Hugging Face's logo

STMicroelectronics
/
resnet

Model card Files
xet
Community
FBAGSTM commited on
Commit
2b65b4c
·
verified ·
1 Parent(s): 5048a89

Release AI-ModelZoo-4.0.0

Browse files
Files changed (1) hide show
  1. README.md +154 -6
README.md CHANGED
@@ -1,6 +1,154 @@
1
- ---
2
- license: other
3
- license_name: sla0044
4
- license_link: >-
5
- https://github.com/STMicroelectronics/stm32ai-modelzoo/blob/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/LICENSE.md
6
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: sla0044
4
+ license_link: >-
5
+ https://github.com/STMicroelectronics/stm32ai-modelzoo/blob/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/LICENSE.md
6
+ ---
7
+ # ResNet v1
8
+
9
+ ## **Use case** : `Image classification`
10
+
11
+ # Model description
12
+
13
+ ResNet models perform image classification - they take images as input and classify the major object in the image into a
14
+ set of pre-defined classes. ResNet models provide very high accuracies with affordable model sizes. They are ideal for cases when high accuracy of classification is required.
15
+ ResNet models consist of residual blocks and came up to counter the effect of deteriorating accuracies with more layers due to network not learning the initial layers.
16
+ ResNet v1 uses post-activation for the residual blocks. The models below have 8 and 32 layers with ResNet v1 architecture.
17
+ (source: https://keras.io/api/applications/resnet/)
18
+ The model is quantized in int8 using tensorflow lite converter.
19
+
20
+
21
+ ## Network information
22
+
23
+ | Network Information | Value |
24
+ |-------------------------|-------------------------------------------------------------------------|
25
+ | Framework | TensorFlow Lite |
26
+ | Quantization | int8 |
27
+ | Provenance | https://www.tensorflow.org/api_docs/python/tf/keras/applications/resnet |
28
+ | Paper | https://arxiv.org/abs/1512.03385 |
29
+
30
+ The models are quantized using tensorflow lite converter.
31
+
32
+ ## Network inputs / outputs
33
+
34
+ For an image resolution of NxM and P classes
35
+
36
+ | Input Shape | Description |
37
+ |----------------|-------------------------------------------------------------|
38
+ | (1, N, M, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
39
+
40
+ | Output Shape | Description |
41
+ |----------------|-------------------------------------------------------------|
42
+ | (1, P) | Per-class confidence for P classes in FLOAT32 |
43
+
44
+ ## Recommended Platforms
45
+
46
+ | Platform | Supported | Optimized |
47
+ |----------|-----------|-----------|
48
+ | STM32L0 | [] | [] |
49
+ | STM32L4 | [x] | [] |
50
+ | STM32U5 | [x] | [] |
51
+ | STM32H7 | [x] | [x] |
52
+ | STM32MP1 | [x] | [x]* |
53
+ | STM32MP2 | [x] | [] |
54
+ | STM32N6 | [x] | [] |
55
+
56
+ * Only for Cifar 100 models
57
+
58
+ # Performances
59
+
60
+ ## Metrics
61
+
62
+ - Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.
63
+ - `tfs` stands for "training from scratch", meaning that the model weights were randomly initialized before training.
64
+ - `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.
65
+ - `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.
66
+
67
+
68
+ ### Reference **MCU** memory footprint based on Cifar 10 dataset (see Accuracy for details on dataset)
69
+
70
+ | Model | Format | Resolution | Series | Activation RAM | Runtime RAM | Weights Flash | Code Flash | Total RAM | Total Flash | STEdgeAI Core version |
71
+ |----------|--------|-------------|---------|----------------|-------------|---------------|------------|-----------|-------------|------------------------|
72
+ | [ResNet v1 8 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar10/resnet8_32_tfs/resnet8_32_tfs_int8.tflite) | Int8 | 32x32x3 | STM32H7 | 62.51 KiB | 1.26 KiB | 76.9 KiB | 36.08 KiB | 63.77 KiB | 112.98 KiB | 3.0.0 |
73
+
74
+
75
+ ### Reference **MCU** inference time based on Cifar 10 dataset (see Accuracy for details on dataset)
76
+
77
+ | Model | Format | Resolution | Board | Execution Engine | Frequency | Inference time (ms) | STEdgeAI Core version |
78
+ |----------------------------------|--------|-------------|------------------|------------------|--------------|---------------------|------------------------|
79
+ | [ResNet v1 8 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar10/resnet8_32_tfs/resnet8_32_tfs_int8.tflite) | Int8 | 32x32x3 | STM32H747I-DISCO | 1 CPU | 400 MHz | 28.71 ms | 3.0.0 |
80
+
81
+
82
+ ### Reference **MPU** inference time based on Flowers dataset (see Accuracy for details on dataset)
83
+
84
+ | Model | Format | Resolution | Quantization | Board | Execution Engine | Frequency | Inference time (ms) | %NPU | %GPU | %CPU | X-LINUX-AI version | Framework |
85
+ |--------------------------------------------------------------------------------------------------------------------------|--------|------------|----------------|-----------------|------------------|-----------|---------------------|-------|-------|------|--------------------|-----------------------|
86
+ | [ResNet v1 8 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar10/resnet8_32_tfs/resnet8_32_tfs_int8.tflite) | Int8 | 32x32x3 | per-channel** | STM32MP257F-DK2 | NPU/GPU | 800 MHz | 2.06 | 21.76 | 78.24 | 0 | v6.1.0 | OpenVX |
87
+ | [ResNet v1 8 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar10/resnet8_32_tfs/resnet8_32_tfs_int8.tflite) | Int8 | 32x32x3 | per-channel | STM32MP157F-DK2 | 2 CPU | 800 MHz | 6.71 | NA | NA | 100 | v6.1.0 | TensorFlowLite 2.18.0 |
88
+ | [ResNet v1 8 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar10/resnet8_32_tfs/resnet8_32_tfs_int8.tflite) | Int8 | 32x32x3 | per-channel | STM32MP135F-DK2 | 1 CPU | 1000 MHz | 10.34 | NA | NA | 100 | v6.1.0 | TensorFlowLite 2.18.0 |
89
+
90
+
91
+ ** **To get the most out of MP25 NPU hardware acceleration, please use per-tensor quantization**
92
+
93
+ ** **Note:** On STM32MP2 devices, per-channel quantized models are internally converted to per-tensor quantization by the compiler using an entropy-based method. This may introduce a slight loss in accuracy compared to the original per-channel models.
94
+
95
+ ### Reference **MCU** memory footprint based on Cifar 100 dataset (see Accuracy for details on dataset)
96
+
97
+ | Model | Format | Resolution | Series | Activation RAM | Runtime RAM | Weights Flash | Code Flash | Total RAM | Total Flash | STEdgeAI Core version |
98
+ |-----------|--------|-------------|---------|----------------|-------------|---------------|------------|-------------|-------------|------------------------|
99
+ | [ResNet v1 32 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar100/resnet32_32_tfs/resnet32_32_tfs_int8.tflite) | Int8 | 32x32x3 | STM32H7 | 45.41 KiB | 24.98 KiB | 464.38 KiB | 78.65 KiB | 70.39 KiB | 543.03 KiB | 3.0.0 |
100
+
101
+
102
+ ### Reference **MCU** inference time based on Cifar 100 dataset (see Accuracy for details on dataset)
103
+
104
+ | Model | Format | Resolution | Board | Execution Engine | Frequency | Inference time (ms) | STEdgeAI Core version |
105
+ |---------|--------|------------|------------------|------------------|--------------|---------------------|------------------------|
106
+ | [ResNet v1 32 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar100/resnet32_32_tfs/resnet32_32_tfs_int8.tflite) | Int8 | 32x32x3 | STM32H747I-DISCO | 1 CPU | 400 MHz | 177.7 ms |3.0.0 |
107
+
108
+
109
+ ### Reference **MPU** inference time based on Flowers dataset (see Accuracy for details on dataset)
110
+ | Model | Format | Resolution | Quantization | Board | Execution Engine | Frequency | Inference time (ms) | %NPU | %GPU | %CPU | X-LINUX-AI version | Framework |
111
+ |---------------------------------------------------------------------------------------------------------------------|----------|------------|---------------|-------------------|------------------|-----------|---------------------|-------|-------|------|--------------------|-----------------------|
112
+ |[ResNet v1 32 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar100/resnet32_32_tfs/resnet32_32_tfs_int8.tflite) | Int8 | 32x32x3 | per-channel | STM32MP257F-DK2 | NPU/GPU | 800 MHz | 9.160 ms | 14.75 | 85.25 | 0 | v6.1.0 | OpenVX |
113
+ |[ResNet v1 32 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar100/resnet32_32_tfs/resnet32_32_tfs_int8.tflite) | Int8 | 32x32x3 | per-channel | STM32MP157F-DK2 | 2 CPU | 800 MHz | 34.78 ms | NA | NA | 100 | v6.1.0 | TensorFlowLite 2.11.0 |
114
+ |[ResNet v1 32 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar100/resnet32_32_tfs/resnet32_32_tfs_int8.tflite) | Int8 | 32x32x3 | per-channel | STM32MP135F-DK2 | 1 CPU | 1000 MHz | 55.32 ms | NA | NA | 100 | v6.1.0 | TensorFlowLite 2.11.0 |
115
+
116
+
117
+ ### Accuracy with Cifar10 dataset
118
+
119
+ Dataset details: [link](https://www.cs.toronto.edu/~kriz/cifar.html) ,
120
+ License [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) , Quotation[[1]](#1) , Number of classes: 10, Number of
121
+ images: 60 000
122
+
123
+ | Model | Format | Resolution | Top 1 Accuracy |
124
+ |------------------------------------------------------------------------------------------------------------------|----------|-------------|----------------|
125
+ | [ResNet v1 8 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar10/resnet8_32_tfs/resnet8_32_tfs.keras) | Float | 32x32x3 | 87.01 % |
126
+ | [ResNet v1 8 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar10/resnet8_32_tfs/resnet8_32_tfs_int8.tflite) | Int8 | 32x32x3 | 85.59 % |
127
+
128
+
129
+ ### Accuracy with Cifar100 dataset
130
+
131
+ Dataset details: [link](https://www.cs.toronto.edu/~kriz/cifar.html) ,
132
+ License [CC0 4.0](https://creativecommons.org/licenses/by/4.0/), Quotation[[2]](#2) , Number of classes:100,
133
+ Number of images: 600 000
134
+
135
+ | Model | Format | Resolution | Top 1 Accuracy |
136
+ |----------------------------------------------------------------------------------------------------------------------|---------|------------|----------------|
137
+ | [ResNet v1 32 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar100/resnet32_32_tfs/resnet32_32_tfs.keras) | Float | 32x32x3 | 67.75 % |
138
+ | [ResNet v1 32 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet/ST_pretrainedmodel_public_dataset/cifar100/resnet32_32_tfs/resnet32_32_tfs_int8.tflite) | Int8 | 32x32x3 | 66.58 % |
139
+
140
+ ## Retraining and Integration in a simple example:
141
+
142
+ Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
143
+
144
+
145
+ # References
146
+
147
+ <a id="1">[1]</a>
148
+ "Tf_flowers : tensorflow datasets," TensorFlow. [Online]. Available: https://www.tensorflow.org/datasets/catalog/tf_flowers.
149
+
150
+ <a id="2">[2]</a>
151
+ J, ARUN PANDIAN; GOPAL, GEETHARAMANI (2019), "Data for: Identification of Plant Leaf Diseases Using a 9-layer Deep Convolutional Neural Network", Mendeley Data, V1, doi: 10.17632/tywbtsjrjv.1
152
+
153
+ <a id="3">[3]</a>
154
+ L. Bossard, M. Guillaumin, and L. Van Gool, "Food-101 -- Mining Discriminative Components with Random Forests." European Conference on Computer Vision, 2014.