Hugging Face's logo

STMicroelectronics
/
sqnxt_pt

Image Classification
Model card Files
xet
Community
FBAGSTM commited on
Commit
8a40fda
·
verified ·
1 Parent(s): 214e212

Release AI-ModelZoo-4.0.0

Browse files
Files changed (1) hide show
  1. README.md +135 -3
README.md CHANGED
@@ -1,3 +1,135 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ pipeline_tag: image-classification
4
+ ---
5
+ # SqueezeNext
6
+
7
+ ## **Use case** : `Image classification`
8
+
9
+ # Model description
10
+
11
+
12
+
13
+ SqueezeNext is the successor to SqueezeNet, offering **improved accuracy through skip connections, bottleneck modules, and separable convolutions**. It is specifically designed for hardware efficiency.
14
+
15
+ The architecture employs a **two-stage bottleneck** with 1x1 squeeze followed by 1x1-3x3 expand patterns, with **skip connections** added for improved gradient flow. **Separable convolutions** further reduce computational cost, and the **hardware-aware design** is optimized for specific hardware platforms.
16
+
17
+ SqueezeNext is ideal for applications requiring SqueezeNet-style compactness with better accuracy, and hardware platforms with specific optimization targets.
18
+
19
+ (source: https://arxiv.org/abs/1803.10615)
20
+
21
+ The model is quantized to **int8** using **ONNX Runtime** and exported for efficient deployment.
22
+
23
+ ## Network information
24
+
25
+
26
+ | Network Information | Value |
27
+ |--------------------|-------|
28
+ | Framework | Torch |
29
+ | MParams | ~0.68–3.17 M |
30
+ | Quantization | Int8 |
31
+ | Provenance | https://github.com/amirgholami/SqueezeNext |
32
+ | Paper | https://arxiv.org/abs/1803.10615 |
33
+
34
+ ## Network inputs / outputs
35
+
36
+
37
+ For an image resolution of NxM and P classes
38
+
39
+ | Input Shape | Description |
40
+ | ----- | ----------- |
41
+ | (1, N, M, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
42
+
43
+ | Output Shape | Description |
44
+ | ----- | ----------- |
45
+ | (1, P) | Per-class confidence for P classes in FLOAT32|
46
+
47
+
48
+ ## Recommended platforms
49
+
50
+
51
+ | Platform | Supported | Recommended |
52
+ |----------|-----------|-----------|
53
+ | STM32L0 |[]|[]|
54
+ | STM32L4 |[]|[]|
55
+ | STM32U5 |[]|[]|
56
+ | STM32H7 |[]|[]|
57
+ | STM32MP1 |[]|[]|
58
+ | STM32MP2 |[]|[]|
59
+ | STM32N6 |[x]|[x]|
60
+
61
+ # Performances
62
+
63
+ ## Metrics
64
+
65
+ - Measures are done with default STEdgeAI Core configuration with enabled input / output allocated option.
66
+ - All the models are trained from scratch on Imagenet dataset
67
+
68
+ ### Reference **NPU** memory footprint on Imagenet dataset (see Accuracy for details on dataset)
69
+ | Model | Dataset | Format | Resolution | Series | Internal RAM (KiB) | External RAM (KiB) | Weights Flash (KiB) | STEdgeAI Core version |
70
+ |-------|---------|--------|------------|--------|--------------|--------------|---------------|----------------------|
71
+ | [sqnxt23_x100_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x100_pt_224/sqnxt23_x100_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 2086.45 | 3025 | 693.67 | 3.0.0 |
72
+ | [sqnxt23_x150_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x150_pt_224/sqnxt23_x150_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 2087.48 | 6806.25 | 1453.99 | 3.0.0 |
73
+ | [sqnxt23_x200_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x200_pt_224/sqnxt23_x200_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 2275.52 | 9075 | 2493.33 | 3.0.0 |
74
+ | [sqnxt23v5_x150_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x150_pt_224/sqnxt23v5_x150_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 2087.48 | 6806.25 | 1879.24 | 3.0.0 |
75
+ | [sqnxt23v5_x200_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x200_pt_224/sqnxt23v5_x200_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 2275.52 | 9075 | 3249.45 | 3.0.0 |
76
+
77
+
78
+ ### Reference **NPU** inference time on Imagenet dataset (see Accuracy for details on dataset)
79
+
80
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
81
+ |--------|---------|--------|--------|-------------|------------------|------------------|---------------------|-------------------------|
82
+ | [sqnxt23_x100_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x100_pt_224/sqnxt23_x100_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 87.07 | 11.49 | 3.0.0 |
83
+ | [sqnxt23_x150_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x150_pt_224/sqnxt23_x150_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 127.46 | 7.85 | 3.0.0 |
84
+ | [sqnxt23_x200_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x200_pt_224/sqnxt23_x200_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 182.12 | 5.49 | 3.0.0 |
85
+ | [sqnxt23v5_x100_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x100_pt_224/sqnxt23v5_x100_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 86.37 | 11.58 | 3.0.0 |
86
+ | [sqnxt23v5_x150_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x150_pt_224/sqnxt23v5_x150_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 126.91 | 7.88 | 3.0.0 |
87
+ | [sqnxt23v5_x200_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x200_pt_224/sqnxt23v5_x200_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 181.01 | 5.52 | 3.0.0 |
88
+
89
+
90
+ ### Accuracy with Imagenet dataset
91
+
92
+ | Model | Format | Resolution | Top 1 Accuracy |
93
+ | --- | --- | --- | --- |
94
+ | [sqnxt23_x100_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x100_pt_224/sqnxt23_x100_pt_224.onnx) | Float | 224x224x3 | 58.18 % |
95
+ | [sqnxt23_x100_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x100_pt_224/sqnxt23_x100_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 57.86 % |
96
+ | [sqnxt23_x150_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x150_pt_224/sqnxt23_x150_pt_224.onnx) | Float | 224x224x3 | 66.17 % |
97
+ | [sqnxt23_x150_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x150_pt_224/sqnxt23_x150_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 65.48 % |
98
+ | [sqnxt23_x200_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x200_pt_224/sqnxt23_x200_pt_224.onnx) | Float | 224x224x3 | 70.56 % |
99
+ | [sqnxt23_x200_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x200_pt_224/sqnxt23_x200_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 70.25 % |
100
+ | [sqnxt23v5_x100_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x100_pt_224/sqnxt23v5_x100_pt_224.onnx) | Float | 224x224x3 | 59.85 % |
101
+ | [sqnxt23v5_x100_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x100_pt_224/sqnxt23v5_x100_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 59.57 % |
102
+ | [sqnxt23v5_x150_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x150_pt_224/sqnxt23v5_x150_pt_224.onnx) | Float | 224x224x3 | 67.32 % |
103
+ | [sqnxt23v5_x150_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x150_pt_224/sqnxt23v5_x150_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 66.78 % |
104
+ | [sqnxt23v5_x200_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x200_pt_224/sqnxt23v5_x200_pt_224.onnx) | Float | 224x224x3 | 71.42 % |
105
+ | [sqnxt23v5_x200_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x200_pt_224/sqnxt23v5_x200_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 71.02 % |
106
+
107
+
108
+ | Model | Format | Resolution | Top 1 Accuracy |
109
+ | --- | --- | --- | --- |
110
+ | [sqnxt23_x100_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x100_pt_224/sqnxt23_x100_pt_224.onnx) | Float | 224x224x3 | 58.18 % |
111
+ | [sqnxt23_x100_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x100_pt_224/sqnxt23_x100_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 57.86 % |
112
+ | [sqnxt23_x150_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x150_pt_224/sqnxt23_x150_pt_224.onnx) | Float | 224x224x3 | 66.17 % |
113
+ | [sqnxt23_x150_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x150_pt_224/sqnxt23_x150_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 65.48 % |
114
+ | [sqnxt23_x200_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x200_pt_224/sqnxt23_x200_pt_224.onnx) | Float | 224x224x3 | 70.56 % |
115
+ | [sqnxt23_x200_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23_x200_pt_224/sqnxt23_x200_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 70.25 % |
116
+ | [sqnxt23v5_x100_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x100_pt_224/sqnxt23v5_x100_pt_224.onnx) | Float | 224x224x3 | 59.85 % |
117
+ | [sqnxt23v5_x100_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x100_pt_224/sqnxt23v5_x100_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 59.57 % |
118
+ | [sqnxt23v5_x150_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x150_pt_224/sqnxt23v5_x150_pt_224.onnx) | Float | 224x224x3 | 67.32 % |
119
+ | [sqnxt23v5_x150_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x150_pt_224/sqnxt23v5_x150_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 66.78 % |
120
+ | [sqnxt23v5_x200_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x200_pt_224/sqnxt23v5_x200_pt_224.onnx) | Float | 224x224x3 | 71.42 % |
121
+ | [sqnxt23v5_x200_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/sqnxt_pt/Public_pretrainedmodel_public_dataset/Imagenet/sqnxt23v5_x200_pt_224/sqnxt23v5_x200_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 71.02 % |
122
+
123
+
124
+
125
+ ## Retraining and Integration in a simple example:
126
+
127
+ Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
128
+
129
+
130
+
131
+ # References
132
+
133
+ <a id="1">[1]</a> - **Dataset**: Imagenet (ILSVRC 2012) — https://www.image-net.org/
134
+
135
+ <a id="2">[2]</a> - **Model**: SqueezeNext — https://github.com/amirgholami/SqueezeNext