wehub-resource-sync
72 lines
3.2 KiB
Markdown
72 lines
3.2 KiB
Markdown
# GroupNormalizationPlugin [DEPRECATED]
|
|
|
|
**This plugin is deprecated since TensorRT 10.12 and will be removed in a future release. No alternatives are planned to be provided.**
|
|
|
|
**Table Of Contents**
|
|
- [Description](#description)
|
|
* [Non-support for Blackwell and later platforms](#non-support-for-blackwell-and-later-platforms)
|
|
* [Structure](#structure)
|
|
- [Parameters](#parameters)
|
|
- [Additional Resources](#additional-resources)
|
|
- [License](#license)
|
|
- [Changelog](#changelog)
|
|
- [Known Issues](#known-issues)
|
|
|
|
## Description
|
|
|
|
The `GroupNormalizationPlugin` implements Group Normalization, which divides channels into groups and computes normalization statistics within each group. This is particularly useful for vision models where batch sizes may be small.
|
|
|
|
### Non-support for Blackwell and later platforms
|
|
|
|
As of TensorRT 10.7, usage of this plugin is not supported on Blackwell or later platforms.
|
|
This plugin can be replaced by TensorRT's native `INormalizationLayer`([C++](https://docs.nvidia.com/deeplearning/tensorrt/api/c_api/classnvinfer1_1_1_i_normalization_layer.html), [Python](https://docs.nvidia.com/deeplearning/tensorrt/operators/docs/Normalization.html)).
|
|
|
|
|
|
**Note:** This plugin remains supported on pre-Blackwell platforms.
|
|
|
|
### Structure
|
|
|
|
The plugin takes three inputs:
|
|
1. Input tensor with shape `[N, C, H, W]` (batch, channels, height, width), where `C` must be divisible by `num_groups`. (See [Parameters](#parameters) for more details on `num_groups`)
|
|
2. Scale parameters (per-channel, shape `[C]`)
|
|
3. Bias parameters (per-channel, shape `[C]`)
|
|
|
|
It produces one output with the same dimensions as the input. The normalization is computed as:
|
|
```
|
|
group_mean = mean(input, group)
|
|
group_var = variance(input, group)
|
|
output = gamma (input - group_mean) / sqrt(group_var + epsilon) + beta
|
|
```
|
|
|
|
Key differences from Instance Normalization:
|
|
- Normalizes across channel groups rather than individual channels
|
|
- More stable for small batch sizes
|
|
- Groups channels to capture cross-channel dependencies
|
|
|
|
## Parameters
|
|
|
|
| Parameter | Type | Description |
|
|
|--------------|---------|-------------|
|
|
| `epsilon` | float | Small value added to variance for numerical stability (default: 1e-5) |
|
|
| `num_groups` | int32 | Number of groups to split channels into; must evenly divide C |
|
|
|
|
## Additional Resources
|
|
|
|
- **Original Paper**: [Group Normalization](https://arxiv.org/abs/1803.08494)
|
|
- **ONNX Operator**: [GroupNormalization](https://github.com/onnx/onnx/blob/main/docs/Operators.md#GroupNormalization)
|
|
- **TensorRT Documentation**: [INormalizationLayer](https://docs.nvidia.com/deeplearning/tensorrt/api/c_api/classnvinfer1_1_1_i_normalization_layer.html)
|
|
- [Master README](../README.md) - Back to main documentation
|
|
|
|
## License
|
|
For terms and conditions for use, reproduction, and distribution, see the [TensorRT Software License Agreement](https://docs.nvidia.com/deeplearning/sdk/tensorrt-sla/index.html).
|
|
|
|
## Changelog
|
|
- **May 2025**: Add deprecation note.
|
|
|
|
- **Feb 2025**: Initial release of this README, Deprecation and non-support notice added.
|
|
|
|
## Known Issues
|
|
- Limited to FP32 precision (native implementation supports mixed precision)
|
|
- No NHWC layout support (native implementation supports multiple layouts)
|
|
- Batch size must be known during network creation
|