| --- |
| license: other |
| tags: |
| - heal |
| - horizon |
| --- |
| |
| # MixVarGENet |
|
|
| MixVarGENet builds a lightweight backbone with mixed-variant blocks (mixvarge); the first two stages use f2/f4 base blocks, the last two stages use f2_gb16 blocks with groups, downsampling stage by stage to stride 32. |
| |
| --- |
| |
| ## Deployment Metrics |
| |
| ### Model Parameters |
| |
| | Model | Model Input | Backbone | Neck | Model Output | |
| |---|---|---|---|---| |
| | MixVarGENet | `1x3x224x224` | MixVarGENet | 5 stage mixvarge blocks | classification logits `(B,1000)` | |
| |
| ### Accuracy Metrics |
| |
| | March | Metric | float | calibration | qat | hbm | |
| | --- | --- | --- | --- | --- | --- | |
| | J6M | Accuracy | 0.716 | 0.7116 | — | 0.7116 | |
| |
| > 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 | 0.36 | 5464.88 | 6.40 | |
| | J6P | 0.33 | 9374.11 | 6.40 | |
| | J6B | - | - | - | |
|
|
| J6B performance is not available for this model. |
|
|
| --- |
|
|
| ## Model Overview |
|
|
| ### Core Design |
|
|
| MixVarGENet builds a lightweight backbone with mixed-variant blocks (mixvarge); the first two stages use f2/f4 base blocks, the last two stages use f2_gb16 blocks with groups, downsampling stage by stage to stride 32. |
| |
| - **Task type**: Image Classification. |
| - **backbone**: MixVarGENet (`type=MixVarGENet`, 5 stages: stride 2/4/8/16/32, `num_classes=1000`), builds a lightweight backbone with mixed-variant blocks (mixvarge), downsampling stage by stage to stride 32. |
| - **neck**: 5 stages of mixvarge blocks: `mixvarge_f2`, `mixvarge_f4`, `mixvarge_f2_gb16` (last two stages use group blocks, gb16), channels 32→32→64→96→160. |
| - **classification head**: MixVarGENet built-in fully connected classification head, directly outputs 1000-class logits. |
| - **Loss function**: `CEWithLabelSmooth` (cross-entropy with label smoothing). |
| - **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 |
|
|
| MixVarGENet is an efficient network architecture proposed by Horizon Robotics, optimized for the Journey series chips. |
|
|
| ### Reference |
|
|
| For more J6 chip deployment details, see https://developer.horizon.auto/blog/14084 |
|
|
