #FLOPs |IN-1K |Download |Config| Log |
|:---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:|
|TransNeXt-Micro|**A-A-A-A**|13.1M|3.3G| 82.6 |[model](https://huggingface.co/DaiShiResearch/transnext-micro-AAAA-256-1k/resolve/main/transnext_micro_AAAA_256_1k.pth?download=true) |[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/classification/configs/transnext_micro_AAAA_256.py)|[log](https://huggingface.co/DaiShiResearch/transnext-micro-AAAA-256-1k/blob/main/transnext_micro_AAAA_256_1k.txt) |

### Object Detection
***Object detection code & weights & configs & training logs are >>>[here](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/ )<<<.***

**COCO object detection and instance segmentation results using the Mask R-CNN method:**

| Backbone | Pretrained Model| Lr Schd| box mAP | mask mAP | #Params | Download |Config| Log |
|:---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:|:---:|
| TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true) |1x|49.9|44.6|47.9M|[model](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-tiny-coco/resolve/main/mask_rcnn_transnext_tiny_fpn_1x_coco_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/maskrcnn/configs/mask_rcnn_transnext_tiny_fpn_1x_coco.py)|[log](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-tiny-coco/raw/main/mask_rcnn_transnext_tiny_fpn_1x_coco_in1k.log.json)|
| TransNeXt-Small | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true) |1x|51.1|45.5|69.3M|[model](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-small-coco/resolve/main/mask_rcnn_transnext_small_fpn_1x_coco_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/maskrcnn/configs/mask_rcnn_transnext_small_fpn_1x_coco.py)|[log](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-small-coco/raw/main/mask_rcnn_transnext_small_fpn_1x_coco_in1k.log.json)|
| TransNeXt-Base | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true) |1x|51.7|45.9|109.2M|[model](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-base-coco/resolve/main/mask_rcnn_transnext_base_fpn_1x_coco_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/maskrcnn/configs/mask_rcnn_transnext_base_fpn_1x_coco.py)|[log](https://huggingface.co/DaiShiResearch/maskrcnn-transnext-base-coco/raw/main/mask_rcnn_transnext_base_fpn_1x_coco_in1k.log.json)|

**COCO object detection results using the DINO method:**

| Backbone | Pretrained Model| scales | epochs | box mAP | #Params | Download |Config| Log |
|:---:|:---:|:---:|:---:| :---:|:---:|:---:|:---:|:---:|
| TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|4scale | 12|55.1|47.8M|[model](https://huggingface.co/DaiShiResearch/dino-4scale-transnext-tiny-coco/resolve/main/dino_4scale_transnext_tiny_12e_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/dino/configs/dino-4scale_transnext_tiny-12e_coco.py)|[log](https://huggingface.co/DaiShiResearch/dino-4scale-transnext-tiny-coco/raw/main/dino_4scale_transnext_tiny_12e_in1k.json)|
| TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|5scale | 12|55.7|48.1M|[model](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-tiny-coco/resolve/main/dino_5scale_transnext_tiny_12e_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/dino/configs/dino-5scale_transnext_tiny-12e_coco.py)|[log](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-tiny-coco/raw/main/dino_5scale_transnext_tiny_12e_in1k.json)|
| TransNeXt-Small | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true)|5scale | 12|56.6|69.6M|[model](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-small-coco/resolve/main/dino_5scale_transnext_small_12e_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/dino/configs/dino-5scale_transnext_small-12e_coco.py)|[log](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-small-coco/raw/main/dino_5scale_transnext_small_12e_in1k.json)|
| TransNeXt-Base | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true)|5scale | 12|57.1|110M|[model](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-base-coco/resolve/main/dino_5scale_transnext_base_12e_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/detection/dino/configs/dino-5scale_transnext_base-12e_coco.py)|[log](https://huggingface.co/DaiShiResearch/dino-5scale-transnext-base-coco/raw/main/dino_5scale_transnext_base_12e_in1k.json)|

### Semantic Segmentation
***Semantic segmentation code & weights & configs & training logs are >>>[here](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/ )<<<.***

**ADE20K semantic segmentation results using the UPerNet method:**

| Backbone | Pretrained Model| Crop Size |Lr Schd| mIoU|mIoU (ms+flip)| #Params | Download |Config| Log |
|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|512x512|160K|51.1|51.5/51.7|59M|[model](https://huggingface.co/DaiShiResearch/upernet-transnext-tiny-ade/resolve/main/upernet_transnext_tiny_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/upernet/configs/upernet_transnext_tiny_512x512_160k_ade20k_ss.py)|[log](https://huggingface.co/DaiShiResearch/upernet-transnext-tiny-ade/blob/main/upernet_transnext_tiny_512x512_160k_ade20k_ss.log.json)|
| TransNeXt-Small | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true)|512x512|160K|52.2|52.5/52.8|80M|[model](https://huggingface.co/DaiShiResearch/upernet-transnext-small-ade/resolve/main/upernet_transnext_small_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/upernet/configs/upernet_transnext_small_512x512_160k_ade20k_ss.py)|[log](https://huggingface.co/DaiShiResearch/upernet-transnext-small-ade/blob/main/upernet_transnext_small_512x512_160k_ade20k_ss.log.json)|
| TransNeXt-Base | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true)|512x512|160K|53.0|53.5/53.7|121M|[model](https://huggingface.co/DaiShiResearch/upernet-transnext-base-ade/resolve/main/upernet_transnext_base_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/upernet/configs/upernet_transnext_base_512x512_160k_ade20k_ss.py)|[log](https://huggingface.co/DaiShiResearch/upernet-transnext-base-ade/blob/main/upernet_transnext_base_512x512_160k_ade20k_ss.log.json)|
* In the context of multi-scale evaluation, TransNeXt reports test results under two distinct scenarios: **interpolation** and **extrapolation** of relative position bias. 

**ADE20K semantic segmentation results using the Mask2Former method:**

| Backbone | Pretrained Model| Crop Size |Lr Schd| mIoU| #Params | Download |Config| Log |
|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| TransNeXt-Tiny | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-tiny-224-1k/resolve/main/transnext_tiny_224_1k.pth?download=true)|512x512|160K|53.4|47.5M|[model](https://huggingface.co/DaiShiResearch/mask2former-transnext-tiny-ade/resolve/main/mask2former_transnext_tiny_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/mask2former/configs/mask2former_transnext_tiny_160k_ade20k-512x512.py)|[log](https://huggingface.co/DaiShiResearch/mask2former-transnext-tiny-ade/raw/main/mask2former_transnext_tiny_512x512_160k_ade20k_in1k.json)|
| TransNeXt-Small | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-small-224-1k/resolve/main/transnext_small_224_1k.pth?download=true)|512x512|160K|54.1|69.0M|[model](https://huggingface.co/DaiShiResearch/mask2former-transnext-small-ade/resolve/main/mask2former_transnext_small_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/mask2former/configs/mask2former_transnext_small_160k_ade20k-512x512.py)|[log](https://huggingface.co/DaiShiResearch/mask2former-transnext-small-ade/raw/main/mask2former_transnext_small_512x512_160k_ade20k_in1k.json)|
| TransNeXt-Base | [ImageNet-1K](https://huggingface.co/DaiShiResearch/transnext-base-224-1k/resolve/main/transnext_base_224_1k.pth?download=true)|512x512|160K|54.7|109M|[model](https://huggingface.co/DaiShiResearch/mask2former-transnext-base-ade/resolve/main/mask2former_transnext_base_512x512_160k_ade20k_in1k.pth?download=true)|[config](https://github.com/DaiShiResearch/TransNeXt/tree/main/segmentation/mask2former/configs/mask2former_transnext_base_160k_ade20k-512x512.py)|[log](https://huggingface.co/DaiShiResearch/mask2former-transnext-base-ade/raw/main/mask2former_transnext_base_512x512_160k_ade20k_in1k.json)|

## Citation

If you find our work helpful, please consider citing the following bibtex. We would greatly appreciate a star for this
project.

    @InProceedings{shi2023transnext,
      author    = {Dai Shi},
      title     = {TransNeXt: Robust Foveal Visual Perception for Vision Transformers},
      booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
      month     = {June},
      year      = {2024},
      pages     = {17773-17783}
    }