| --- |
| license: other |
| tags: |
| - heal |
| - horizon |
| --- |
| |
| # HENet-tinyM |
|
|
| HENet uses depthwise conv to build lightweight blocks with LayerScale for stable training; the tiny-M configuration has widths 64/128/192/384 and depths 4/3/8/6, with GroupDWCB in the first stage and parameter count in the tiny range. |
|
|
| --- |
|
|
| ## Deployment Metrics |
|
|
| ### Model Parameters |
|
|
| | Model | Model Input | Backbone | Neck | Model Output | |
| |---|---|---|---|---| |
| | HENet-tinym | `1x3x224x224` | HENet | — | classification logits `(B,1000)` | |
|
|
| ### Accuracy Metrics |
|
|
| | March | Metric | float | calibration | qat | hbm | |
| | --- | --- | --- | --- | --- | --- | |
| | J6M | Accuracy | 0.784 | 0.7795 | — | 0.7801 | |
| | | TopKAccuracy(5) | — | — | — | — | |
|
|
| > Data tested with `march = March.NASH_M` (J6M); this task has no QAT stage (`—` in the qat column). |
| > |
| > HEAL versions: heal 0.0.2 / hbdk4-compiler 4.11.11 / horizon_plugin_pytorch 3.3.10. |
| |
| ### Performance Metrics |
| |
| > **Performance test methodology**: FPS is measured with 8 threads on a single core; Latency is measured with single core, single thread; Memory is peak DDR usage. |
| |
| | March | latency (ms) | fps | Memory Usage | |
| |---|---|---|---| |
| | J6M | 0.51 | 3435.50 | 12.90 | |
| | J6P | 0.41 | 11410.39 | 13.10 | |
| | J6B | - | - | - | |
| |
| J6B performance is not available for this model. |
| |
| --- |
| |
| ## Model Overview |
| |
| ### Core Design |
| |
| HENet uses depthwise conv to build lightweight blocks with LayerScale for stable training; the tiny-M configuration has widths 64/128/192/384 and depths 4/3/8/6, with GroupDWCB in the first stage and parameter count in the tiny range. |
| |
| - **Task type**: Image Classification. |
| - **backbone**: HENet (tiny-M configuration, `depth/block_nums=[4,3,8,6]`, `width/embed_dims=[64,128,192,384]`, `block_cls=["GroupDWCB","GroupDWCB","AltDWCB","DWCB"]`, GELU activation, LayerScale, S2D downsampling, `num_classes=1000`), uses depthwise conv to build lightweight blocks with LayerScale for stable training; first stage uses GroupDWCB. |
| - **neck**: — (HENet has a built-in fully connected classification head; no separate neck). |
| - **classification head**: HENet built-in fully connected classification head (`include_top=True`, `feature_mix_channel=1024`). |
| - **Loss function**: `SoftTargetCrossEntropy` (soft-label cross-entropy with mixup). |
| - **Model input**: single RGB image at resolution `224 × 224` (`1x3x224x224`). |
| - **Model output**: 1000-class prediction logits; argmax gives the predicted class. |
|
|
| ### Official Repo and Paper |
|
|
| HybridEfficient Network is an efficient backbone designed by Horizon Robotics for the Journey series chips. |
|
|
| ### Reference |
|
|
| For more J6 chip deployment details, see https://developer.horizon.auto/blog/14094 |
|
|
