Hugging Face's logo

STMicroelectronics
/
mobilenetv2_reid

Model card Files
xet
Community
FBAGSTM commited on
Commit
de6f06b
·
verified ·
1 Parent(s): 4b1b5cc

Release AI-ModelZoo-4.0.0

Browse files
Files changed (1) hide show
  1. README.md +129 -6
README.md CHANGED
@@ -1,6 +1,129 @@
1
- ---
2
- license: other
3
- license_name: sla0044
4
- license_link: >-
5
- https://github.com/STMicroelectronics/stm32ai-modelzoo/blob/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/LICENSE.md
6
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: sla0044
4
+ license_link: >-
5
+ https://github.com/STMicroelectronics/stm32ai-modelzoo/blob/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/LICENSE.md
6
+ ---
7
+ # MobileNet v2
8
+
9
+ ## **Use case** : `Re-Identification`
10
+
11
+ # Model description
12
+
13
+
14
+ MobileNet v2 is very similar to the original MobileNet, except that it uses inverted residual blocks with bottlenecking features.
15
+
16
+ It has a drastically lower parameter count than the original MobileNet.
17
+
18
+ MobileNet models support any input size greater than 32 x 32, with larger image sizes offering better performance.
19
+ Alpha parameter: float, larger than zero, controls the width of the network. This is known as the width multiplier in the MobileNetV2 paper, but the name is kept for consistency with applications.
20
+
21
+ If alpha < 1.0, proportionally decreases the number of filters in each layer.
22
+
23
+ If alpha > 1.0, proportionally increases the number of filters in each layer.
24
+
25
+ If alpha = 1.0, default number of filters from the paper are used at each layer.
26
+
27
+ (source: https://keras.io/api/applications/mobilenet/)
28
+
29
+ The model is quantized in int8 using tensorflow lite converter.
30
+
31
+ ## Network information
32
+
33
+
34
+ | Network Information | Value |
35
+ |-------------------------|-----------------|
36
+ | Framework | TensorFlow Lite |
37
+ | MParams alpha=0.35 | 1.66 M |
38
+ | Quantization | int8 |
39
+ | Provenance | https://www.tensorflow.org/api_docs/python/tf/keras/applications/mobilenet_v2 |
40
+ | Paper | https://arxiv.org/pdf/1801.04381.pdf |
41
+
42
+ The models are quantized using tensorflow lite converter.
43
+
44
+
45
+ ## Network inputs / outputs
46
+
47
+
48
+ For an image resolution of NxM and P classes
49
+
50
+ | Input Shape | Description |
51
+ | ----- | ----------- |
52
+ | (1, N, M, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
53
+
54
+ | Output Shape | Description |
55
+ | ----- | ----------- |
56
+ | (1, P) | Per-class confidence for P classes in FLOAT32|
57
+
58
+
59
+ ## Recommended platforms
60
+
61
+
62
+ | Platform | Supported | Recommended |
63
+ |----------|-----------|-----------|
64
+ | STM32L0 |[]|[]|
65
+ | STM32L4 |[x]|[]|
66
+ | STM32U5 |[x]|[]|
67
+ | STM32H7 |[x]|[x]|
68
+ | STM32MP1 |[x]|[x]|
69
+ | STM32MP2 |[x]|[x]|
70
+ | STM32N6 |[x]|[x]|
71
+
72
+ # Performances
73
+
74
+ ## Metricss
75
+
76
+ - Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.
77
+ - `tfs` stands for "training from scratch", meaning that the model weights were randomly initialized before training.
78
+ - `tl` stands for "transfer learning", meaning that the model backbone weights were initialized from a pre-trained model, then only the last layer was unfrozen during the training.
79
+ - `fft` stands for "full fine-tuning", meaning that the full model weights were initialized from a transfer learning pre-trained model, and all the layers were unfrozen during the training.
80
+
81
+
82
+ ### Reference **NPU** memory footprint on DeepSportradar dataset (see Accuracy for details on dataset)
83
+ |Model | Dataset | Format | Resolution | Series | Internal RAM | External RAM | Weights Flash | STEdgeAI Core version |
84
+ |----------|------------------|--------|-------------|------------------|------------------|---------------------|---------------|-------------------------|
85
+ | [MobileNet v2 0.35 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a035_256_128_fft/mobilenetv2_a035_256_128_fft_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32N6 | 480 | 0.0 | 553.58 | 3.0.0 |
86
+ | [MobileNet v2 1.0 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a100_256_128_fft/mobilenetv2_a100_256_128_fft_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32N6 | 1440 | 0.0 | 2786.52 | 3.0.0 |
87
+
88
+
89
+ ### Reference **NPU** inference time on DeepSportradar dataset (see Accuracy for details on dataset)
90
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
91
+ |--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------| -----------------------|
92
+ | [MobileNet v2 0.35 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a035_256_128_fft/mobilenetv2_a035_256_128_fft_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32N6570-DK | NPU/MCU | 4.15 | 241 | 3.0.0 |
93
+ | [MobileNet v2 1.0 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a100_256_128_fft/mobilenetv2_a100_256_128_fft_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32N6570-DK | NPU/MCU | 13.37 | 74.8 | 3.0.0 |
94
+
95
+
96
+ ### Reference **MCU** memory footprint based on DeepSportradar dataset (see Accuracy for details on dataset)
97
+ |Model | Dataset | Format | Resolution | Series | Internal RAM | External RAM | Weights Flash | STEdgeAI Core version |
98
+ |----------|------------------|--------|-------------|------------------|------------------|---------------------|---------------|-------------------------|
99
+ | [MobileNet v2 0.35 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a035_256_128_fft/mobilenetv2_a035_256_128_fft_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32H7 | 461.32 | 0.0 | 400.59 | 3.0.0 |
100
+ | [MobileNet v2 1.0 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a100_256_128_fft/mobilenetv2_a100_256_128_fft_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32H7 | 398.25 | 804.38 | 2205.06 | 3.0.0 |
101
+
102
+ ### Reference **MCU** inference time on DeepSportradar dataset (see Accuracy for details on dataset)
103
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
104
+ |--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------| -----------------------|
105
+ | [MobileNet v2 0.35 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a035_256_128_fft/mobilenetv2_a035_256_128_fft_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32H747I-DISCO | 1 CPU | 190.3 | 5.3 | 3.0.0 |
106
+ | [MobileNet v2 1.0 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a100_256_128_fft/mobilenetv2_a100_256_128_fft_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32H747I-DISCO | 1 CPU | 729.45 | 1.37 | 3.0.0 |
107
+
108
+
109
+
110
+ ### Performance with DeepSportradar ReID dataset
111
+
112
+
113
+ Dataset details: [link](https://github.com/DeepSportradar/player-reidentification-challenge) , License [Apache-2.0](https://github.com/DeepSportradar/player-reidentification-challenge?tab=Apache-2.0-1-ov-file#readme) , Number of identities: 486 (train: 436, test: 50), Number of images: 9529 (train: 8569, test_query: 50, test_gallery: 910)
114
+
115
+ | Model | Format | Resolution | mAP | rank-1 accuracy |rank-5 accuracy |rank-10 accuracy |
116
+ |-------|--------|------------|----------------|-----------------|----------------|-----------------|
117
+ | [MobileNet v2 0.35 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a035_256_128_fft/mobilenetv2_a035_256_128_fft_int8.tflite) | Int8 | 256x128 | 73.43 % | 96.0 % | 96.0 % | 98.0 % |
118
+ | [MobileNet v2 1.0 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/mobilenetv2/ST_pretrainedmodel_public_dataset/DeepSportradar/mobilenetv2_a100_256_128_fft/mobilenetv2_a100_256_128_fft_int8.tflite) | Int8 | 256x128 | 72.5 % | 94.0 % | 98.0 % | 98.0 % |
119
+
120
+
121
+ ## Retraining and Integration in a simple example:
122
+
123
+ Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
124
+
125
+
126
+ # References
127
+
128
+ <a id="1">[1]</a>
129
+ The DeepSportradar Player Re-Identification Challenge (2023) [Online]. Available: https://github.com/DeepSportradar/player-reidentification-challenge.