Hugging Face's logo

STMicroelectronics
/
resnet50v2

Image Classification
Model card Files
xet
Community
FBAGSTM commited on
Commit
51e373d
·
verified ·
1 Parent(s): ac55db6

Release AI-ModelZoo-4.0.0

Browse files
Files changed (1) hide show
  1. README.md +32 -28
README.md CHANGED
@@ -14,7 +14,7 @@ pipeline_tag: image-classification
14
 
15
  ResNets family is a well known architecture that uses skip connections to enable stronger gradients in much deeper networks. This variant has 50 layers.
16
 
17
- The model is quantized in int8 using tensorflow lite converter.
18
 
19
  ## Network information
20
 
@@ -67,34 +67,38 @@ For an image resolution of NxM and P classes
67
  - `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.
68
 
69
 
70
- ### Reference **NPU** memory footprint on food-101 and ImageNet dataset (see Accuracy for details on dataset)
71
- |Model | Dataset | Format | Resolution | Series | Internal RAM | External RAM | Weights Flash | STM32Cube.AI version | STEdgeAI Core version |
72
- |----------|------------------|--------|-------------|------------------|--------------|--------------|---------------|----------------------|-------------------------|
73
- | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food-101/resnet50_v2_224_fft/resnet50_v2_224_fft_int8.tflite) | food-101 | Int8 | 224x224x3 | STM32N6 | 2308.06 | 3136 | 23833.61 | 10.2.0 | 2.2.0 |
74
- | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/ImageNet/resnet50_v2_224/resnet50_v2_224_int8.tflite) | ImageNet | Int8 | 224x224x3 | STM32N6 | 2308.06 | 3136.0 | 25633.55 | 10.2.0 | 2.2.0 |
 
 
75
 
76
- ### Reference **NPU** inference time on food-101 and ImageNet dataset (see Accuracy for details on dataset)
77
- | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STM32Cube.AI version | STEdgeAI Core version |
78
- |--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------|----------------------|-------------------------|
79
- | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food-101/resnet50_v2_224_fft/resnet50_v2_224_fft_int8.tflite) | food-101 | Int8 | 224x224x3 | STM32N6570-DK | NPU/MCU | 226.16 | 4.42 | 10.2.0 | 2.2.0 |
80
- | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/ImageNet/resnet50_v2_224/resnet50_v2_224_int8.tflite) | ImageNet | Int8 | 224x224x3 | STM32N6570-DK | NPU/MCU | 231.59 | 4.31 | 10.2.0 | 2.2.0 |
 
 
81
 
82
 
83
- ### Reference **MCU** memory footprint based on Food-101 and ImageNet dataset (see Accuracy for details on dataset)
84
 
85
- | Model | Format | Resolution | Series | Activation RAM | Runtime RAM | Weights Flash | Code Flash | Total RAM | Total Flash | STM32Cube.AI version |
86
- |--------------------------------------------------------------------------------------------------------------------------------------|--------|------------|---------|----------------|-------------|---------------|------------|-------------|-----------------|-----------------------|
87
- | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food-101/resnet50_v2_224_fft/resnet50_v2_224_fft_int8.tflite) | Int8 | 224x224x3 | STM32H7 | 2142.07 KiB | 41.03 KiB | 23240.96 KiB | 225.32 KiB | 2183.1 KiB | 23466.28 KiB | 10.2.0 |
88
- | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/ImageNet/resnet50_v2_224/resnet50_v2_224_int8.tflite) | Int8 | 224x224x3 | STM32H7 | 2142.07 KiB | 41.03 KiB | 25042.47 KiB | 225.32 KiB | 2183.1 KiB | 25267.79 KiB | 10.2.0 |
89
 
90
 
91
- ### Reference **MCU** inference time based on Food-101 and ImageNet dataset (see Accuracy for details on dataset)
92
 
93
 
94
- | Model | Format | Resolution | Board | Execution Engine | Frequency | Inference time (ms) | STM32Cube.AI version |
95
- |-------------------|--------|------------|------------------|------------------|-----------|---------------------|-----------------------|
96
- | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food-101/resnet50_v2_224_fft/resnet50_v2_224_fft_int8.tflite) | Int8 | 224x224x3 | STM32H747I-DISCO | 1 CPU | 400 MHz | 11360.76 ms | 10.2.0 |
97
- | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/ImageNet/resnet50_v2_224/resnet50_v2_224_int8.tflite) | Int8 | 224x224x3 | STM32H747I-DISCO | 1 CPU | 400 MHz | 11370.07 | 10.2.0 |
98
 
99
 
100
 
@@ -104,11 +108,11 @@ Dataset details: [link](https://data.vision.ee.ethz.ch/cvl/datasets_extra/food-1
104
 
105
  | Model | Format | Resolution | Top 1 Accuracy |
106
  |-------|--------|------------|----------------|
107
- | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food-101/resnet50_v2_224_fft/resnet50_v2_224_fft.h5) | Float | 224x224x3 | 71.53 % |
108
- | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food-101/resnet50_v2_224_fft/resnet50_v2_224_fft_int8.tflite) | Int8 | 224x224x3 | 70.07 % |
 
109
 
110
-
111
- ### Accuracy with ImageNet dataset
112
 
113
  Dataset details: [link](https://www.image-net.org), Quotation[[4]](#4).
114
  Number of classes: 1000.
@@ -117,9 +121,9 @@ For the sake of simplicity, the accuracy reported here was estimated on the 5000
117
 
118
  |model | Format | Resolution | Top 1 Accuracy |
119
  |---------|--------|------------|----------------|
120
- | [ResNet50 v2 ](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/ImageNet/resnet50_v2_224/resnet50_v2_224.h5) | Float | 224x224x3 | 66.38 % |
121
- | [ResNet50 v2 ](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/ImageNet/resnet50_v2_224/resnet50_v2_224_int8.tflite) | Int8 | 224x224x3 | 65.99 % |
122
-
123
 
124
 
125
  ## Retraining and Integration in a simple example:
 
14
 
15
  ResNets family is a well known architecture that uses skip connections to enable stronger gradients in much deeper networks. This variant has 50 layers.
16
 
17
+ The model is quantized in int8 using tensorflow lite converter. A mixed precision version is also provided using onnx-runtime and our own quantization scripts.
18
 
19
  ## Network information
20
 
 
67
  - `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.
68
 
69
 
70
+ ### Reference **NPU** memory footprint on food101 and imagenet dataset (see Accuracy for details on dataset)
71
+ |Model | Dataset | Format | Resolution | Series | Internal RAM | External RAM | Weights Flash | STEdgeAI Core version |
72
+ |----------|------------------|--------|-------------|------------------|--------------|--------------|----------------------|-------------------------|
73
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft_int8.tflite) | food101 | Int8 | 224x224x3 | STM32N6 | 2308.06 | 3136 | 23833.67 | 3.0.0 |
74
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft_qdq_w4_91.4%_w8_8.6%_a8_100%_acc_80.17.onnx) | food101 | Int8/Int4 | 224x224x3 | STM32N6 | 2308.06 | 2352 | 13268.39 | 3.0.0 |
75
+ | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224_int8.tflite) | imagenet | Int8 | 224x224x3 | STM32N6 | 2308.06 | 3136.0 | 25633.61 | 3.0.0 |
76
+ | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224_qdq_w4_35.98%_w8_64.02%_a8_100%_acc_67.45.onnx) | imagenet | Int8/Int4 | 224x224x3 | STM32N6 | 2308.06 | 2352 | 21154.53 | 3.0.0 |
77
 
78
+ ### Reference **NPU** inference time on food101 and imagenet dataset (see Accuracy for details on dataset)
79
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
80
+ |--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------|-------------------------|
81
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft_int8.tflite) | food101 | Int8 | 224x224x3 | STM32N6570-DK | NPU/MCU | 238.49 | 4.19 | 3.0.0 |
82
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft_qdq_w4_91.4%_w8_8.6%_a8_100%_acc_80.17.onnx) | food101 | Int8/Int4 | 224x224x3 | STM32N6570-DK | NPU/MCU | 267.33 | 3.74 | 3.0.0 |
83
+ | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224_int8.tflite) | imagenet | Int8 | 224x224x3 | STM32N6570-DK | NPU/MCU | 243.04 | 4.11 | 3.0.0 |
84
+ | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224_qdq_w4_35.98%_w8_64.02%_a8_100%_acc_67.45.onnx) | imagenet | Int8/Int4 | 224x224x3 | STM32N6570-DK | NPU/MCU | 286.06 | 3.5 | 3.0.0 |
85
 
86
 
87
+ ### Reference **MCU** memory footprint based on Food-101 and imagenet dataset (see Accuracy for details on dataset)
88
 
89
+ | Model | Dataset | Format | Resolution | Series | Activation RAM | Runtime RAM | Weights Flash | Code Flash | Total RAM | Total Flash | STEdgeAI Core version |
90
+ |-----------|---------------------------------------------------------------------------------------------------------------------------|--------|------------|---------|----------------|-------------|---------------|------------|-------------|-----------------|-----------------------|
91
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft_int8.tflite) | food101 | Int8 | 224x224x3 | STM32H7 | 1816.2 KiB | 14.56 KiB | 23240.96 KiB | 169.12 KiB | 1830.76 KiB | 23410.08 KiB | 3.0.0 |
92
+ | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224_int8.tflite) | imagenet | Int8 | 224x224x3 | STM32H7 | 2142.07 KiB | 41.03 KiB | 25042.47 KiB | 225.32 KiB | 2183.1 KiB | 25267.79 KiB | 3.0.0 |
93
 
94
 
95
+ ### Reference **MCU** inference time based on Food-101 and imagenet dataset (see Accuracy for details on dataset)
96
 
97
 
98
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Frequency | Inference time (ms) | STEdgeAI Core version |
99
+ |-------------------|-----|---|------------|------------------|------------------|-----------|---------------------|-----------------------|
100
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft_int8.tflite) | food101 | Int8 | 224x224x3 | STM32H747I-DISCO | 1 CPU | 400 MHz | 11314.82 | 3.0.0 |
101
+ | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224_int8.tflite) | imagenet | Int8 | 224x224x3 | STM32H747I-DISCO | 1 CPU | 400 MHz | 11370.07 | 3.0.0 |
102
 
103
 
104
 
 
108
 
109
  | Model | Format | Resolution | Top 1 Accuracy |
110
  |-------|--------|------------|----------------|
111
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft.keras) | Float | 224x224x3 | 82.2 % |
112
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft_int8.tflite) | Int8 | 224x224x3 | 81.03 % |
113
+ | [ResNet50 v2 fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/ST_pretrainedmodel_public_dataset/food101/resnet50v2_224_fft/resnet50v2_224_fft_qdq_w4_91.4%_w8_8.6%_a8_100%_acc_80.17.onnx) | Int8/Int4 | 224x224x3 | 80.17 % |
114
 
115
+ ### Accuracy with imagenet dataset
 
116
 
117
  Dataset details: [link](https://www.image-net.org), Quotation[[4]](#4).
118
  Number of classes: 1000.
 
121
 
122
  |model | Format | Resolution | Top 1 Accuracy |
123
  |---------|--------|------------|----------------|
124
+ | [ResNet50 v2 ](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224.keras) | Float | 224x224x3 | 68.73 % |
125
+ | [ResNet50 v2 ](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224_int8.tflite) | Int8 | 224x224x3 | 67.99 % |
126
+ | [ResNet50 v2](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/resnet50v2/Public_pretrainedmodel_public_dataset/imagenet/resnet50v2_224/resnet50v2_224_qdq_w4_35.98%_w8_64.02%_a8_100%_acc_67.45.onnx) | Int8/Int4 | 224x224x3 | 67.45 % |
127
 
128
 
129
  ## Retraining and Integration in a simple example: