File size: 5,142 Bytes
1218a5a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 | ---
license: apache-2.0
pipeline_tag: image-classification
---
# HardNet
## **Use case** : `Image classification`
# Model description
Harmonic DenseNet (HardNet) is a memory-efficient variant of DenseNet that optimizes for both **computational efficiency and memory access cost**. It introduces a harmonic pattern in the dense connections to reduce redundant feature computations.
HardNet features **harmonic dense connections** that reduce connection patterns to minimize memory bandwidth, while maintaining the benefits of DenseNet's feature reuse. The architecture combines **depthwise separable convolutions** with dense blocks for enhanced efficiency.
Designed for practical hardware deployment, HardNet provides DenseNet-like feature richness with lower memory cost on edge devices.
(source: https://arxiv.org/abs/1909.00948)
The model is quantized to **int8** using **ONNX Runtime** and exported for efficient deployment.
## Network information
| Network Information | Value |
|--------------------|-------|
| Framework | Torch |
| MParams | ~3.43 M |
| Quantization | Int8 |
| Provenance | https://github.com/PingoLH/Pytorch-HarDNet |
| Paper | https://arxiv.org/abs/1909.00948 |
## 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 |
|-------|---------|--------|------------|--------|--------------|--------------|---------------|----------------------|
| [hardnet39ds_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/hardnet_pt/Public_pretrainedmodel_public_dataset/Imagenet/hardnet39ds_pt_224/hardnet39ds_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 1476.12 | 0 | 3516.67 | 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 |
|-------|---------|--------|--------|------------|-------|-----------------|-------------------|---------------------|
| [hardnet39ds_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/hardnet_pt/Public_pretrainedmodel_public_dataset/Imagenet/hardnet39ds_pt_224/hardnet39ds_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 65.81 | 15.19 | 3.0.0 |
### Accuracy with Imagenet dataset
| Model | Format | Resolution | Top 1 Accuracy |
| --- | --- | --- | --- |
| [hardnet39ds_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/hardnet_pt/Public_pretrainedmodel_public_dataset/Imagenet/hardnet39ds_pt_224/hardnet39ds_pt_224.onnx) | Float | 224x224x3 | 74.38 % |
| [hardnet39ds_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/hardnet_pt/Public_pretrainedmodel_public_dataset/Imagenet/hardnet39ds_pt_224/hardnet39ds_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 73.61 % |
Dataset details: [link](https://www.image-net.org)
Number of classes: 1000.
To perform the quantization, we calibrated the activations with a random subset of the training set.
For the sake of simplicity, the accuracy reported here was estimated on the 50000 labelled images of the validation set.
| Model | Format | Resolution | Top 1 Accuracy |
| --- | --- | --- | --- |
| [hardnet39ds_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/hardnet_pt/Public_pretrainedmodel_public_dataset/Imagenet/hardnet39ds_pt_224/hardnet39ds_pt_224.onnx) | Float | 224x224x3 | 74.38 % |
| [hardnet39ds_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/hardnet_pt/Public_pretrainedmodel_public_dataset/Imagenet/hardnet39ds_pt_224/hardnet39ds_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 73.61 % |
## Retraining and Integration in a simple example:
Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
# References
<a id="1">[1]</a> - **Dataset**: Imagenet (ILSVRC 2012) — https://www.image-net.org/
<a id="2">[2]</a> - **Model**: HarDNet — https://github.com/PingoLH/Pytorch-HarDNet |