# blocks

Python API: `luxonis_train.nodes.blocks`

Layer blocks that the nodes build on.

The package holds convolution blocks and stacks, RepVGG-style blocks, attention and pooling blocks, and parts of the detection and
segmentation heads. It also holds the ResNet blocks and the U-Net encoder and decoder blocks.

[GeneralReparameterizableBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md)
subclasses
[Reparameterizable](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/reparameterizable.md).
It trains with parallel branches and fuses them into a single convolution when its node enters export mode.

When `aimet_torch` is installed, the import of the package tells its quantization to ignore
[DropPath](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md)
and
[UpscaleOnline](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md).

## Child Pages

 * [blocks](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md)
 * [reparameterizable](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/reparameterizable.md)
 * [resnet](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/resnet.md)
 * [unet](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/unet.md)
 * [utils](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/utils.md)