SqueezeNext
Use case : Image classification
Model description
SqueezeNext is the successor to SqueezeNet, offering improved accuracy through skip connections, bottleneck modules, and separable convolutions. It is specifically designed for hardware efficiency.
The architecture employs a two-stage bottleneck with 1x1 squeeze followed by 1x1-3x3 expand patterns, with skip connections added for improved gradient flow. Separable convolutions further reduce computational cost, and the hardware-aware design is optimized for specific hardware platforms.
SqueezeNext is ideal for applications requiring SqueezeNet-style compactness with better accuracy, and hardware platforms with specific optimization targets.
(source: https://arxiv.org/abs/1803.10615)
The model is quantized to int8 using ONNX Runtime and exported for efficient deployment.
Network information
| Network Information | Value |
|---|---|
| Framework | Torch |
| MParams | ~0.68–3.17 M |
| Quantization | Int8 |
| Provenance | https://github.com/amirgholami/SqueezeNext |
| Paper | https://arxiv.org/abs/1803.10615 |
Network inputs / outputs
For an image resolution of NxM and P classes
| Input Shape | Description |
|---|---|
| (1, N, M, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
| Output Shape | Description |
|---|---|
| (1, P) | Per-class confidence for P classes in FLOAT32 |
Recommended platforms
| Platform | Supported | Recommended |
|---|---|---|
| STM32L0 | [] | [] |
| STM32L4 | [] | [] |
| STM32U5 | [] | [] |
| STM32H7 | [] | [] |
| STM32MP1 | [] | [] |
| STM32MP2 | [] | [] |
| STM32N6 | [x] | [x] |
Performances
Metrics
- Measures are done with default STEdgeAI Core configuration with enabled input / output allocated option.
- All the models are trained from scratch on Imagenet dataset
Reference NPU memory footprint on Imagenet dataset (see Accuracy for details on dataset)
| Model | Dataset | Format | Resolution | Series | Internal RAM (KiB) | External RAM (KiB) | Weights Flash (KiB) | STEdgeAI Core version |
|---|---|---|---|---|---|---|---|---|
| sqnxt23_x100_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6 | 2086.45 | 3025 | 693.67 | 3.0.0 |
| sqnxt23_x150_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6 | 2087.48 | 6806.25 | 1453.99 | 3.0.0 |
| sqnxt23_x200_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6 | 2275.52 | 9075 | 2493.33 | 3.0.0 |
| sqnxt23v5_x150_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6 | 2087.48 | 6806.25 | 1879.24 | 3.0.0 |
| sqnxt23v5_x200_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6 | 2275.52 | 9075 | 3249.45 | 3.0.0 |
Reference NPU inference time on Imagenet dataset (see Accuracy for details on dataset)
| Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
|---|---|---|---|---|---|---|---|---|
| sqnxt23_x100_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 87.07 | 11.49 | 3.0.0 |
| sqnxt23_x150_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 127.46 | 7.85 | 3.0.0 |
| sqnxt23_x200_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 182.12 | 5.49 | 3.0.0 |
| sqnxt23v5_x100_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 86.37 | 11.58 | 3.0.0 |
| sqnxt23v5_x150_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 126.91 | 7.88 | 3.0.0 |
| sqnxt23v5_x200_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 181.01 | 5.52 | 3.0.0 |
Accuracy with Imagenet dataset
| Model | Format | Resolution | Top 1 Accuracy |
|---|---|---|---|
| sqnxt23_x100_pt | Float | 224x224x3 | 58.18 % |
| sqnxt23_x100_pt | Int8 | 224x224x3 | 57.86 % |
| sqnxt23_x150_pt | Float | 224x224x3 | 66.17 % |
| sqnxt23_x150_pt | Int8 | 224x224x3 | 65.48 % |
| sqnxt23_x200_pt | Float | 224x224x3 | 70.56 % |
| sqnxt23_x200_pt | Int8 | 224x224x3 | 70.25 % |
| sqnxt23v5_x100_pt | Float | 224x224x3 | 59.85 % |
| sqnxt23v5_x100_pt | Int8 | 224x224x3 | 59.57 % |
| sqnxt23v5_x150_pt | Float | 224x224x3 | 67.32 % |
| sqnxt23v5_x150_pt | Int8 | 224x224x3 | 66.78 % |
| sqnxt23v5_x200_pt | Float | 224x224x3 | 71.42 % |
| sqnxt23v5_x200_pt | Int8 | 224x224x3 | 71.02 % |
| Model | Format | Resolution | Top 1 Accuracy |
|---|---|---|---|
| sqnxt23_x100_pt | Float | 224x224x3 | 58.18 % |
| sqnxt23_x100_pt | Int8 | 224x224x3 | 57.86 % |
| sqnxt23_x150_pt | Float | 224x224x3 | 66.17 % |
| sqnxt23_x150_pt | Int8 | 224x224x3 | 65.48 % |
| sqnxt23_x200_pt | Float | 224x224x3 | 70.56 % |
| sqnxt23_x200_pt | Int8 | 224x224x3 | 70.25 % |
| sqnxt23v5_x100_pt | Float | 224x224x3 | 59.85 % |
| sqnxt23v5_x100_pt | Int8 | 224x224x3 | 59.57 % |
| sqnxt23v5_x150_pt | Float | 224x224x3 | 67.32 % |
| sqnxt23v5_x150_pt | Int8 | 224x224x3 | 66.78 % |
| sqnxt23v5_x200_pt | Float | 224x224x3 | 71.42 % |
| sqnxt23v5_x200_pt | Int8 | 224x224x3 | 71.02 % |
Retraining and Integration in a simple example:
Please refer to the stm32ai-modelzoo-services GitHub here
References
[1] - Dataset: Imagenet (ILSVRC 2012) — https://www.image-net.org/
[2] - Model: SqueezeNext — https://github.com/amirgholami/SqueezeNext