Hugging Face's logo

suous
/
recnext_t_share_channel.dist_300e_in1k

Image Classification
timm
PyTorch
Safetensors
Transformers
English
vision
Model card Files
xet
Community
suous commited on
Commit
ecad43c
·
verified ·
1 Parent(s): 67b0dcd

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +590 -3
README.md CHANGED
@@ -1,9 +1,596 @@
1
  ---
 
 
 
 
 
 
 
 
 
2
  tags:
 
3
  - image-classification
 
4
  - timm
5
  - transformers
6
- library_name: timm
7
- license: apache-2.0
8
  ---
9
- # Model card for recnext_t_share_channel.dist_300e_in1k
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ datasets:
3
+ - imagenet-1k
4
+ language: en
5
+ library_name: timm
6
+ license: apache-2.0
7
+ metrics:
8
+ - accuracy
9
+ model_name: recnext_t_share_channel
10
+ pipeline_tag: image-classification
11
  tags:
12
+ - vision
13
  - image-classification
14
+ - pytorch
15
  - timm
16
  - transformers
 
 
17
  ---
18
+
19
+ # Model Card for RecNeXt-CHANNEL (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
+
27
+ <div style="display: flex; justify-content: space-between;">
28
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/refs/heads/main/figures/RecConvA.png" alt="RecConvA" style="width: 52%;">
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
35
+ - **Model Series**: M
36
+ - **Model Stats**:
37
+ - **Parameters**: N/A
38
+ - **MACs**: N/A
39
+ - **Latency**: N/A (iPhone 13, iOS 18)
40
+ - **Image Size**: 224x224
41
+
42
+ - **Architecture Configuration**:
43
+ - **Embedding Dimensions**: N/A
44
+ - **Depths**: N/A
45
+ - **MLP Ratio**: 2
46
+
47
+ - **Paper**: [RecConv: Efficient Recursive Convolutions for Multi-Frequency Representations](https://arxiv.org/abs/2412.19628)
48
+
49
+ - **Code**: https://github.com/suous/RecNeXt
50
+
51
+ - **Dataset**: ImageNet-1K
52
+
53
+ ## Recent Updates
54
+
55
+ **UPDATES** 🔥
56
+ - **2025/07/23**: Added a simple architecture, the overall design follows [LSNet](https://github.com/jameslahm/lsnet).
57
+ - **2025/07/04**: Uploaded classification models to [HuggingFace](https://huggingface.co/suous)🤗.
58
+ - **2025/07/01**: Added more comparisons with [LSNet](https://github.com/jameslahm/lsnet).
59
+ - **2025/06/27**: Added **A** series code and logs, replacing convolution with linear attention.
60
+ - **2025/03/19**: Added more ablation study results, including using attention with RecConv design.
61
+ - **2025/01/02**: Uploaded checkpoints and training logs of RecNeXt-M0.
62
+ - **2024/12/29**: Uploaded checkpoints and training logs of RecNeXt-M1 - M5.
63
+
64
+ ## Model Usage
65
+
66
+ ### Image Classification
67
+
68
+ ```python
69
+ from urllib.request import urlopen
70
+ from PIL import Image
71
+ import timm
72
+ import torch
73
+
74
+ img = Image.open(urlopen(
75
+ 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
76
+ ))
77
+
78
+ model = timm.create_model('recnext_t_share_channel', pretrained=True, distillation=True)
79
+ model = model.eval()
80
+
81
+ # get model specific transforms (normalization, resize)
82
+ data_config = timm.data.resolve_model_data_config(model)
83
+ transforms = timm.data.create_transform(**data_config, is_training=False)
84
+
85
+ output = model(transforms(img).unsqueeze(0)) # unsqueeze single image into batch of 1
86
+
87
+ top5_probabilities, top5_class_indices = torch.topk(output.softmax(dim=1) * 100, k=5)
88
+ ```
89
+
90
+ ### Converting to Inference Mode
91
+
92
+ ```python
93
+ import utils
94
+
95
+ # Convert training-time model to inference structure, fuse batchnorms
96
+ utils.replace_batchnorm(model)
97
+ ```
98
+ ## Model Comparison
99
+
100
+ ### Classification
101
+
102
+ We introduce two series of models: the **A** series uses linear attention and nearest interpolation, while the **M** series employs convolution and bilinear interpolation for simplicity and broader hardware compatibility (e.g., to address suboptimal nearest interpolation support in some iOS versions).
103
+
104
+ > **dist**: distillation; **base**: without distillation (all models are trained over 300 epochs).
105
+
106
+ | model | top_1_accuracy | params | gmacs | npu_latency | cpu_latency | throughput | fused_weights | training_logs |
107
+ |-------|----------------|--------|-------|-------------|-------------|------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
108
+ | 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://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) |
109
+ | 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) |
110
+ | 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) |
111
+ | 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) |
112
+ | 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://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) |
113
+ | 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://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) |
114
+ | 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://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) |
115
+ | 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://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) |
116
+ | 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://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) |
117
+ | 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://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) |
118
+ | 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) |
119
+ | 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://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) |
120
+
121
+ ### Comparison with [LSNet](https://github.com/jameslahm/lsnet)
122
+
123
+ 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.
124
+ Our motivation for doing so is to reduce the computational cost of token mixers and minimize feature redundancy in the final stage.
125
+
126
+ ![Architecture](https://raw.githubusercontent.com/suous/RecNeXt/refs/heads/main/lsnet/figures/architecture.png)
127
+
128
+ #### With **Shared-Channel Blocks**
129
+
130
+ | model | top_1_accuracy | params | gmacs | npu_latency | cpu_latency | throughput | fused_weights | training_logs |
131
+ |-------|----------------|--------|-------|-------------|-------------|------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
132
+ | 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://github.com/suous/RecNeXt/blob/main/lsnet/logs/distill/recnext_t_share_channel_distill_300e.txt) \| [norm](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/normal/recnext_t_share_channel_without_distill_300e.txt) |
133
+ | 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://github.com/suous/RecNeXt/blob/main/lsnet/logs/distill/recnext_s_share_channel_distill_300e.txt) \| [norm](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/normal/recnext_s_share_channel_without_distill_300e.txt) |
134
+ | 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://github.com/suous/RecNeXt/blob/main/lsnet/logs/distill/recnext_b_share_channel_distill_300e.txt) \| [norm](https://github.com/suous/RecNeXt/blob/main/lsnet/logs/normal/recnext_b_share_channel_without_distill_300e.txt) |
135
+
136
+ #### Without **Shared-Channel Blocks**
137
+
138
+ | model | top_1_accuracy | params | gmacs | npu_latency | cpu_latency | throughput | fused_weights | training_logs |
139
+ |-------|----------------|--------|-------|-------------|-------------|------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
140
+ | 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://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) |
141
+ | 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://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) |
142
+ | 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://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) |
143
+
144
+ > The NPU latency is measured on an iPhone 13 with models compiled by Core ML Tools.
145
+ > The CPU latency is accessed on a Quad-core ARM Cortex-A57 processor in ONNX format.
146
+ > And the throughput is tested on an Nvidia RTX3090 with maximum power-of-two batch size that fits in memory.
147
+
148
+
149
+ ## Latency Measurement
150
+
151
+ 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/).
152
+
153
+ <details>
154
+ <summary>
155
+ RecNeXt-M0
156
+ </summary>
157
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m0_224x224.png" alt="recnext_m0">
158
+ </details>
159
+
160
+ <details>
161
+ <summary>
162
+ RecNeXt-M1
163
+ </summary>
164
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m1_224x224.png" alt="recnext_m1">
165
+ </details>
166
+
167
+ <details>
168
+ <summary>
169
+ RecNeXt-M2
170
+ </summary>
171
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m2_224x224.png" alt="recnext_m2">
172
+ </details>
173
+
174
+ <details>
175
+ <summary>
176
+ RecNeXt-M3
177
+ </summary>
178
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m3_224x224.png" alt="recnext_m3">
179
+ </details>
180
+
181
+ <details>
182
+ <summary>
183
+ RecNeXt-M4
184
+ </summary>
185
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m4_224x224.png" alt="recnext_m4">
186
+ </details>
187
+
188
+ <details>
189
+ <summary>
190
+ RecNeXt-M5
191
+ </summary>
192
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_m5_224x224.png" alt="recnext_m5">
193
+ </details>
194
+
195
+ <details>
196
+ <summary>
197
+ RecNeXt-A0
198
+ </summary>
199
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a0_224x224.png" alt="recnext_a0">
200
+ </details>
201
+
202
+ <details>
203
+ <summary>
204
+ RecNeXt-A1
205
+ </summary>
206
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a1_224x224.png" alt="recnext_a1">
207
+ </details>
208
+
209
+ <details>
210
+ <summary>
211
+ RecNeXt-A2
212
+ </summary>
213
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a2_224x224.png" alt="recnext_a2">
214
+ </details>
215
+
216
+ <details>
217
+ <summary>
218
+ RecNeXt-A3
219
+ </summary>
220
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a3_224x224.png" alt="recnext_a3">
221
+ </details>
222
+
223
+ <details>
224
+ <summary>
225
+ RecNeXt-A4
226
+ </summary>
227
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a4_224x224.png" alt="recnext_a4">
228
+ </details>
229
+
230
+ <details>
231
+ <summary>
232
+ RecNeXt-A5
233
+ </summary>
234
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/latency/recnext_a5_224x224.png" alt="recnext_a5">
235
+ </details>
236
+
237
+ <details>
238
+ <summary>
239
+ RecNeXt-T
240
+ </summary>
241
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/figures/latency/recnext_t_224x224.png" alt="recnext_t">
242
+ </details>
243
+
244
+ <details>
245
+ <summary>
246
+ RecNeXt-S
247
+ </summary>
248
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/figures/latency/recnext_s_224x224.png" alt="recnext_s">
249
+ </details>
250
+
251
+ <details>
252
+ <summary>
253
+ RecNeXt-B
254
+ </summary>
255
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/lsnet/figures/latency/recnext_b_224x224.png" alt="recnext_b">
256
+ </details>
257
+
258
+ Tips: export the model to Core ML model
259
+ ```
260
+ python export_coreml.py --model recnext_m1 --ckpt pretrain/recnext_m1_distill_300e.pth
261
+ ```
262
+ Tips: measure the throughput on GPU
263
+ ```
264
+ python speed_gpu.py --model recnext_m1
265
+ ```
266
+
267
+ ## ImageNet (Training and Evaluation)
268
+
269
+ ### Prerequisites
270
+ `conda` virtual environment is recommended.
271
+ ```
272
+ conda create -n recnext python=3.8
273
+ pip install -r requirements.txt
274
+ ```
275
+
276
+ ### Data preparation
277
+
278
+ 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:
279
+
280
+ ```bash
281
+ # script to extract ImageNet dataset: https://github.com/pytorch/examples/blob/main/imagenet/extract_ILSVRC.sh
282
+ # ILSVRC2012_img_train.tar (about 138 GB)
283
+ # ILSVRC2012_img_val.tar (about 6.3 GB)
284
+ ```
285
+
286
+ ```
287
+ # organize the ImageNet dataset as follows:
288
+ imagenet
289
+ ├── train
290
+ │ ├── n01440764
291
+ │ │ ├── n01440764_10026.JPEG
292
+ │ │ ├── n01440764_10027.JPEG
293
+ │ │ ├── ......
294
+ │ ├── ......
295
+ ├── val
296
+ │ ├── n01440764
297
+ │ │ ├── ILSVRC2012_val_00000293.JPEG
298
+ │ │ ├── ILSVRC2012_val_00002138.JPEG
299
+ │ │ ├── ......
300
+ │ ├── ......
301
+ ```
302
+
303
+ ### Training
304
+ To train RecNeXt-M1 on an 8-GPU machine:
305
+
306
+ ```
307
+ python -m torch.distributed.launch --nproc_per_node=8 --master_port 12346 --use_env main.py --model recnext_m1 --data-path ~/imagenet --dist-eval
308
+ ```
309
+ Tips: specify your data path and model name!
310
+
311
+ ### Testing
312
+ For example, to test RecNeXt-M1:
313
+ ```
314
+ python main.py --eval --model recnext_m1 --resume pretrain/recnext_m1_distill_300e.pth --data-path ~/imagenet
315
+ ```
316
+
317
+ Use pretrained model without knowledge distillation from [HuggingFace](https://huggingface.co/suous) 🤗.
318
+ ```bash
319
+ python main.py --eval --model recnext_m1 --data-path ~/imagenet --pretrained --distillation-type none
320
+ ```
321
+
322
+ Use pretrained model with knowledge distillation from [HuggingFace](https://huggingface.co/suous) 🤗.
323
+ ```bash
324
+ python main.py --eval --model recnext_m1 --data-path ~/imagenet --pretrained --distillation-type hard
325
+ ```
326
+
327
+ ### Fused model evaluation
328
+ 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)
329
+ ```
330
+ python fuse_eval.py --model recnext_m1 --resume pretrain/recnext_m1_distill_300e_fused.pt --data-path ~/imagenet
331
+ ```
332
+
333
+ ### Extract model for publishing
334
+
335
+ ```
336
+ # without distillation
337
+ python publish.py --model_name recnext_m1 --checkpoint_path pretrain/checkpoint_best.pth --epochs 300
338
+
339
+ # with distillation
340
+ python publish.py --model_name recnext_m1 --checkpoint_path pretrain/checkpoint_best.pth --epochs 300 --distillation
341
+
342
+ # fused model
343
+ python publish.py --model_name recnext_m1 --checkpoint_path pretrain/checkpoint_best.pth --epochs 300 --fused
344
+ ```
345
+
346
+ ## Downstream Tasks
347
+ [Object Detection and Instance Segmentation](https://github.com/suous/RecNeXt/blob/main/detection/README.md)<br>
348
+
349
+ | model | $AP^b$ | $AP_{50}^b$ | $AP_{75}^b$ | $AP^m$ | $AP_{50}^m$ | $AP_{75}^m$ | Latency | Ckpt | Log |
350
+ |:-----------|:------:|:-----------:|:-----------:|:------:|:-----------:|:-----------:|:-------:|:---------------------------------------------------------------------------------:|:----------------------------------------------------------------------------------------------:|
351
+ | RecNeXt-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) |
352
+ | RecNeXt-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) |
353
+ | RecNeXt-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) |
354
+ | RecNeXt-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) |
355
+ | RecNeXt-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) |
356
+ | RecNeXt-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) |
357
+
358
+ [Semantic Segmentation](https://github.com/suous/RecNeXt/blob/main/segmentation/README.md)
359
+
360
+ | Model | mIoU | Latency | Ckpt | Log |
361
+ |:-----------|:----:|:-------:|:-----------------------------------------------------------------------------------:|:---------------------------------------------------------------------------------------------------:|
362
+ | 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) |
363
+ | 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) |
364
+ | 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) |
365
+ | 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) |
366
+ | 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) |
367
+ | 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) |
368
+
369
+ ## Ablation Study
370
+
371
+ ### Overall Experiments
372
+
373
+ ![ablation](https://raw.githubusercontent.com/suous/RecNeXt/main/figures/ablation.png)
374
+
375
+ <details>
376
+ <summary>
377
+ <span style="font-size: larger; ">Ablation Logs</span>
378
+ </summary>
379
+
380
+ <pre>
381
+ logs/ablation
382
+ ├── 224
383
+ │ ├── <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>
384
+ │ ├── <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>
385
+ │ ├── <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>
386
+ │ ├── <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>
387
+ │ ├── <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>
388
+ │ ├── <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>
389
+ │ ├── <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>
390
+ │ ├── <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>
391
+ │ └── <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>
392
+ └── 384
393
+ ├── <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>
394
+ ├── <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>
395
+ ├── <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>
396
+ ├── <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>
397
+ ├── <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>
398
+ ├── <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>
399
+ ├── <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>
400
+ ├── <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>
401
+ ├── <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>
402
+ ├── <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>
403
+ └── <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>
404
+ </pre>
405
+ </details>
406
+
407
+ <details>
408
+ <summary>
409
+ <span style="font-size: larger; ">RecConv Recurrent Aggregation</span>
410
+ </summary>
411
+
412
+ ```python
413
+ class RecConv2d(nn.Module):
414
+ def __init__(self, in_channels, kernel_size=5, bias=False, level=1, mode='nearest'):
415
+ super().__init__()
416
+ self.level = level
417
+ self.mode = mode
418
+ kwargs = {
419
+ 'in_channels': in_channels,
420
+ 'out_channels': in_channels,
421
+ 'groups': in_channels,
422
+ 'kernel_size': kernel_size,
423
+ 'padding': kernel_size // 2,
424
+ 'bias': bias
425
+ }
426
+ self.n = nn.Conv2d(stride=2, **kwargs)
427
+ self.a = nn.Conv2d(**kwargs) if level >1 else None
428
+ self.b = nn.Conv2d(**kwargs)
429
+ self.c = nn.Conv2d(**kwargs)
430
+ self.d = nn.Conv2d(**kwargs)
431
+
432
+ def forward(self, x):
433
+ # 1. Generate Multi-scale Features.
434
+ fs = [x]
435
+ for _ in range(self.level):
436
+ fs.append(self.n(fs[-1]))
437
+
438
+ # 2. Multi-scale Recurrent Aggregation.
439
+ h = None
440
+ for i, o in reversed(list(zip(fs[1:], fs[:-1]))):
441
+ h = self.a(h) + self.b(i) if h is not None else self.b(i)
442
+ h = nn.functional.interpolate(h, size=o.shape[2:], mode=self.mode)
443
+ return self.c(h) + self.d(x)
444
+ ```
445
+ </details>
446
+
447
+ ### RecConv Variants
448
+
449
+ <div style="display: flex; justify-content: space-between;">
450
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/RecConvB.png" alt="RecConvB" style="width: 49%;">
451
+ <img src="https://raw.githubusercontent.com/suous/RecNeXt/main/figures/RecConvC.png" alt="RecConvC" style="width: 49%;">
452
+ </div>
453
+
454
+
455
+ <details>
456
+ <summary>
457
+ <span style="font-size: larger; ">RecConv Variant Details</span>
458
+ </summary>
459
+
460
+ - **RecConv using group convolutions**
461
+
462
+ ```python
463
+ # RecConv Variant A
464
+ # recursive decomposition on both spatial and channel dimensions
465
+ # downsample and upsample through group convolutions
466
+ class RecConv2d(nn.Module):
467
+ def __init__(self, in_channels, kernel_size=5, bias=False, level=2):
468
+ super().__init__()
469
+ self.level = level
470
+ kwargs = {'kernel_size': kernel_size, 'padding': kernel_size // 2, 'bias': bias}
471
+ downs = []
472
+ for l in range(level):
473
+ i_channels = in_channels // (2 ** l)
474
+ o_channels = in_channels // (2 ** (l+1))
475
+ downs.append(nn.Conv2d(in_channels=i_channels, out_channels=o_channels, groups=o_channels, stride=2, **kwargs))
476
+ self.downs = nn.ModuleList(downs)
477
+
478
+ convs = []
479
+ for l in range(level+1):
480
+ channels = in_channels // (2 ** l)
481
+ convs.append(nn.Conv2d(in_channels=channels, out_channels=channels, groups=channels, **kwargs))
482
+ self.convs = nn.ModuleList(reversed(convs))
483
+
484
+ # this is the simplest modification, only support resoltions like 256, 384, etc
485
+ kwargs['kernel_size'] = kernel_size + 1
486
+ ups = []
487
+ for l in range(level):
488
+ i_channels = in_channels // (2 ** (l+1))
489
+ o_channels = in_channels // (2 ** l)
490
+ ups.append(nn.ConvTranspose2d(in_channels=i_channels, out_channels=o_channels, groups=i_channels, stride=2, **kwargs))
491
+ self.ups = nn.ModuleList(reversed(ups))
492
+
493
+ def forward(self, x):
494
+ i = x
495
+ features = []
496
+ for down in self.downs:
497
+ x, s = down(x), x.shape[2:]
498
+ features.append((x, s))
499
+
500
+ x = 0
501
+ for conv, up, (f, s) in zip(self.convs, self.ups, reversed(features)):
502
+ x = up(conv(f + x))
503
+ return self.convs[self.level](i + x)
504
+ ```
505
+
506
+ - **RecConv using channel-wise concatenation**
507
+
508
+ ```python
509
+ # recursive decomposition on both spatial and channel dimensions
510
+ # downsample using channel-wise split, followed by depthwise convolution with a stride of 2
511
+ # upsample through channel-wise concatenation
512
+ class RecConv2d(nn.Module):
513
+ def __init__(self, in_channels, kernel_size=5, bias=False, level=2):
514
+ super().__init__()
515
+ self.level = level
516
+ kwargs = {'kernel_size': kernel_size, 'padding': kernel_size // 2, 'bias': bias}
517
+ downs = []
518
+ for l in range(level):
519
+ channels = in_channels // (2 ** (l+1))
520
+ downs.append(nn.Conv2d(in_channels=channels, out_channels=channels, groups=channels, stride=2, **kwargs))
521
+ self.downs = nn.ModuleList(downs)
522
+
523
+ convs = []
524
+ for l in range(level+1):
525
+ channels = in_channels // (2 ** l)
526
+ convs.append(nn.Conv2d(in_channels=channels, out_channels=channels, groups=channels, **kwargs))
527
+ self.convs = nn.ModuleList(reversed(convs))
528
+
529
+ . # this is the simplest modification, only support resoltions like 256, 384, etc
530
+ kwargs['kernel_size'] = kernel_size + 1
531
+ ups = []
532
+ for l in range(level):
533
+ channels = in_channels // (2 ** (l+1))
534
+ ups.append(nn.ConvTranspose2d(in_channels=channels, out_channels=channels, groups=channels, stride=2, **kwargs))
535
+ self.ups = nn.ModuleList(reversed(ups))
536
+
537
+ def forward(self, x):
538
+ features = []
539
+ for down in self.downs:
540
+ r, x = torch.chunk(x, 2, dim=1)
541
+ x, s = down(x), x.shape[2:]
542
+ features.append((r, s))
543
+
544
+ for conv, up, (r, s) in zip(self.convs, self.ups, reversed(features)):
545
+ x = torch.cat([r, up(conv(x))], dim=1)
546
+ return self.convs[self.level](x)
547
+ ```
548
+ </details>
549
+
550
+ ### RecConv Beyond
551
+
552
+ We apply RecConv to [MLLA](https://github.com/LeapLabTHU/MLLA) small variants, replacing linear attention and downsampling layers.
553
+ Result in higher throughput and less training memory usage.
554
+
555
+ <pre>
556
+ mlla/logs
557
+ ├── 1_mlla_nano
558
+ │ ├── <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>
559
+ │ ├── <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>
560
+ │ ├── <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>
561
+ │ ├── <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>
562
+ │ └── <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>
563
+ └── 2_mlla_mini
564
+ ├── <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>
565
+ ├── <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>
566
+ ├── <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>
567
+ ├── <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>
568
+ └── <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>
569
+ </pre>
570
+
571
+ ## Limitations
572
+
573
+ 1. RecNeXt exhibits the lowest **throughput** among models of comparable parameter size due to extensive use of bilinear interpolation, which can be mitigated by employing transposed convolution.
574
+ 2. The recursive decomposition may introduce **numerical instability** during mixed precision training, which can be alleviated by using fixed-point or BFloat16 arithmetic.
575
+ 3. **Compatibility issues** with bilinear interpolation and transposed convolution on certain iOS versions may also result in performance degradation.
576
+
577
+ ## Acknowledgement
578
+
579
+ 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), and [MogaNet](https://github.com/Westlake-AI/MogaNet).
580
+
581
+ 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)).
582
+
583
+ Thanks for the great implementations!
584
+
585
+ ## Citation
586
+
587
+ ```BibTeX
588
+ @misc{zhao2024recnext,
589
+ title={RecConv: Efficient Recursive Convolutions for Multi-Frequency Representations},
590
+ author={Mingshu Zhao and Yi Luo and Yong Ouyang},
591
+ year={2024},
592
+ eprint={2412.19628},
593
+ archivePrefix={arXiv},
594
+ primaryClass={cs.CV}
595
+ }
596
+ ```