The STM32 model zoo includes several models for image classification use cases pre-trained on custom and public datasets. Under each model directory, you can find the following model categories:
Public_pretrainedmodel_public_datasetcontains public image classification models trained on public datasets.ST_pretrainedmodel_custom_datasetcontains different image classification models trained on ST custom datasets using our training scripts.ST_pretrainedmodel_public_datasetcontains different image classification models trained on various public datasets following the training section in STM32 model zoo.
The table below summarizes the performance of the models, as well as their memory footprints generated using STM32Cube.AI for deployment purposes.
By default, the results are provided for quantized Int8 models. When nothing is precised in the model name, training is done using transfer learning technique from a pre-trained model. Else, "tfs" stands for "training from scratch".
| Models | Implementation | Dataset | Input Resolution | Top 1 Accuracy (%) | MACCs (M) | Activation RAM (KiB) | Weights Flash (KiB) | STM32Cube.AI version | Source |
|---|---|---|---|---|---|---|---|---|---|
| MobileNet v1 0.25 | TensorFlow | Flowers | 224x224x3 | 83.24% | 41 | 202.14 | 214.69 | 7.3.0 | link |
| MobileNet v1 0.5 | TensorFlow | Flowers | 224x224x3 | 89.51% | 149.3 | 404.28 | 812.61 | 7.3.0 | link |
| MobileNet v1 0.25 tfs | TensorFlow | Flowers | 96x96x3 | 82.15% | 7.550 | 38.64 | 214.69 | 7.3.0 | link |
| MobileNet v1 0.25 tfs | TensorFlow | Flowers | 96x96x1 | 72.75% | 7.550 | 38.64 | 214.69 | 7.3.0 | link |
| MobileNet v2 0.35 | TensorFlow | Flowers | 128x128x3 | 86.38% | 19 | 224.5 | 406.86 | 7.3.0 | link |
| MobileNet v2 0.35 | TensorFlow | Flowers | 224x224x3 | 89.37% | 58.4 | 686.5 | 406.86 | 7.3.0 | link |
| FdMobileNet 0.25 tfs | TensorFlow | Flowers | 224x224x3 | 85.97% | 12.1 | 152.05 | 128.32 | 7.3.0 | link |
| FdMobileNet 0.25 tfs | TensorFlow | Flowers | 128x128x3 | 84.2% | 3.953 | 51.18 | 128.32 | 7.3.0 | link |
| ST FdMobileNet v1 tfs | TensorFlow | Flowers | 224x224x3 | 86.92% | 24.74 | 206.87 | 144.93 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST FdMobileNet v1 tfs | TensorFlow | Flowers | 128x128x3 | 84.74% | 8.08 | 70.75 | 144.93 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST EfficientNet LC v1 tfs | TensorFlow | Flowers | 224x224x3 | 90.46% | 62.49 | 308.8 | 505.41 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST EfficientNet LC v1 tfs | TensorFlow | Flowers | 128x128x3 | 87.19% | 20.46 | 104.61 | 505.41 | 7.3.0 | Please Contact Edge.ai@st.com |
| SqueezeNet v1.0 tfs | TensorFlow | Flowers | 128x128x3 | 81.34% | 175.00 | 450.0 | 535.69 | 7.3.0 | link |
| SqueezeNet v1.1 tfs | TensorFlow | Flowers | 128x128x3 | 78.61% | 82.09 | 240.25 | 716.70 | 7.3.0 | link |
| SqueezeNet v1.1 tfs | TensorFlow | Flowers | 224x224x3 | 81.06% | 266,33 | 756.25 | 733.83 | 7.3.0 | link |
| MobileNet v1 0.25 | TensorFlow | Plant-village | 224x224x3 | 86.22% | 41.1 | 202.14 | 223.32 | 7.3.0 | link |
| MobileNet v1 0.5 | TensorFlow | Plant-village | 224x224x3 | 92.01% | 149.3 | 404.28 | 829.75 | 7.3.0 | link |
| MobileNet v2 0.35 | TensorFlow | Plant-village | 128x128x3 | 91.35% | 19.1 | 224.5 | 449.5 | 7.3.0 | link |
| MobileNet v2 0.35 | TensorFlow | Plant-village | 224x224x3 | 91.62% | 58.5 | 686.5 | 449.5 | 7.3.0 | link |
| FdMobileNet 0.25 tfs | TensorFlow | Plant-village | 224x224x3 | 99.59% | 12.1 | 152.05 | 136.95 | 7.3.0 | link |
| FdMobileNet 0.25 tfs | TensorFlow | Plant-village | 128x128x3 | 98.35% | 3.96 | 51.18 | 136.95 | 7.3.0 | link |
| ST FdMobileNet v1 tfs | TensorFlow | Plant-village | 224x224x3 | 99.79% | 24.75 | 206.87 | 153.56 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST FdMobileNet v1 tfs | TensorFlow | Plant-village | 128x128x3 | 96.08% | 8.09 | 70.75 | 153.56 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST EfficientNet LC v1 tfs | TensorFlow | Plant-village | 224x224x3 | 99.76% | 62.5 | 308.8 | 524.67 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST EfficientNet LC v1 tfs | TensorFlow | Plant-village | 128x128x3 | 99.4% | 20.5 | 104.61 | 524.67 | 7.3.0 | Please Contact Edge.ai@st.com |
| SqueezeNet v1.0 tfs | TensorFlow | Plant-village | 128x128x3 | 98.29% | 175.63 | 450.0 | 552.88 | 7.3.0 | link |
| SqueezeNet v1.1 tfs | TensorFlow | Plant-village | 128x128x3 | 99.04% | 82.09 | 240.25 | 716.70 | 7.3.0 | link |
| SqueezeNet v1.1 tfs | TensorFlow | Plant-village | 224x224x3 | 99.32% | 269,28 | 686.5 | 449.5 | 7.3.0 | link |
| MobileNet v1 0.25 | TensorFlow | Food-101 | 224x224x3 | 38.02% | 41.1 | 202.14 | 239.07 | 7.3.0 | link |
| MobileNet v1 0.5 | TensorFlow | Food-101 | 224x224x3 | 48.5% | 149.3 | 404.28 | 860.99 | 7.3.0 | link |
| MobileNet v2 0.35 | TensorFlow | Food-101 | 128x128x3 | 41.58% | 19.2 | 224.5 | 527.24 | 7.3.0 | link |
| MobileNet v2 0.35 | TensorFlow | Food-101 | 224x224x3 | 48.67% | 58.6 | 686.5 | 527.24 | 7.3.0 | link |
| FdMobileNet 0.25 tfs | TensorFlow | Food-101 | 224x224x3 | 56.36% | 12.125 | 152.05 | 152.69 | 7.3.0 | link |
| FdMobileNet 0.25 tfs | TensorFlow | Food-101 | 128x128x3 | 44.76% | 3.979 | 51.18 | 152.69 | 7.3.0 | link |
| ST FdMobileNet v1 tfs | TensorFlow | Food-101 | 224x224x3 | 63.6% | 24.77 | 206.87 | 169.3 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST FdMobileNet v1 tfs | TensorFlow | Food-101 | 128x128x3 | 52.48% | 8.108 | 70.75 | 169.3 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST EfficientNet LC v1 tfs | TensorFlow | Food-101 | 224x224x3 | 72.53% | 62.55 | 308.8 | 559.79 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST EfficientNet LC v1 tfs | TensorFlow | Food-101 | 128x128x3 | 61.3% | 20.52 | 104.61 | 559.79 | 7.3.0 | Please Contact Edge.ai@st.com |
| SqueezeNet v1.0 tfs | TensorFlow | Food-101 | 128x128x3 | 57.8% | 176.78 | 450 | 584.24 | 7.3.0 | link |
| SqueezeNet v1.1 tfs | TensorFlow | Food-101 | 128x128x3 | 52.18% | 83.65 | 240.25 | 765.082 | 7.3.0 | link |
| SqueezeNet v1.1 tfs | TensorFlow | Food-101 | 224x224x3 | 56.43% | 274.65 | 756.25 | 765.082 | 7.3.0 | link |
| MobileNet v2 0.35 | TensorFlow | person | 128x128x3 | 91.05% | 19,09 | 245.5 | 401.84 | 7.3.0 | link |
| MobileNet v2 0.35 tfs | TensorFlow | person | 128x128x3 | 86.22% | 19,09 | 245.5 | 401.84 | 7.3.0 | link |
| MobileNet v1 0.25 | TensorFlow | visual wake word | 96x96x3 | 85.2% | 7,494 | 40.19 | 213.93 | 7.3.0 | link |
| ResNet v1 8 tfs | TensorFlow | cifar10 | 32x32x3 | 84.95% | 12.54 | 55.28 | 76.89 | 7.3.0 | link |
| ST ResNet 8 Hybrid v1 tfs * | TensorFlow | cifar10 | 32x32x3 | 85.89% | 11.45 | 72 | 66.28 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST ResNet 8 Hybrid v2 tfs * | TensorFlow | cifar10 | 32x32x3 | 85.24% | 10.10 | 72 | 85.8 | 7.3.0 | Please Contact Edge.ai@st.com |
| ST ResNet 8 Hybrid v3 tfs * | TensorFlow | cifar10 | 32x32x3 | 84.37% | 8.13 | 72 | 66.28 | 7.3.0 | Please Contact Edge.ai@st.com |
| ResNet v1 32 tfs | TensorFlow | cifar100 | 32x32x3 | 66.3 % | 69.28 | 55.28 | 464.37 | 7.3.0 | link |
| ST MNIST Byclass v1 tfs | TensorFlow | EMNIST-Byclass | 28x28x1 | 93.38% | 1.081 | 14.13 | 10.08 | 7.3.0 | link |
* Quantization on <= 8 bits model results