Hugging Face's logo

suous
/
recnext_a3.dist_300e_in1k

Image Classification
timm
PyTorch
Safetensors
Transformers
English
vision
Model card Files
xet
Community
1
suous commited on
Commit
32b2b45
·
verified ·
1 Parent(s): 09466a7

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +490 -36
README.md CHANGED
@@ -18,6 +18,9 @@ tags:
18
 
19
  # Model Card for RecNeXt-A3 (With Knowledge Distillation)
20
 
 
 
 
21
  [![license](https://img.shields.io/github/license/suous/RecNeXt)](https://github.com/suous/RecNeXt/blob/main/LICENSE)
22
  [![arXiv](https://img.shields.io/badge/arXiv-2412.19628-red)](https://arxiv.org/abs/2412.19628)
23
 
@@ -26,10 +29,6 @@ tags:
26
  <img src="https://raw.githubusercontent.com/suous/RecNeXt/refs/heads/main/figures/code.png" alt="code" style="width: 46%;">
27
  </div>
28
 
29
- ## Abstract
30
-
31
- Recent advances in vision transformers (ViTs) have demonstrated the advantage of global modeling capabilities, prompting widespread integration of large-kernel convolutions for enlarging the effective receptive field (ERF). However, the quadratic scaling of parameter count and computational complexity (FLOPs) with respect to kernel size poses significant efficiency and optimization challenges. This paper introduces RecConv, a recursive decomposition strategy that efficiently constructs multi-frequency representations using small-kernel convolutions. RecConv establishes a linear relationship between parameter growth and decomposing levels which determines the effective receptive field $k\times 2^\ell$ for a base kernel $k$ and $\ell$ levels of decomposition, while maintaining constant FLOPs regardless of the ERF expansion. Specifically, RecConv achieves a parameter expansion of only $\ell+2$ times and a maximum FLOPs increase of $5/3$ times, compared to the exponential growth ($4^\ell$) of standard and depthwise convolutions. RecNeXt-M3 outperforms RepViT-M1.1 by 1.9 $AP^{box}$ on COCO with similar FLOPs. This innovation provides a promising avenue towards designing efficient and compact networks across various modalities. Codes and models can be found at this https URL .
32
-
33
  ## Model Details
34
 
35
  - **Model Type**: Image Classification / Feature Extraction
@@ -38,6 +37,7 @@ Recent advances in vision transformers (ViTs) have demonstrated the advantage of
38
  - **Parameters**: 9.0M
39
  - **MACs**: 1.4G
40
  - **Latency**: 2.4ms (iPhone 13, iOS 18)
 
41
  - **Image Size**: 224x224
42
 
43
  - **Architecture Configuration**:
@@ -51,6 +51,17 @@ Recent advances in vision transformers (ViTs) have demonstrated the advantage of
51
 
52
  - **Dataset**: ImageNet-1K
53
 
 
 
 
 
 
 
 
 
 
 
 
54
  ## Model Usage
55
 
56
  ### Image Classification
@@ -93,46 +104,476 @@ We introduce two series of models: the **A** series uses linear attention and ne
93
 
94
  > **dist**: distillation; **base**: without distillation (all models are trained over 300 epochs).
95
 
96
- | model | top_1_accuracy | params | gmacs | npu_latency | cpu_latency | throughput | fused_weights | training_logs |
97
- |-------|----------------|--------|-------|-------------|-------------|------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
98
- | M0 | 74.7* \| 73.2 | 2.5M | 0.4 | 1.0ms | 189ms | 763 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m0_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m0_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_m0_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_m0_without_distill_300e.txt) |
99
- | M1 | 79.2* \| 78.0 | 5.2M | 0.9 | 1.4ms | 361ms | 384 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m1_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m1_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_m1_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_m1_without_distill_300e.txt) |
100
- | M2 | 80.3* \| 79.2 | 6.8M | 1.2 | 1.5ms | 431ms | 325 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m2_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m2_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_m2_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_m2_without_distill_300e.txt) |
101
- | M3 | 80.9* \| 79.6 | 8.2M | 1.4 | 1.6ms | 482ms | 314 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m3_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m3_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_m3_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_m3_without_distill_300e.txt) |
102
- | M4 | 82.5* \| 81.1 | 14.1M | 2.4 | 2.4ms | 843ms | 169 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m4_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m4_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_m4_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_m4_without_distill_300e.txt) |
103
- | M5 | 83.3* \| 81.6 | 22.9M | 4.7 | 3.4ms | 1487ms | 104 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m5_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m5_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_m5_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_m5_without_distill_300e.txt) |
104
- | A0 | 75.0* \| 73.6 | 2.8M | 0.4 | 1.4ms | 177ms | 4902 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a0_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a0_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_a0_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_a0_without_distill_300e.txt) |
105
- | A1 | 79.6* \| 78.3 | 5.9M | 0.9 | 1.9ms | 334ms | 2746 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a1_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a1_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_a1_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_a1_without_distill_300e.txt) |
106
- | A2 | 80.8* \| 79.6 | 7.9M | 1.2 | 2.2ms | 413ms | 2327 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a2_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a2_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_a2_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_a2_without_distill_300e.txt) |
107
- | A3 | 81.1* \| 80.1 | 9.0M | 1.4 | 2.4ms | 447ms | 2206 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a3_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a3_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_a3_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_a3_without_distill_300e.txt) |
108
- | A4 | 82.5* \| 81.6 | 15.8M | 2.4 | 3.6ms | 764ms | 1265 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a4_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a4_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_a4_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_a4_without_distill_300e.txt) |
109
- | A5 | 83.5* \| 83.1 | 25.7M | 4.7 | 5.6ms | 1376ms | 721 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a5_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a5_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/logs/distill/recnext_a5_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/logs/normal/recnext_a5_without_distill_300e.txt) |
110
 
111
  ### Comparison with [LSNet](https://github.com/jameslahm/lsnet)
112
 
113
- | model | top_1_accuracy | params | gmacs | npu_latency | cpu_latency | throughput | fused_weights | training_logs |
114
- |-------|----------------|--------|-------|-------------|-------------|------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
115
- | T | 76.6* \| 75.1 | 12.1M | 0.3 | 1.8ms | 109ms | 14181 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_t_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_t_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/distill/recnext_t_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/normal/recnext_t_without_distill_300e.txt) |
116
- | S | 79.6* \| 78.3 | 15.8M | 0.7 | 2.0ms | 188ms | 8234 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_s_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_s_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/distill/recnext_s_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/normal/recnext_s_without_distill_300e.txt) |
117
- | B | 81.4* \| 80.3 | 19.3M | 1.1 | 2.5ms | 290ms | 4385 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_b_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_b_without_distill_300e_fused.pt) | [dist](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/distill/recnext_b_distill_300e.txt) \| [base](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/normal/recnext_b_without_distill_300e.txt) |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
118
 
119
  > The NPU latency is measured on an iPhone 13 with models compiled by Core ML Tools.
120
  > The CPU latency is accessed on a Quad-core ARM Cortex-A57 processor in ONNX format.
121
  > And the throughput is tested on an Nvidia RTX3090 with maximum power-of-two batch size that fits in memory.
122
 
123
 
124
- ## Citation
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
125
 
126
- ```BibTeX
127
- @misc{zhao2024recnext,
128
- title={RecConv: Efficient Recursive Convolutions for Multi-Frequency Representations},
129
- author={Mingshu Zhao and Yi Luo and Yong Ouyang},
130
- year={2024},
131
- eprint={2412.19628},
132
- archivePrefix={arXiv},
133
- primaryClass={cs.CV}
134
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
135
  ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
136
 
137
  ## Limitations
138
 
@@ -142,8 +583,21 @@ We introduce two series of models: the **A** series uses linear attention and ne
142
 
143
  ## Acknowledgement
144
 
145
- Classification (ImageNet) code base is partly built with [LeViT](https://github.com/facebookresearch/LeViT), [PoolFormer](https://github.com/sail-sg/poolformer), [EfficientFormer](https://github.com/snap-research/EfficientFormer), [RepViT](https://github.com/THU-MIG/RepViT), and [MogaNet](https://github.com/Westlake-AI/MogaNet).
 
 
 
 
146
 
147
- The detection and segmentation pipeline is from [MMCV](https://github.com/open-mmlab/mmcv) ([MMDetection](https://github.com/open-mmlab/mmdetection) and [MMSegmentation](https://github.com/open-mmlab/mmsegmentation)).
148
 
149
- Thanks for the great implementations!
 
 
 
 
 
 
 
 
 
 
18
 
19
  # Model Card for RecNeXt-A3 (With Knowledge Distillation)
20
 
21
+ ## Abstract
22
+ Recent advances in vision transformers (ViTs) have demonstrated the advantage of global modeling capabilities, prompting widespread integration of large-kernel convolutions for enlarging the effective receptive field (ERF). However, the quadratic scaling of parameter count and computational complexity (FLOPs) with respect to kernel size poses significant efficiency and optimization challenges. This paper introduces RecConv, a recursive decomposition strategy that efficiently constructs multi-frequency representations using small-kernel convolutions. RecConv establishes a linear relationship between parameter growth and decomposing levels which determines the effective receptive field $k\times 2^\ell$ for a base kernel $k$ and $\ell$ levels of decomposition, while maintaining constant FLOPs regardless of the ERF expansion. Specifically, RecConv achieves a parameter expansion of only $\ell+2$ times and a maximum FLOPs increase of $5/3$ times, compared to the exponential growth ($4^\ell$) of standard and depthwise convolutions. RecNeXt-M3 outperforms RepViT-M1.1 by 1.9 $AP^{box}$ on COCO with similar FLOPs. This innovation provides a promising avenue towards designing efficient and compact networks across various modalities. Codes and models can be found at https://github.com/suous/RecNeXt.
23
+
24
  [![license](https://img.shields.io/github/license/suous/RecNeXt)](https://github.com/suous/RecNeXt/blob/main/LICENSE)
25
  [![arXiv](https://img.shields.io/badge/arXiv-2412.19628-red)](https://arxiv.org/abs/2412.19628)
26
 
 
29
  <img src="https://raw.githubusercontent.com/suous/RecNeXt/refs/heads/main/figures/code.png" alt="code" style="width: 46%;">
30
  </div>
31
 
 
 
 
 
32
  ## Model Details
33
 
34
  - **Model Type**: Image Classification / Feature Extraction
 
37
  - **Parameters**: 9.0M
38
  - **MACs**: 1.4G
39
  - **Latency**: 2.4ms (iPhone 13, iOS 18)
40
+ - **Throughput**: 2151 (RTX 3090)
41
  - **Image Size**: 224x224
42
 
43
  - **Architecture Configuration**:
 
51
 
52
  - **Dataset**: ImageNet-1K
53
 
54
+ ## Recent Updates
55
+
56
+ **UPDATES** 🔥
57
+ - **2025/07/23**: Added a simple architecture, the overall design follows [LSNet](https://github.com/jameslahm/lsnet).
58
+ - **2025/07/04**: Uploaded classification models to [HuggingFace](https://huggingface.co/suous)🤗.
59
+ - **2025/07/01**: Added more comparisons with [LSNet](https://github.com/jameslahm/lsnet).
60
+ - **2025/06/27**: Added **A** series code and logs, replacing convolution with linear attention.
61
+ - **2025/03/19**: Added more ablation study results, including using attention with RecConv design.
62
+ - **2025/01/02**: Uploaded checkpoints and training logs of RecNeXt-M0.
63
+ - **2024/12/29**: Uploaded checkpoints and training logs of RecNeXt-M1 - M5.
64
+
65
  ## Model Usage
66
 
67
  ### Image Classification
 
104
 
105
  > **dist**: distillation; **base**: without distillation (all models are trained over 300 epochs).
106
 
107
+ | model | top_1_accuracy | params | gmacs | npu_latency | cpu_latency | throughput | fused_weights | training_logs |
108
+ |-------|----------------|--------|-------|-------------|-------------|------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
109
+ | M0 | 74.7* \| 73.2 | 2.5M | 0.4 | 1.0ms | 189ms | 750 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m0_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m0_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_m0_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_m0_without_distill_300e.txt) |
110
+ | M1 | 79.2* \| 78.0 | 5.2M | 0.9 | 1.4ms | 361ms | 384 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m1_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m1_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_m1_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_m1_without_distill_300e.txt) |
111
+ | M2 | 80.3* \| 79.2 | 6.8M | 1.2 | 1.5ms | 431ms | 325 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m2_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m2_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_m2_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_m2_without_distill_300e.txt) |
112
+ | M3 | 80.9* \| 79.6 | 8.2M | 1.4 | 1.6ms | 482ms | 314 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m3_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m3_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_m3_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_m3_without_distill_300e.txt) |
113
+ | M4 | 82.5* \| 81.4 | 14.1M | 2.4 | 2.4ms | 843ms | 169 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m4_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m4_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_m4_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_m4_without_distill_300e.txt) |
114
+ | M5 | 83.3* \| 82.9 | 22.9M | 4.7 | 3.4ms | 1487ms | 104 | [dist](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m5_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m5_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_m5_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_m5_without_distill_300e.txt) |
115
+ | A0 | 75.0* \| 73.6 | 2.8M | 0.4 | 1.4ms | 177ms | 4891 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a0_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a0_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_a0_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_a0_without_distill_300e.txt) |
116
+ | A1 | 79.6* \| 78.3 | 5.9M | 0.9 | 1.9ms | 334ms | 2730 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a1_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a1_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_a1_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_a1_without_distill_300e.txt) |
117
+ | A2 | 80.8* \| 79.6 | 7.9M | 1.2 | 2.2ms | 413ms | 2331 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a2_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a2_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_a2_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_a2_without_distill_300e.txt) |
118
+ | A3 | 81.1* \| 80.1 | 9.0M | 1.4 | 2.4ms | 447ms | 2151 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a3_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a3_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_a3_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_a3_without_distill_300e.txt) |
119
+ | A4 | 82.5* \| 81.6 | 15.8M | 2.4 | 3.6ms | 764ms | 1265 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a4_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a4_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_a4_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_a4_without_distill_300e.txt) |
120
+ | A5 | 83.5* \| 83.1 | 25.7M | 4.7 | 5.6ms | 1376ms | 733 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a5_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a5_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/distill/recnext_a5_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/logs/normal/recnext_a5_without_distill_300e.txt) |
121
 
122
  ### Comparison with [LSNet](https://github.com/jameslahm/lsnet)
123
 
124
+ We present a simple architecture, the overall design follows [LSNet](https://github.com/jameslahm/lsnet). This framework centers around sharing channel features from the previous layers.
125
+ Our motivation for doing so is to reduce the computational cost of token mixers and minimize feature redundancy in the final stage.
126
+
127
+ ![Architecture](https://raw.githubusercontent.com/suous/RecNeXt/refs/heads/main/lsnet/figures/architecture.png)
128
+
129
+ #### With **Shared-Channel Blocks**
130
+
131
+ | model | top_1_accuracy | params | gmacs | npu_latency | cpu_latency | throughput | fused_weights | training_logs |
132
+ |-------|----------------|--------|-------|-------------|-------------|------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
133
+ | T | 76.8 \| 75.2 | 12.1M | 0.3 | 1.8ms | 105ms | 13957 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_t_share_channel_distill_300e_fused.pt) \| [norm](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_t_share_channel_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/distill/recnext_t_share_channel_distill_300e.txt) \| [norm](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/normal/recnext_t_share_channel_without_distill_300e.txt) |
134
+ | S | 79.5 \| 78.3 | 15.8M | 0.7 | 2.0ms | 182ms | 8034 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_s_share_channel_distill_300e_fused.pt) \| [norm](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_s_share_channel_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/distill/recnext_s_share_channel_distill_300e.txt) \| [norm](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/normal/recnext_s_share_channel_without_distill_300e.txt) |
135
+ | B | 81.5 \| 80.3 | 19.2M | 1.1 | 2.5ms | 296ms | 4472 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_b_share_channel_distill_300e_fused.pt) \| [norm](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_b_share_channel_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/distill/recnext_b_share_channel_distill_300e.txt) \| [norm](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/normal/recnext_b_share_channel_without_distill_300e.txt) |
136
+
137
+ #### Without **Shared-Channel Blocks**
138
+
139
+ | model | top_1_accuracy | params | gmacs | npu_latency | cpu_latency | throughput | fused_weights | training_logs |
140
+ |-------|----------------|--------|-------|-------------|-------------|------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
141
+ | T | 76.6* \| 75.1 | 12.1M | 0.3 | 1.8ms | 109ms | 13878 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_t_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_t_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/distill/recnext_t_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/normal/recnext_t_without_distill_300e.txt) |
142
+ | S | 79.6* \| 78.3 | 15.8M | 0.7 | 2.0ms | 188ms | 7989 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_s_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_s_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/distill/recnext_s_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/normal/recnext_s_without_distill_300e.txt) |
143
+ | B | 81.4* \| 80.3 | 19.3M | 1.1 | 2.5ms | 290ms | 4450 | [dist](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_b_distill_300e_fused.pt) \| [base](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_b_without_distill_300e_fused.pt) | [dist](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/distill/recnext_b_distill_300e.txt) \| [base](https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/logs/normal/recnext_b_without_distill_300e.txt) |
144
 
145
  > The NPU latency is measured on an iPhone 13 with models compiled by Core ML Tools.
146
  > The CPU latency is accessed on a Quad-core ARM Cortex-A57 processor in ONNX format.
147
  > And the throughput is tested on an Nvidia RTX3090 with maximum power-of-two batch size that fits in memory.
148
 
149
 
150
+ ## Latency Measurement
151
+
152
+ The latency reported in RecNeXt for iPhone 13 (iOS 18) uses the benchmark tool from [XCode 14](https://developer.apple.com/videos/play/wwdc2022/10027/).
153
+
154
+ <details>
155
+ <summary>
156
+ RecNeXt-M0
157
+ </summary>
158
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m0_224x224.png" alt="recnext_m0">
159
+ </details>
160
+
161
+ <details>
162
+ <summary>
163
+ RecNeXt-M1
164
+ </summary>
165
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m1_224x224.png" alt="recnext_m1">
166
+ </details>
167
+
168
+ <details>
169
+ <summary>
170
+ RecNeXt-M2
171
+ </summary>
172
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m2_224x224.png" alt="recnext_m2">
173
+ </details>
174
+
175
+ <details>
176
+ <summary>
177
+ RecNeXt-M3
178
+ </summary>
179
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m3_224x224.png" alt="recnext_m3">
180
+ </details>
181
+
182
+ <details>
183
+ <summary>
184
+ RecNeXt-M4
185
+ </summary>
186
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m4_224x224.png" alt="recnext_m4">
187
+ </details>
188
+
189
+ <details>
190
+ <summary>
191
+ RecNeXt-M5
192
+ </summary>
193
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m5_224x224.png" alt="recnext_m5">
194
+ </details>
195
+
196
+ <details>
197
+ <summary>
198
+ RecNeXt-A0
199
+ </summary>
200
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a0_224x224.png" alt="recnext_a0">
201
+ </details>
202
+
203
+ <details>
204
+ <summary>
205
+ RecNeXt-A1
206
+ </summary>
207
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a1_224x224.png" alt="recnext_a1">
208
+ </details>
209
+
210
+ <details>
211
+ <summary>
212
+ RecNeXt-A2
213
+ </summary>
214
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a2_224x224.png" alt="recnext_a2">
215
+ </details>
216
+
217
+ <details>
218
+ <summary>
219
+ RecNeXt-A3
220
+ </summary>
221
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a3_224x224.png" alt="recnext_a3">
222
+ </details>
223
+
224
+ <details>
225
+ <summary>
226
+ RecNeXt-A4
227
+ </summary>
228
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a4_224x224.png" alt="recnext_a4">
229
+ </details>
230
+
231
+ <details>
232
+ <summary>
233
+ RecNeXt-A5
234
+ </summary>
235
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a5_224x224.png" alt="recnext_a5">
236
+ </details>
237
+
238
+ <details>
239
+ <summary>
240
+ RecNeXt-T
241
+ </summary>
242
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/figures/latency/recnext_t_224x224.png" alt="recnext_t">
243
+ </details>
244
+
245
+ <details>
246
+ <summary>
247
+ RecNeXt-S
248
+ </summary>
249
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/figures/latency/recnext_s_224x224.png" alt="recnext_s">
250
+ </details>
251
+
252
+ <details>
253
+ <summary>
254
+ RecNeXt-B
255
+ </summary>
256
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/figures/latency/recnext_b_224x224.png" alt="recnext_b">
257
+ </details>
258
+
259
+ Tips: export the model to Core ML model
260
+ ```
261
+ python export_coreml.py --model recnext_m1 --ckpt pretrain/recnext_m1_distill_300e.pth
262
+ ```
263
+ Tips: measure the throughput on GPU
264
+ ```
265
+ python speed_gpu.py --model recnext_m1
266
+ ```
267
 
268
+ ## ImageNet (Training and Evaluation)
269
+
270
+ ### Prerequisites
271
+ `conda` virtual environment is recommended.
272
+ ```
273
+ conda create -n recnext python=3.8
274
+ pip install -r requirements.txt
275
+ ```
276
+
277
+ ### Data preparation
278
+
279
+ Download and extract ImageNet train and val images from http://image-net.org/. The training and validation data are expected to be in the `train` folder and `val` folder respectively:
280
+
281
+ ```bash
282
+ # script to extract ImageNet dataset: https://github.com/pytorch/examples/blob/main/imagenet/extract_ILSVRC.sh
283
+ # ILSVRC2012_img_train.tar (about 138 GB)
284
+ # ILSVRC2012_img_val.tar (about 6.3 GB)
285
+ ```
286
+
287
+ ```
288
+ # organize the ImageNet dataset as follows:
289
+ imagenet
290
+ ├── train
291
+ │ ├── n01440764
292
+ │ │ ├── n01440764_10026.JPEG
293
+ │ │ ├── n01440764_10027.JPEG
294
+ │ │ ├── ......
295
+ │ ├── ......
296
+ ├── val
297
+ │ ├── n01440764
298
+ │ │ ├── ILSVRC2012_val_00000293.JPEG
299
+ │ │ ├── ILSVRC2012_val_00002138.JPEG
300
+ │ │ ├── ......
301
+ │ ├── ......
302
+ ```
303
+
304
+ ### Training
305
+ To train RecNeXt-M1 on an 8-GPU machine:
306
+
307
+ ```
308
+ python -m torch.distributed.launch --nproc_per_node=8 --master_port 12346 --use_env main.py --model recnext_m1 --data-path ~/imagenet --dist-eval
309
  ```
310
+ Tips: specify your data path and model name!
311
+
312
+ ### Testing
313
+ For example, to test RecNeXt-M1:
314
+ ```
315
+ python main.py --eval --model recnext_m1 --resume pretrain/recnext_m1_distill_300e.pth --data-path ~/imagenet
316
+ ```
317
+
318
+ Use pretrained model without knowledge distillation from [HuggingFace](https://huggingface.co/suous) 🤗.
319
+ ```bash
320
+ python main.py --eval --model recnext_m1 --data-path ~/imagenet --pretrained --distillation-type none
321
+ ```
322
+
323
+ Use pretrained model with knowledge distillation from [HuggingFace](https://huggingface.co/suous) 🤗.
324
+ ```bash
325
+ python main.py --eval --model recnext_m1 --data-path ~/imagenet --pretrained --distillation-type hard
326
+ ```
327
+
328
+ ### Fused model evaluation
329
+ For example, to evaluate RecNeXt-M1 with the fused model: [![Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/suous/RecNeXt/blob/main/demo/fused_model_evaluation.ipynb)
330
+ ```
331
+ python fuse_eval.py --model recnext_m1 --resume pretrain/recnext_m1_distill_300e_fused.pt --data-path ~/imagenet
332
+ ```
333
+
334
+ ### Extract model for publishing
335
+
336
+ ```
337
+ # without distillation
338
+ python publish.py --model_name recnext_m1 --checkpoint_path pretrain/checkpoint_best.pth --epochs 300
339
+
340
+ # with distillation
341
+ python publish.py --model_name recnext_m1 --checkpoint_path pretrain/checkpoint_best.pth --epochs 300 --distillation
342
+
343
+ # fused model
344
+ python publish.py --model_name recnext_m1 --checkpoint_path pretrain/checkpoint_best.pth --epochs 300 --fused
345
+ ```
346
+
347
+ ## Downstream Tasks
348
+ [Object Detection and Instance Segmentation](https://github.com/suous/RecNeXt/blob/main/detection/README.md)<br>
349
+
350
+ | model | $AP^b$ | $AP_{50}^b$ | $AP_{75}^b$ | $AP^m$ | $AP_{50}^m$ | $AP_{75}^m$ | Latency | Ckpt | Log |
351
+ |:------|:------:|:-----------:|:-----------:|:------:|:-----------:|:-----------:|:-------:|:---------------------------------------------------------------------------------:|:----------------------------------------------------------------------------------------------:|
352
+ | M3 | 41.7 | 63.4 | 45.4 | 38.6 | 60.5 | 41.4 | 5.2ms | [M3](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m3_coco.pth) | [M3](https://raw.githubusercontent.com/suous/RecNeXt/main/detection/logs/recnext_m3_coco.json) |
353
+ | M4 | 43.5 | 64.9 | 47.7 | 39.7 | 62.1 | 42.4 | 7.6ms | [M4](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m4_coco.pth) | [M4](https://raw.githubusercontent.com/suous/RecNeXt/main/detection/logs/recnext_m4_coco.json) |
354
+ | M5 | 44.6 | 66.3 | 49.0 | 40.6 | 63.5 | 43.5 | 12.4ms | [M5](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m5_coco.pth) | [M5](https://raw.githubusercontent.com/suous/RecNeXt/main/detection/logs/recnext_m5_coco.json) |
355
+ | A3 | 42.1 | 64.1 | 46.2 | 38.8 | 61.1 | 41.6 | 8.3ms | [A3](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a3_coco.pth) | [A3](https://raw.githubusercontent.com/suous/RecNeXt/main/detection/logs/recnext_a3_coco.json) |
356
+ | A4 | 43.5 | 65.4 | 47.6 | 39.8 | 62.4 | 42.9 | 14.0ms | [A4](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a4_coco.pth) | [A4](https://raw.githubusercontent.com/suous/RecNeXt/main/detection/logs/recnext_a4_coco.json) |
357
+ | A5 | 44.4 | 66.3 | 48.9 | 40.3 | 63.3 | 43.4 | 25.3ms | [A5](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a5_coco.pth) | [A5](https://raw.githubusercontent.com/suous/RecNeXt/main/detection/logs/recnext_a5_coco.json) |
358
+
359
+ [Semantic Segmentation](https://github.com/suous/RecNeXt/blob/main/segmentation/README.md)
360
+
361
+ | Model | mIoU | Latency | Ckpt | Log |
362
+ |:-----------|:----:|:-------:|:-----------------------------------------------------------------------------------:|:---------------------------------------------------------------------------------------------------:|
363
+ | RecNeXt-M3 | 41.0 | 5.6ms | [M3](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m3_ade20k.pth) | [M3](https://raw.githubusercontent.com/suous/RecNeXt/main/segmentation/logs/recnext_m3_ade20k.json) |
364
+ | RecNeXt-M4 | 43.6 | 7.2ms | [M4](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m4_ade20k.pth) | [M4](https://raw.githubusercontent.com/suous/RecNeXt/main/segmentation/logs/recnext_m4_ade20k.json) |
365
+ | RecNeXt-M5 | 46.0 | 12.4ms | [M5](https://github.com/suous/RecNeXt/releases/download/v1.0/recnext_m5_ade20k.pth) | [M5](https://raw.githubusercontent.com/suous/RecNeXt/main/segmentation/logs/recnext_m5_ade20k.json) |
366
+ | RecNeXt-A3 | 41.9 | 8.4ms | [A3](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a3_ade20k.pth) | [A3](https://raw.githubusercontent.com/suous/RecNeXt/main/segmentation/logs/recnext_a3_ade20k.json) |
367
+ | RecNeXt-A4 | 43.0 | 14.0ms | [A4](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a4_ade20k.pth) | [A4](https://raw.githubusercontent.com/suous/RecNeXt/main/segmentation/logs/recnext_a4_ade20k.json) |
368
+ | RecNeXt-A5 | 46.5 | 25.3ms | [A5](https://github.com/suous/RecNeXt/releases/download/v2.0/recnext_a5_ade20k.pth) | [A5](https://raw.githubusercontent.com/suous/RecNeXt/main/segmentation/logs/recnext_a5_ade20k.json) |
369
+
370
+ ## Ablation Study
371
+
372
+ ### Overall Experiments
373
+
374
+ ![ablation](https://raw.githubusercontent.com/suous/RecNeXt/main/figures/ablation.png)
375
+
376
+ <details>
377
+ <summary>
378
+ <span style="font-size: larger; ">Ablation Logs</span>
379
+ </summary>
380
+
381
+ <pre>
382
+ logs/ablation
383
+ ├── 224
384
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_3x3_7464.txt">recnext_m1_120e_224x224_3x3_7464.txt</a>
385
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_7x7_7552.txt">recnext_m1_120e_224x224_7x7_7552.txt</a>
386
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_bxb_7541.txt">recnext_m1_120e_224x224_bxb_7541.txt</a>
387
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_rec_3x3_7548.txt">recnext_m1_120e_224x224_rec_3x3_7548.txt</a>
388
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_rec_5x5_7603.txt">recnext_m1_120e_224x224_rec_5x5_7603.txt</a>
389
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_rec_7x7_7567.txt">recnext_m1_120e_224x224_rec_7x7_7567.txt</a>
390
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_rec_7x7_nearest_7571.txt">recnext_m1_120e_224x224_rec_7x7_nearest_7571.txt</a>
391
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_rec_7x7_nearest_ssm_7593.txt">recnext_m1_120e_224x224_rec_7x7_nearest_ssm_7593.txt</a>
392
+ │ └── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/224/recnext_m1_120e_224x224_rec_7x7_unpool_7548.txt">recnext_m1_120e_224x224_rec_7x7_unpool_7548.txt</a>
393
+ └── 384
394
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_3x3_7635.txt">recnext_m1_120e_384x384_3x3_7635.txt</a>
395
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_7x7_7742.txt">recnext_m1_120e_384x384_7x7_7742.txt</a>
396
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_bxb_7800.txt">recnext_m1_120e_384x384_bxb_7800.txt</a>
397
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_rec_3x3_7772.txt">recnext_m1_120e_384x384_rec_3x3_7772.txt</a>
398
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_rec_5x5_7811.txt">recnext_m1_120e_384x384_rec_5x5_7811.txt</a>
399
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_rec_7x7_7803.txt">recnext_m1_120e_384x384_rec_7x7_7803.txt</a>
400
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_rec_convtrans_3x3_basic_7726.txt">recnext_m1_120e_384x384_rec_convtrans_3x3_basic_7726.txt</a>
401
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_rec_convtrans_5x5_basic_7787.txt">recnext_m1_120e_384x384_rec_convtrans_5x5_basic_7787.txt</a>
402
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_rec_convtrans_7x7_basic_7824.txt">recnext_m1_120e_384x384_rec_convtrans_7x7_basic_7824.txt</a>
403
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_rec_convtrans_7x7_group_7791.txt">recnext_m1_120e_384x384_rec_convtrans_7x7_group_7791.txt</a>
404
+ └── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/logs/ablation/384/recnext_m1_120e_384x384_rec_convtrans_7x7_split_7683.txt">recnext_m1_120e_384x384_rec_convtrans_7x7_split_7683.txt</a>
405
+ </pre>
406
+ </details>
407
+
408
+ <details>
409
+ <summary>
410
+ <span style="font-size: larger; ">RecConv Recurrent Aggregation</span>
411
+ </summary>
412
+
413
+ ```python
414
+ class RecConv2d(nn.Module):
415
+ def __init__(self, in_channels, kernel_size=5, bias=False, level=1, mode='nearest'):
416
+ super().__init__()
417
+ self.level = level
418
+ self.mode = mode
419
+ kwargs = {
420
+ 'in_channels': in_channels,
421
+ 'out_channels': in_channels,
422
+ 'groups': in_channels,
423
+ 'kernel_size': kernel_size,
424
+ 'padding': kernel_size // 2,
425
+ 'bias': bias
426
+ }
427
+ self.n = nn.Conv2d(stride=2, **kwargs)
428
+ self.a = nn.Conv2d(**kwargs) if level >1 else None
429
+ self.b = nn.Conv2d(**kwargs)
430
+ self.c = nn.Conv2d(**kwargs)
431
+ self.d = nn.Conv2d(**kwargs)
432
+
433
+ def forward(self, x):
434
+ # 1. Generate Multi-scale Features.
435
+ fs = [x]
436
+ for _ in range(self.level):
437
+ fs.append(self.n(fs[-1]))
438
+
439
+ # 2. Multi-scale Recurrent Aggregation.
440
+ h = None
441
+ for i, o in reversed(list(zip(fs[1:], fs[:-1]))):
442
+ h = self.a(h) + self.b(i) if h is not None else self.b(i)
443
+ h = nn.functional.interpolate(h, size=o.shape[2:], mode=self.mode)
444
+ return self.c(h) + self.d(x)
445
+ ```
446
+ </details>
447
+
448
+ ### RecConv Variants
449
+
450
+ <div style="display: flex; justify-content: space-between;">
451
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/RecConvB.png" alt="RecConvB" style="width: 49%;">
452
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/RecConvC.png" alt="RecConvC" style="width: 49%;">
453
+ </div>
454
+
455
+
456
+ <details>
457
+ <summary>
458
+ <span style="font-size: larger; ">RecConv Variant Details</span>
459
+ </summary>
460
+
461
+ - **RecConv using group convolutions**
462
+
463
+ ```python
464
+ # RecConv Variant A
465
+ # recursive decomposition on both spatial and channel dimensions
466
+ # downsample and upsample through group convolutions
467
+ class RecConv2d(nn.Module):
468
+ def __init__(self, in_channels, kernel_size=5, bias=False, level=2):
469
+ super().__init__()
470
+ self.level = level
471
+ kwargs = {'kernel_size': kernel_size, 'padding': kernel_size // 2, 'bias': bias}
472
+ downs = []
473
+ for l in range(level):
474
+ i_channels = in_channels // (2 ** l)
475
+ o_channels = in_channels // (2 ** (l+1))
476
+ downs.append(nn.Conv2d(in_channels=i_channels, out_channels=o_channels, groups=o_channels, stride=2, **kwargs))
477
+ self.downs = nn.ModuleList(downs)
478
+
479
+ convs = []
480
+ for l in range(level+1):
481
+ channels = in_channels // (2 ** l)
482
+ convs.append(nn.Conv2d(in_channels=channels, out_channels=channels, groups=channels, **kwargs))
483
+ self.convs = nn.ModuleList(reversed(convs))
484
+
485
+ # this is the simplest modification, only support resoltions like 256, 384, etc
486
+ kwargs['kernel_size'] = kernel_size + 1
487
+ ups = []
488
+ for l in range(level):
489
+ i_channels = in_channels // (2 ** (l+1))
490
+ o_channels = in_channels // (2 ** l)
491
+ ups.append(nn.ConvTranspose2d(in_channels=i_channels, out_channels=o_channels, groups=i_channels, stride=2, **kwargs))
492
+ self.ups = nn.ModuleList(reversed(ups))
493
+
494
+ def forward(self, x):
495
+ i = x
496
+ features = []
497
+ for down in self.downs:
498
+ x, s = down(x), x.shape[2:]
499
+ features.append((x, s))
500
+
501
+ x = 0
502
+ for conv, up, (f, s) in zip(self.convs, self.ups, reversed(features)):
503
+ x = up(conv(f + x))
504
+ return self.convs[self.level](i + x)
505
+ ```
506
+
507
+ - **RecConv using channel-wise concatenation**
508
+
509
+ ```python
510
+ # recursive decomposition on both spatial and channel dimensions
511
+ # downsample using channel-wise split, followed by depthwise convolution with a stride of 2
512
+ # upsample through channel-wise concatenation
513
+ class RecConv2d(nn.Module):
514
+ def __init__(self, in_channels, kernel_size=5, bias=False, level=2):
515
+ super().__init__()
516
+ self.level = level
517
+ kwargs = {'kernel_size': kernel_size, 'padding': kernel_size // 2, 'bias': bias}
518
+ downs = []
519
+ for l in range(level):
520
+ channels = in_channels // (2 ** (l+1))
521
+ downs.append(nn.Conv2d(in_channels=channels, out_channels=channels, groups=channels, stride=2, **kwargs))
522
+ self.downs = nn.ModuleList(downs)
523
+
524
+ convs = []
525
+ for l in range(level+1):
526
+ channels = in_channels // (2 ** l)
527
+ convs.append(nn.Conv2d(in_channels=channels, out_channels=channels, groups=channels, **kwargs))
528
+ self.convs = nn.ModuleList(reversed(convs))
529
+
530
+ . # this is the simplest modification, only support resoltions like 256, 384, etc
531
+ kwargs['kernel_size'] = kernel_size + 1
532
+ ups = []
533
+ for l in range(level):
534
+ channels = in_channels // (2 ** (l+1))
535
+ ups.append(nn.ConvTranspose2d(in_channels=channels, out_channels=channels, groups=channels, stride=2, **kwargs))
536
+ self.ups = nn.ModuleList(reversed(ups))
537
+
538
+ def forward(self, x):
539
+ features = []
540
+ for down in self.downs:
541
+ r, x = torch.chunk(x, 2, dim=1)
542
+ x, s = down(x), x.shape[2:]
543
+ features.append((r, s))
544
+
545
+ for conv, up, (r, s) in zip(self.convs, self.ups, reversed(features)):
546
+ x = torch.cat([r, up(conv(x))], dim=1)
547
+ return self.convs[self.level](x)
548
+ ```
549
+ </details>
550
+
551
+ ### RecConv Beyond
552
+
553
+ We apply RecConv to [MLLA](https://github.com/LeapLabTHU/MLLA) small variants, replacing linear attention and downsampling layers.
554
+ Result in higher throughput and less training memory usage.
555
+
556
+ <details>
557
+ <summary>
558
+ <span style="font-size: larger; ">Ablation Logs</span>
559
+ </summary>
560
+
561
+ <pre>
562
+ mlla/logs
563
+ ├── 1_mlla_nano
564
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/1_mlla_nano/01_baseline.txt">01_baseline.txt</a>
565
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/1_mlla_nano/02_recconv_5x5_conv_trans.txt">02_recconv_5x5_conv_trans.txt</a>
566
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/1_mlla_nano/03_recconv_5x5_nearest_interp.txt">03_recconv_5x5_nearest_interp.txt</a>
567
+ │ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/1_mlla_nano/04_recattn_nearest_interp.txt">04_recattn_nearest_interp.txt</a>
568
+ │ └── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/1_mlla_nano/05_recattn_nearest_interp_simplify.txt">05_recattn_nearest_interp_simplify.txt</a>
569
+ └── 2_mlla_mini
570
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/2_mlla_mini/01_baseline.txt">01_baseline.txt</a>
571
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/2_mlla_mini/02_recconv_5x5_conv_trans.txt">02_recconv_5x5_conv_trans.txt</a>
572
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/2_mlla_mini/03_recconv_5x5_nearest_interp.txt">03_recconv_5x5_nearest_interp.txt</a>
573
+ ├── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/2_mlla_mini/04_recattn_nearest_interp.txt">04_recattn_nearest_interp.txt</a>
574
+ └── <a style="text-decoration:none" href="https://raw.githubusercontent.com/suous/RecNeXt/main/mlla/logs/2_mlla_mini/05_recattn_nearest_interp_simplify.txt">05_recattn_nearest_interp_simplify.txt</a>
575
+ </pre>
576
+ </details>
577
 
578
  ## Limitations
579
 
 
583
 
584
  ## Acknowledgement
585
 
586
+ Classification (ImageNet) code base is partly built with [LeViT](https://github.com/facebookresearch/LeViT), [PoolFormer](https://github.com/sail-sg/poolformer), [EfficientFormer](https://github.com/snap-research/EfficientFormer), [RepViT](https://github.com/THU-MIG/RepViT), [LSNet](https://github.com/jameslahm/lsnet), [MLLA](https://github.com/LeapLabTHU/MLLA), and [MogaNet](https://github.com/Westlake-AI/MogaNet).
587
+
588
+ The detection and segmentation pipeline is from [MMCV](https://github.com/open-mmlab/mmcv) ([MMDetection](https://github.com/open-mmlab/mmdetection) and [MMSegmentation](https://github.com/open-mmlab/mmsegmentation)).
589
+
590
+ Thanks for the great implementations!
591
 
592
+ ## Citation
593
 
594
+ ```BibTeX
595
+ @misc{zhao2024recnext,
596
+ title={RecConv: Efficient Recursive Convolutions for Multi-Frequency Representations},
597
+ author={Mingshu Zhao and Yi Luo and Yong Ouyang},
598
+ year={2024},
599
+ eprint={2412.19628},
600
+ archivePrefix={arXiv},
601
+ primaryClass={cs.CV}
602
+ }
603
+ ```