# blocks Python API: `luxonis_train.nodes.necks.reppan_neck.blocks` The upsampling and downsampling blocks of the RepPAN neck. ## Classes ### CSPDownBlock Bottom-up fusion step of [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) with a CSP block. The encode block is a [CSPStackRepBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) that maps the concatenation to `out_channels`. [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) uses this step when `block` is `"CSPStackRepBlock"`, as in the `"m"` and `"l"` variants. > **Example** > ```pycon >>> import torch >>> block = CSPDownBlock( ... 4, 8, in_channels_next=12, out_channels=16, n_repeats=2, e=0.25 ... ) >>> block.encode_block.conv_1.out_channels 4 >>> x0, x1 = torch.zeros(1, 4, 8, 8), torch.zeros(1, 12, 4, 4) >>> block(x0, x1).shape torch.Size([1, 16, 4, 4]) ``` #### Methods ##### init ```python def __init__(in_channels: int, downsample_out_channels: int, in_channels_next: int, out_channels: int, n_repeats: int, e: float): ``` Initialize the downsampling layer and the CSP block. Parameters * `in_channels` (`int`): Number of channels of the fine input. * `downsample_out_channels` (`int`): Number of channels after the downsampling convolution. * `in_channels_next` (`int`): Number of channels of the lateral input, which the step concatenates. * `out_channels` (`int`): Number of output channels. * `n_repeats` (`int`): Controls the number of RepVGG-style blocks in the [CSPStackRepBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md). Each [BottleRep](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) holds two of them, so the stack has `max(1, n_repeats // 2)` [BottleRep](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) blocks and `2 * max(1, n_repeats // 2)` RepVGG-style blocks. * `e` (`float`): Fraction of `out_channels` in each of the two paths of the [CSPStackRepBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md). ### CSPUpBlock Top-down fusion step of [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) with a CSP block. The encode block is a [CSPStackRepBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) that maps the concatenation to `out_channels`. [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) uses this step when `block` is `"CSPStackRepBlock"`, as in the `"m"` and `"l"` variants. > **Example** > ```pycon >>> block = CSPUpBlock( ... 32, in_channels_next=16, out_channels=8, n_repeats=4, e=0.5 ... ) >>> block.encode_block.conv_1.out_channels 4 >>> len(block.encode_block.rep_stack) 2 ``` #### Methods ##### init ```python def __init__(in_channels: int, in_channels_next: int, out_channels: int, n_repeats: int, e: float): ``` Initialize the upsampling layers and the CSP block. Parameters * `in_channels` (`int`): Number of channels of the coarse input. * `in_channels_next` (`int`): Number of channels of the finer input, which the step concatenates. * `out_channels` (`int`): Number of output channels. * `n_repeats` (`int`): Controls the number of RepVGG-style blocks in the [CSPStackRepBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md). Each [BottleRep](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) holds two of them, so the stack has `max(1, n_repeats // 2)` [BottleRep](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) blocks and `2 * max(1, n_repeats // 2)` RepVGG-style blocks. * `e` (`float`): Fraction of `out_channels` in each of the two paths of the [CSPStackRepBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md). ### PANDownBlockBase Base class of the bottom-up fusion steps of [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md). A `3x3` [ConvBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) with stride `2`, padding `1`, batch norm, and ReLU halves the height and the width of the fine input. An odd size rounds up. The block concatenates the result with the lateral input along the channel axis and runs `encode_block` on it. A subclass selects `encode_block`. #### Methods ##### init ```python def __init__(in_channels: int, downsample_out_channels: int, encode_block: nn.Module): ``` Initialize the downsampling convolution. Parameters * `in_channels` (`int`): Number of channels of the fine input. * `downsample_out_channels` (`int`): Number of channels after the downsampling convolution. * `encode_block` (`nn.Module`): Block that runs on the concatenation. Its input has `downsample_out_channels` plus the channels of the lateral input. ##### forward ```python def forward(x0: Tensor, x1: Tensor) -> Tensor: ``` Downsample the fine map and fuse it with the lateral map. > **Example** > ```pycon >>> import torch >>> block = RepDownBlock( ... 4, 8, in_channels_next=12, out_channels=16, n_repeats=1 ... ) >>> x0, x1 = torch.zeros(1, 4, 8, 8), torch.zeros(1, 12, 4, 4) >>> block(x0, x1).shape torch.Size([1, 16, 4, 4]) ``` Parameters * `x0` (`Tensor`): Fine map of shape `[B, in_channels, H, W]`. * `x1` (`Tensor`): Lateral map of shape `[B, C1, H / 2, W / 2]`. An odd `H` or `W` rounds up. In the subclasses, `C1` is `in_channels_next`. Returns * `Tensor`: The output of `encode_block`. In the subclasses, it has the shape `[B, out_channels, H / 2, W / 2]`. #### Attributes ##### downsample ##### encode_block ### PANUpBlockBase Base class of the top-down fusion steps of [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md). A `1x1` [ConvBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) with batch norm and ReLU maps the coarse input to `out_channels`. A `2x2` transposed convolution with stride `2` then doubles the height and the width. The block concatenates the result with the finer input along the channel axis and runs `encode_block` on it. A subclass selects `encode_block`. #### Methods ##### init ```python def __init__(in_channels: int, out_channels: int, encode_block: nn.Module): ``` Build the `1x1` convolution and the upsampling layer. Parameters * `in_channels` (`int`): Number of channels of the coarse input. * `out_channels` (`int`): Number of channels after the `1x1` convolution. The upsampling layer keeps this number. * `encode_block` (`nn.Module`): Block that runs on the concatenation. Its input has `out_channels` plus the channels of the finer input. ##### forward ```python def forward(x0: Tensor, x1: Tensor) -> tuple[Tensor, Tensor]: ``` Upsample the coarse map and fuse it with the finer map. > **Example** > ```pycon >>> import torch >>> block = RepUpBlock( ... 32, in_channels_next=16, out_channels=8, n_repeats=1 ... ) >>> x0, x1 = torch.zeros(1, 32, 4, 4), torch.zeros(1, 16, 8, 8) >>> conv_out, out = block(x0, x1) >>> conv_out.shape, out.shape (torch.Size([1, 8, 4, 4]), torch.Size([1, 8, 8, 8])) ``` Parameters * `x0` (`Tensor`): Coarse map of shape `[B, in_channels, H, W]`. * `x1` (`Tensor`): Finer map of shape `[B, C1, 2 * H, 2 * W]`. In the subclasses, `C1` is `in_channels_next`. Returns * `tuple[Tensor, Tensor]`: The tuple `(conv_out, out)`. `conv_out` is the output of the `1x1` convolution, of shape `[B, out_channels, H, W]`. [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) gives it to a bottom-up step as the lateral input. `out` is the output of `encode_block`. In the subclasses, `out` has the shape `[B, out_channels, 2 * H, 2 * W]`. #### Attributes ##### conv ##### encode_block ##### upsample ### RepDownBlock Bottom-up fusion step of [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) with RepVGG-style blocks. The encode block is a [BlockRepeater](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) of `n_repeats` [GeneralReparameterizableBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) layers. The first layer maps the concatenation to `out_channels`. The other layers keep that number. [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) uses this step when `block` is `"RepBlock"`, as in the `"n"` and `"s"` variants. > **Example** > ```pycon >>> import torch >>> block = RepDownBlock( ... 4, 8, in_channels_next=12, out_channels=16, n_repeats=3 ... ) >>> block.encode_block(torch.zeros(1, 20, 4, 4)).shape torch.Size([1, 16, 4, 4]) ``` #### Methods ##### init ```python def __init__(in_channels: int, downsample_out_channels: int, in_channels_next: int, out_channels: int, n_repeats: int): ``` Initialize the downsampling layer and the RepVGG-style stack. Parameters * `in_channels` (`int`): Number of channels of the fine input. * `downsample_out_channels` (`int`): Number of channels after the downsampling convolution. * `in_channels_next` (`int`): Number of channels of the lateral input, which the step concatenates. * `out_channels` (`int`): Number of output channels. * `n_repeats` (`int`): Number of [GeneralReparameterizableBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) layers. A value below `1` still builds one layer. ### RepUpBlock Top-down fusion step of [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) with RepVGG-style blocks. The encode block is a [BlockRepeater](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) of `n_repeats` [GeneralReparameterizableBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) layers. The first layer maps the concatenation to `out_channels`. The other layers keep that number. [RepPANNeck](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/necks/reppan_neck/reppan_neck.md) uses this step when `block` is `"RepBlock"`, as in the `"n"` and `"s"` variants. > **Example** > ```pycon >>> import torch >>> block = RepUpBlock( ... 32, in_channels_next=16, out_channels=8, n_repeats=3 ... ) >>> block.encode_block(torch.zeros(1, 24, 4, 4)).shape torch.Size([1, 8, 4, 4]) ``` #### Methods ##### init ```python def __init__(in_channels: int, in_channels_next: int, out_channels: int, n_repeats: int): ``` Initialize the upsampling layers and the RepVGG-style stack. Parameters * `in_channels` (`int`): Number of channels of the coarse input. * `in_channels_next` (`int`): Number of channels of the finer input, which the step concatenates. * `out_channels` (`int`): Number of output channels. * `n_repeats` (`int`): Number of [GeneralReparameterizableBlock](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/blocks/blocks.md) layers. A value below `1` still builds one layer.