Hugging Face's logo

STMicroelectronics
/
efficientnetv2

Image Classification
Model card Files
xet
Community
FBAGSTM commited on
Commit
183c4b9
·
verified ·
1 Parent(s): 85a1469

Release AI-ModelZoo-4.0.0

Browse files
Files changed (1) hide show
  1. README.md +80 -49
README.md CHANGED
@@ -12,13 +12,12 @@ pipeline_tag: image-classification
12
  # Model description
13
 
14
 
15
- EfficientNet v2 family is one of the best topologies for image classification. It has been obtained through neural architecture search with a special care given to training time
16
- and number of parameters reduction.
17
 
18
  This family of networks comprises various subtypes: B0 (224x224), B1 (240x240), B2 (260x260), B3 (300x300), S (384x384) ranked by depth and width increasing order.
19
  There are also M, L, XL variants but too large to be executed efficiently on STM32N6.
20
 
21
- All these networks are already available on https://www.tensorflow.org/api_docs/python/tf/keras/applications/ pre-trained on ImageNet.
22
 
23
 
24
  ## Network information
@@ -72,49 +71,73 @@ For an image resolution of NxM and P classes
72
  * Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.
73
  * `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.
74
 
75
- ### Reference **NPU** memory footprint on food-101 and ImageNet dataset (see Accuracy for details on dataset)
76
- |Model | Dataset | Format | Resolution | Series | Internal RAM (KiB) | External RAM (KiB) | Weights Flash (KiB) | STM32Cube.AI version | STEdgeAI Core version |
77
- |----------|------------------|--------|-------------|------------------|------------------|---------------------|---------------------|----------------------|-------------------------|
78
- | [efficientnet_v2B0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B0_224_fft/efficientnet_v2B0_224_fft_qdq_int8.onnx) | food-101 | Int8 | 224x224x3 | STM32N6 | 1834.44 |0.0| 7552.02 | 10.2.0 | 2.2.0 |
79
- | [efficientnet_v2B1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B1_240_fft/efficientnet_v2B1_240_fft_qdq_int8.onnx) | food-101 | Int8 | 240x240x3 | STM32N6 | 2589.97 |0.0| 8332.27 | 10.2.0 | 2.2.0 |
80
- | [efficientnet_v2B2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B2_260_fft/efficientnet_v2B2_260_fft_qdq_int8.onnx) | food-101 | Int8 | 260x260x3 | STM32N6 | 2629.56 |528.12| 10525.95 | 10.2.0 | 2.2.0 |
81
- | [efficientnet_v2S_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2S_384_fft/efficientnet_v2S_384_fft_qdq_int8.onnx) | food-101 | Int8 | 384x384x3 | STM32N6 | 2700 | 6912 | 24451.31 | 10.2.0 | 2.2.0 |
82
- | [efficientnet_v2B0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B0_224/efficientnet_v2B0_224_qdq_int8.onnx) | ImageNet | Int8 | 224x224x3 | STM32N6 | 1834.44 | 0.0 | 8179.67 | 10.2.0 | 2.2.0 |
83
- | [efficientnet_v2B1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B1_240/efficientnet_v2B1_240_qdq_int8.onnx) | ImageNet | Int8 | 240x240x3 | STM32N6 | 2589.97 | 0.0 | 9459.92 | 10.2.0 | 2.2.0 |
84
- | [efficientnet_v2B2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B2_260/efficientnet_v2B2_260_qdq_int8.onnx) | ImageNet | Int8 | 260x260x3 | STM32N6 | 2629.56 | 528.12 | 11765.99 | 10.2.0 | 2.2.0 |
85
- | [efficientnet_v2S_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2S_384/efficientnet_v2S_384_qdq_int8.onnx) | ImageNet | Int8 | 384x384x3 | STM32N6 | 2700 | 6912 | 25579.03 | 10.2.0 | 2.2.0 |
86
-
87
-
88
- ### Reference **NPU** inference time on food-101 and ImageNet dataset (see Accuracy for details on dataset)
89
- | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STM32Cube.AI version | STEdgeAI Core version |
90
- |--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------|----------------------|-------------------------|
91
- | [efficientnet_v2B0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B0_224_fft/efficientnet_v2B0_224_fft_qdq_int8.onnx) | food-101 | Int8 | 224x224x3 | STM32N6570-DK | NPU/MCU | 52.05 | 19.21 | 10.2.0 | 2.2.0 |
92
- | [efficientnet_v2B1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B1_240_fft/efficientnet_v2B1_240_fft_qdq_int8.onnx) | food-101 | Int8 | 240x240x3 | STM32N6570-DK | NPU/MCU | 70.91 | 14.1 | 10.2.0 | 2.2.0 |
93
- | [efficientnet_v2B2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B2_260_fft/efficientnet_v2B2_260_fft_qdq_int8.onnx) | food-101 | Int8 | 260x260x3 | STM32N6570-DK | NPU/MCU | 142.62 | 7.01 | 10.2.0 | 2.2.0 |
94
- | [efficientnet_v2S_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2S_384_fft/efficientnet_v2S_384_fft_qdq_int8.onnx) | food-101 | Int8 | 384x384x3 | STM32N6570-DK | NPU/MCU | 816.34 | 1.22 | 10.2.0 | 2.2.0 |
95
- | [efficientnet_v2B0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B0_224/efficientnet_v2B0_224_qdq_int8.onnx) | ImageNet | Int8 | 224x224x3 | STM32N6570-DK | NPU/MCU | 55.27 | 18.09 | 10.2.0 | 2.2.0 |
96
- | [efficientnet_v2B1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B1_240/efficientnet_v2B1_240_qdq_int8.onnx) | ImageNet | Int8 | 240x240x3 | STM32N6570-DK | NPU/MCU | 74.48 | 13.34 | 10.2.0 | 2.2.0 |
97
- | [efficientnet_v2B2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B2_260/efficientnet_v2B2_260_qdq_int8.onnx) | ImageNet | Int8 | 260x260x3 | STM32N6570-DK | NPU/MCU | 145.27 | 6.88 | 10.2.0 | 2.2.0 |
98
- | [efficientnet_v2S_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2S_384/efficientnet_v2S_384_qdq_int8.onnx) | ImageNet | Int8 | 384x384x3 | STM32N6570-DK | NPU/MCU | 785.01 | 1.27 | 10.2.0 | 2.2.0 |
99
-
100
- * The deployment of all the models listed in the table is supported, except for the efficientnet_v2S_384 model, for which support is coming soon.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
101
  ### Accuracy with Food-101 dataset
102
 
103
  Dataset details: [link](https://data.vision.ee.ethz.ch/cvl/datasets_extra/food-101/), Quotation[[3]](#3) , Number of classes: 101 , Number of images: 101 000
104
 
105
  | Model | Format | Resolution | Top 1 Accuracy |
106
  |--------------------------------------------------------------------------------------------------------------------------------------------------|--------|-----------|----------------|
107
- | [efficientnet_v2B0_224_fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B0_224_fft/efficientnet_v2B0_224_fft.h5) | Float | 224x224x3 | 81.35 % |
108
- | [efficientnet_v2B0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B0_224_fft/efficientnet_v2B0_224_fft_qdq_int8.onnx) | Int8 | 224x224x3 | 81.1 % |
109
- | [efficientnet_v2B1_240_fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B1_240_fft/efficientnet_v2B1_240_fft.h5) | Float | 240x240x3 | 83.23 % |
110
- | [efficientnet_v2B1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B1_240_fft/efficientnet_v2B1_240_fft_qdq_int8.onnx) | Int8 | 240x240x3 | 82.95 % |
111
- | [efficientnet_v2B2_260_fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B2_260_fft/efficientnet_v2B2_260_fft.h5) | Float | 260x260x3 | 84.35 % |
112
- | [efficientnet_v2B2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2B2_260_fft/efficientnet_v2B2_260_fft_qdq_int8.onnx) | Int8 | 260x260x3 | 84.04 % |
113
- | [efficientnet_v2S_384_fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2S_384_fft/efficientnet_v2S_384_fft.h5) | Float | 384x384x3 | 88.16 % |
114
- | [efficientnet_v2S_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food-101/efficientnet_v2S_384_fft/efficientnet_v2S_384_fft_qdq_int8.onnx) | Int8 | 384x384x3 | 87.34 % |
115
-
116
-
117
- ### Accuracy with ImageNet
 
 
 
 
118
 
119
  Dataset details: [link](https://www.image-net.org), Quotation[[4]](#4).
120
  Number of classes: 1000.
@@ -123,14 +146,22 @@ For the sake of simplicity, the accuracy reported here was estimated on the 1000
123
 
124
  | Model | Format | Resolution | Top 1 Accuracy |
125
  |------------------------------------------------------------------------------------------------------------------------------------------|--------|------------|----------------|
126
- | [efficientnet_v2B0_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B0_224/efficientnet_v2B0_224.h5) | Float | 224x224x3 | 73.94 % |
127
- | [efficientnet_v2B0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B0_224/efficientnet_v2B0_224_qdq_int8.onnx) | Int8 | 224x224x3 | 72.21 % |
128
- | [efficientnet_v2B1_240](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B1_240/efficientnet_v2B1_240.h5) | Float | 240x240x3 | 76.14 % |
129
- | [efficientnet_v2B1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B1_240/efficientnet_v2B1_240_qdq_int8.onnx) | Int8 | 240x240x3 | 75.5 % |
130
- | [efficientnet_v2B2_260](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B2_260/efficientnet_v2B2_260.h5) | Float | 260x260x3 | 76.58 % |
131
- | [efficientnet_v2B2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2B2_260/efficientnet_v2B2_260_qdq_int8.onnx) | Int8 | 260x260x3 | 76.26 % |
132
- | [efficientnet_v2S_384](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2S_384/efficientnet_v2S_384.h5) | Float | 384x384x3 | 83.52 % |
133
- | [efficientnet_v2S_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/efficientnet_v2S_384/efficientnet_v2S_384_qdq_int8.onnx) | Int8 | 384x384x3 | 83.07 % |
 
 
 
 
 
 
 
 
134
 
135
 
136
  ## Retraining and Integration in a simple example:
@@ -151,4 +182,4 @@ L. Bossard, M. Guillaumin, and L. Van Gool, "Food-101 -- Mining Discriminative C
151
 
152
  <a id="4">[4]</a>
153
  Olga Russakovsky*, Jia Deng*, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg and Li Fei-Fei.
154
- (* = equal contribution) ImageNet Large Scale Visual Recognition Challenge.
 
12
  # Model description
13
 
14
 
15
+ EfficientNet v2 family is one of the best topologies for image classification. It has been obtained through neural architecture search with a special care given to training time and number of parameters reduction.
 
16
 
17
  This family of networks comprises various subtypes: B0 (224x224), B1 (240x240), B2 (260x260), B3 (300x300), S (384x384) ranked by depth and width increasing order.
18
  There are also M, L, XL variants but too large to be executed efficiently on STM32N6.
19
 
20
+ All these networks are already available on https://www.tensorflow.org/api_docs/python/tf/keras/applications/ pre-trained on imagenet.
21
 
22
 
23
  ## Network information
 
71
  * Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.
72
  * `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.
73
 
74
+ ### Reference **NPU** memory footprint on food101 and imagenet dataset (see Accuracy for details on dataset)
75
+ |Model | Dataset | Format | Resolution | Series | Internal RAM (KiB) | External RAM (KiB) | Weights Flash (KiB) | STEdgeAI Core version |
76
+ |-----------|---------------|----------|------------|-----------|--------------------|--------------------|---------------------|-----------------------|
77
+ | [efficientnetv2b0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b0_224_fft/efficientnetv2b0_224_fft_qdq_int8.onnx) | food101 | Int8 | 224x224x3 | STM32N6 | 1911.56 |0.0| 6839.39 | 3.0.0 |
78
+ | [efficientnetv2b0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b0_224_fft/efficientnetv2b0_224_fft_qdq_w4_90.1%_w8_9.9%_a8_100%_acc_84.47.onnx) | food101 | Int8/Int4 | 224x224x3 | STM32N6 | 1911.56 |0.0| 4237.52 | 3.0.0 |
79
+ | [efficientnetv2b1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b1_240_fft/efficientnetv2b1_240_fft_qdq_int8.onnx) | food101 | Int8 | 240x240x3 | STM32N6 | 2604.03 |0.0| 8089.27 | 3.0.0 |
80
+ | [efficientnetv2b1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b1_240_fft/efficientnetv2b1_240_fft_qdq_w4_91.8%_w8_8.2%_a8_100%_acc_85.71.onnx) | food101 | Int8/Int4 | 240x240x3 | STM32N6 | 2604.03 |0.0| 4995.39 | 3.0.0 |
81
+ | [efficientnetv2b2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b2_260_fft/efficientnetv2b2_260_fft_qdq_int8.onnx) | food101 | Int8 | 260x260x3 | STM32N6 | 2712.19 |528.12| 10328.52 | 3.0.0 |
82
+ | [efficientnetv2b2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b2_260_fft/efficientnetv2b2_260_fft_qdq_w4_81.26%_w8_18.74%_a8_100%_acc_87.24.onnx) | food101 | Int8/Int4 | 260x260x3 | STM32N6 | 2712.19 |528.12| 6865.39 | 3.0.0 |
83
+ | [efficientnetv2s_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2s_384_fft/efficientnetv2s_384_fft_qdq_int8.onnx) | food101 | Int8 | 384x384x3 | STM32N6 | 2757 | 3456 | 24262.34 | 3.0.0 |
84
+ | [efficientnetv2s_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2s_384_fft/efficientnetv2s_384_fft_qdq_w4_95.95%_w8_4.05%_a8_100%_acc_89.87.onnx) | food101 | Int8/Int4 | 384x384x3 | STM32N6 | 2757 | 3456 | 14836.94 | 3.0.0 |
85
+ | [efficientnetv2b0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b0_224/efficientnetv2b0_224_qdq_int8.onnx) | imagenet | Int8 | 224x224x3 | STM32N6 | 1911.56 | 0.0 | 7967.05 | 3.0.0 |
86
+ | [efficientnetv2b0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b0_224/efficientnetv2b0_224_qdq_w4_65.43%_w8_34.57%_a8_100%_acc_73.38.onnx) | imagenet | Int8/Int4 | 224x224x3 | STM32N6 | 1911.56 | 0.0 | 5710.05 | 3.0.0 |
87
+ | [efficientnetv2b1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b1_240/efficientnetv2b1_240_qdq_int8.onnx) | imagenet | Int8 | 240x240x3 | STM32N6 | 2604.03 | 0.0 | 9216.92 | 3.0.0 |
88
+ | [efficientnetv2b1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b1_240/efficientnetv2b1_240_qdq_w4_73.1%_w8_26.9%_a8_100%_acc_73.92.onnx) | imagenet | Int8/Int4 | 240x240x3 | STM32N6 | 2604.03 | 0.0 | 6342.67 | 3.0.0 |
89
+ | [efficientnetv2b2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b2_260/efficientnetv2b2_260_qdq_int8.onnx) | imagenet | Int8 | 260x260x3 | STM32N6 | 2712.19 | 528.12 | 11568.55 | 3.0.0 |
90
+ | [efficientnetv2b2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b2_260/efficientnetv2b2_260_qdq_w4_67.53%_w8_32.47%_a8_100%_acc_74.71.onnx) | imagenet | Int8/Int4 | 260x260x3 | STM32N6 | 2712.19 | 528.12 | 8273.17 | 3.0.0 |
91
+ | [efficientnetv2b3_300 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b3_300/efficientnetv2b3_300_qdq_int8.onnx) | imagenet | Int8 | 300x300x3 | STM32N6 | 2574.47 | 1757.81 | 16510.05 | 3.0.0 |
92
+ | [efficientnetv2b3_300 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b3_300/efficientnetv2b3_300_qdq_w4_88.31%_w8_11.69%_a8_100%_acc_78.11.onnx) | imagenet | Int8/Int4 | 300x300x3 | STM32N6 | 2574.47 | 1757.81 | 10376.74 | 3.0.0 |
93
+ | [efficientnetv2s_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2s_384/efficientnetv2s_384_qdq_int8.onnx) | imagenet | Int8 | 384x384x3 | STM32N6 | 2800 | 2592 | 25390 | 3.0.0 |
94
+ | [efficientnetv2s_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2s_384/efficientnetv2s_384_qdq_w4_95.63%_w8_4.37%_a8_100%_acc_82.25.onnx) | imagenet | Int8/Int4 | 384x384x3 | STM32N6 | 2800 | 2592 | 15458.97 | 3.0.0 |
95
+
96
+
97
+
98
+ ### Reference **NPU** inference time on food101 and imagenet dataset (see Accuracy for details on dataset)
99
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
100
+ |--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------|------------------------|
101
+ | [efficientnetv2b0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b0_224_fft/efficientnetv2b0_224_fft_qdq_int8.onnx) | food101 | Int8 | 224x224x3 | STM32N6570-DK | NPU/MCU | 62.48 | 16 | 3.0.0 |
102
+ | [efficientnetv2b0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b0_224_fft/efficientnetv2b0_224_fft_qdq_w4_90.1%_w8_9.9%_a8_100%_acc_84.47.onnx) | food101 | Int8/Int4 | 224x224x3 | STM32N6570-DK | NPU/MCU | 57.05 | 17.53 | 3.0.0 |
103
+ | [efficientnetv2b1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b1_240_fft/efficientnetv2b1_240_fft_qdq_int8.onnx) | food101 | Int8 | 240x240x3 | STM32N6570-DK | NPU/MCU | 86.55 | 11.55 | 3.0.0 |
104
+ | [efficientnetv2b1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b1_240_fft/efficientnetv2b1_240_fft_qdq_w4_91.8%_w8_8.2%_a8_100%_acc_85.71.onnx) | food101 | Int8/Int4 | 240x240x3 | STM32N6570-DK | NPU/MCU | 80.5 | 12.42 | 3.0.0 |
105
+ | [efficientnetv2b2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b2_260_fft/efficientnetv2b2_260_fft_qdq_int8.onnx) | food101 | Int8 | 260x260x3 | STM32N6570-DK | NPU/MCU | 147.21 | 6.79 | 3.0.0 |
106
+ | [efficientnetv2b2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b2_260_fft/efficientnetv2b2_260_fft_qdq_w4_81.26%_w8_18.74%_a8_100%_acc_87.24.onnx) | food101 | Int8/Int4 | 260x260x3 | STM32N6570-DK | NPU/MCU | 140.38 | 7.12 | 3.0.0 |
107
+ | [efficientnetv2s_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2s_384_fft/efficientnetv2s_384_fft_qdq_int8.onnx) | food101 | Int8 | 384x384x3 | STM32N6570-DK | NPU/MCU | 1089.83 | 0.92 | 3.0.0 |
108
+ | [efficientnetv2s_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2s_384_fft/efficientnetv2s_384_fft_qdq_w4_95.95%_w8_4.05%_a8_100%_acc_89.87.onnx) | food101 | Int8/Int4 | 384x384x3 | STM32N6570-DK | NPU/MCU | 1078.35 | 0.93 | 3.0.0 |
109
+ | [efficientnetv2b0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b0_224/efficientnetv2b0_224_qdq_int8.onnx) | imagenet | Int8 | 224x224x3 | STM32N6570-DK | NPU/MCU | 65.44 | 15.28 | 3.0.0 |
110
+ | [efficientnetv2b0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b0_224/efficientnetv2b0_224_qdq_w4_65.43%_w8_34.57%_a8_100%_acc_73.38.onnx) | imagenet | Int8/Int4 | 224x224x3 | STM32N6570-DK | NPU/MCU | 59.54 | 16.80 | 3.0.0 |
111
+ | [efficientnetv2b1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b1_240/efficientnetv2b1_240_qdq_int8.onnx) | imagenet | Int8 | 240x240x3 | STM32N6570-DK | NPU/MCU | 89.71 | 11.15 | 3.0.0 |
112
+ | [efficientnetv2b1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b1_240/efficientnetv2b1_240_qdq_w4_73.1%_w8_26.9%_a8_100%_acc_73.92.onnx) | imagenet | Int8/Int4 | 240x240x3 | STM32N6570-DK | NPU/MCU | 83.2 | 12.02 | 3.0.0 |
113
+ | [efficientnetv2b2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b2_260/efficientnetv2b2_260_qdq_int8.onnx) | imagenet | Int8 | 260x260x3 | STM32N6570-DK | NPU/MCU | 150.04 | 6.66 | 3.0.0 |
114
+ | [efficientnetv2b2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b2_260/efficientnetv2b2_260_qdq_w4_67.53%_w8_32.47%_a8_100%_acc_74.71.onnx) | imagenet | Int8/Int4 | 260x260x3 | STM32N6570-DK | NPU/MCU | 141.94 | 7.05 | 3.0.0 |
115
+ | [efficientnetv2b3_300 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b3_300/efficientnetv2b3_300_qdq_int8.onnx) | imagenet | Int8 | 300x300x3 | STM32N6570-DK | NPU/MCU | 224.03 | 4.46 | 3.0.0 |
116
+ | [efficientnetv2b3_300 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b3_300/efficientnetv2b3_300_qdq_w4_88.31%_w8_11.69%_a8_100%_acc_78.11.onnx) | imagenet | Int8/Int4 | 300x300x3 | STM32N6570-DK | NPU/MCU | 219.31 | 4.56 | 3.0.0 |
117
+ | [efficientnetv2s_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2s_384/efficientnetv2s_384_qdq_int8.onnx) | imagenet | Int8 | 384x384x3 | STM32N6570-DK | NPU/MCU | 839.14 | 1.19 | 3.0.0 |
118
+ | [efficientnetv2s_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2s_384/efficientnetv2s_384_qdq_w4_95.63%_w8_4.37%_a8_100%_acc_82.25.onnx) | imagenet | Int8/Int4 | 384x384x3 | STM32N6570-DK | NPU/MCU | 826.23 | 1.21 | 3.0.0 |
119
+
120
  ### Accuracy with Food-101 dataset
121
 
122
  Dataset details: [link](https://data.vision.ee.ethz.ch/cvl/datasets_extra/food-101/), Quotation[[3]](#3) , Number of classes: 101 , Number of images: 101 000
123
 
124
  | Model | Format | Resolution | Top 1 Accuracy |
125
  |--------------------------------------------------------------------------------------------------------------------------------------------------|--------|-----------|----------------|
126
+ | [efficientnetv2b0_224_fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b0_224_fft/efficientnetv2b0_224_fft.keras) | Float | 224x224x3 | 86.59 % |
127
+ | [efficientnetv2b0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b0_224_fft/efficientnetv2b0_224_fft_qdq_int8.onnx) | Int8 | 224x224x3 | 85.98 % |
128
+ | [efficientnetv2b0_224_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b0_224_fft/efficientnetv2b0_224_fft_qdq_w4_90.1%_w8_9.9%_a8_100%_acc_84.47.onnx)| Int8/Int4 | 224x224x3 | 84.47 % |
129
+ | [efficientnetv2b1_240_fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b1_240_fft/efficientnetv2b1_240_fft.keras) | Float | 240x240x3 | 87.71 % |
130
+ | [efficientnetv2b1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b1_240_fft/efficientnetv2b1_240_fft_qdq_int8.onnx) | Int8 | 240x240x3 | 87.09 % |
131
+ | [efficientnetv2b1_240_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b1_240_fft/efficientnetv2b1_240_fft_qdq_w4_91.8%_w8_8.2%_a8_100%_acc_85.71.onnx) | Int8/Int4 | 240x240x3 | 85.71 % |
132
+ | [efficientnetv2b2_260_fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b2_260_fft/efficientnetv2b2_260_fft.keras) | Float | 260x260x3 | 88.67 % |
133
+ | [efficientnetv2b2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b2_260_fft/efficientnetv2b2_260_fft_qdq_int8.onnx) | Int8 | 260x260x3 | 88.44 % |
134
+ | [efficientnetv2b2_260_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2b2_260_fft/efficientnetv2b2_260_fft_qdq_w4_81.26%_w8_18.74%_a8_100%_acc_87.24.onnx) | Int8/Int4 | 260x260x3 | 87.24 % |
135
+ | [efficientnetv2s_384_fft](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2s_384_fft/efficientnetv2s_384_fft.keras) | Float | 384x384x3 | 91.69 % |
136
+ | [efficientnetv2s_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2s_384_fft/efficientnetv2s_384_fft_qdq_int8.onnx) | Int8 | 384x384x3 | 91.34 % |
137
+ | [efficientnetv2s_384_fft onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/ST_pretrainedmodel_public_dataset/food101/efficientnetv2s_384_fft/efficientnetv2s_384_fft_qdq_w4_95.95%_w8_4.05%_a8_100%_acc_89.87.onnx) | Int8/Int4 | 384x384x3 | 89.87 % |
138
+
139
+
140
+ ### Accuracy with imagenet
141
 
142
  Dataset details: [link](https://www.image-net.org), Quotation[[4]](#4).
143
  Number of classes: 1000.
 
146
 
147
  | Model | Format | Resolution | Top 1 Accuracy |
148
  |------------------------------------------------------------------------------------------------------------------------------------------|--------|------------|----------------|
149
+ | [efficientnetv2b0_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b0_224/efficientnetv2b0_224.keras) | Float | 224x224x3 | 75.18 % |
150
+ | [efficientnetv2b0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b0_224/efficientnetv2b0_224_qdq_int8.onnx) | Int8 | 224x224x3 | 73.75 % |
151
+ | [efficientnetv2b0_224 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b0_224/efficientnetv2b0_224_qdq_w4_65.43%_w8_34.57%_a8_100%_acc_73.38.onnx) | Int8/Int4 | 224x224x3 | 73.38 % |
152
+ | [efficientnetv2b1_240](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b1_240/efficientnetv2b1_240.keras) | Float | 240x240x3 | 76.14 % |
153
+ | [efficientnetv2b1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b1_240/efficientnetv2b1_240_qdq_int8.onnx) | Int8 | 240x240x3 | 75.19 % |
154
+ | [efficientnetv2b1_240 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b1_240/efficientnetv2b1_240_qdq_w4_73.1%_w8_26.9%_a8_100%_acc_73.92.onnx) | Int8/Int4 | 240x240x3 | 73.92 % |
155
+ | [efficientnetv2b2_260](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b2_260/efficientnetv2b2_260.keras) | Float | 260x260x3 | 76.58 % |
156
+ | [efficientnetv2b2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b2_260/efficientnetv2b2_260_qdq_int8.onnx) | Int8 | 260x260x3 | 76.14 % |
157
+ |[efficientnetv2b2_260 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b2_260/efficientnetv2b2_260_qdq_w4_67.53%_w8_32.47%_a8_100%_acc_74.71.onnx) | Int8/Int4 | 260x260x3 | 74.71 % |
158
+ | [efficientnetv2b3_300](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b3_300/efficientnetv2b3_300.keras) | Float | 300x300x3 | 79.18 % |
159
+ | [efficientnetv2b3_300 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b3_300/efficientnetv2b3_300_qdq_int8.onnx) | Int8 | 300x300x3 | 79.05 % |
160
+ | [efficientnetv2b3_300 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2b3_300/efficientnetv2b3_300_qdq_w4_88.31%_w8_11.69%_a8_100%_acc_78.11.onnx) | Int8/Int4 | 300x300x3 | 78.11 % |
161
+ | [efficientnetv2s_384](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2s_384/efficientnetv2s_384.keras) | Float | 384x384x3 | 83.52 % |
162
+ | [efficientnetv2s_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2s_384/efficientnetv2s_384_qdq_int8.onnx) | Int8 | 384x384x3 | 83.07 % |
163
+ | [efficientnetv2s_384 onnx](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/efficientnetv2/Public_pretrainedmodel_public_dataset/ImageNet/imagenet/efficientnetv2s_384/efficientnetv2s_384_qdq_w4_95.63%_w8_4.37%_a8_100%_acc_82.25.onnx) | Int8/Int4 | 384x384x3 | 82.25 % |
164
+
165
 
166
 
167
  ## Retraining and Integration in a simple example:
 
182
 
183
  <a id="4">[4]</a>
184
  Olga Russakovsky*, Jia Deng*, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg and Li Fei-Fei.
185
+ (* = equal contribution) imagenet Large Scale Visual Recognition Challenge.