InceptionNeXt¶

ConvNet Image Classification

class lucid.models.InceptionNeXt(num_classes: int = 1000, depths: list[int] = [3, 3, 9, 3], dims: list[int] = [96, 192, 384, 768], token_mixers: ~typing.Type[~lucid.nn.module.Module] = <class 'lucid.nn.modules.linear.Identity'>, mlp_ratios: list[int] = [4, 4, 4, 3], head_fn: ~typing.Type[~lucid.nn.module.Module] = <class 'lucid.models.imgclf.inception_next._MLPHead'>, drop_rate: float = 0.0, drop_path_rate: float = 0.0, ls_init_value: float = 1e-06)¶

The InceptionNeXt class implements a modernized version of the Inception architecture, leveraging depthwise convolutions, multi-layer perceptrons (MLPs), and advanced token mixing techniques for enhanced performance and efficiency.

Class Signature¶

class InceptionNeXt(nn.Module):
    def __init__(
        self,
        num_classes: int = 1000,
        depths: list[int] = [3, 3, 9, 3],
        dims: list[int] = [96, 192, 384, 768],
        token_mixers: Type[nn.Module] = nn.Identity,
        mlp_ratios: list[int] = [4, 4, 4, 3],
        head_fn: Type[nn.Module] = _MLPHead,
        drop_rate: float = 0.0,
        drop_path_rate: float = 0.0,
        ls_init_value: float = 1e-6,
    ) -> None

Parameters¶

num_classes (int, optional): Number of output classes for the final classification layer. Defaults to 1000.
depths (list[int], optional): Number of blocks at each stage of the model. Defaults to [3, 3, 9, 3].
dims (list[int], optional): Dimensionality of feature maps at different stages. Defaults to [96, 192, 384, 768].
token_mixers (Type[nn.Module], optional): Type of token mixer module applied in each block. Defaults to nn.Identity.
mlp_ratios (list[int], optional): Expansion ratios for the MLP layers at different stages. Defaults to [4, 4, 4, 3].
head_fn (Type[nn.Module], optional): Function to construct the classification head. Defaults to _MLPHead.
drop_rate (float, optional): Dropout rate applied to MLP layers. Defaults to 0.0.
drop_path_rate (float, optional): Stochastic depth drop path rate. Defaults to 0.0.
ls_init_value (float, optional): Initial value for layer scale. Defaults to 1e-6.

Examples¶

from lucid.models import InceptionNeXt

# Instantiate InceptionNeXt-Tiny
model_tiny = InceptionNeXt(num_classes=1000, depths=[3, 3, 9, 3], dims=[96, 192, 384, 768])

# Forward pass with a random input
input_tensor = lucid.random.randn(1, 3, 224, 224)  # Batch size of 1, ImageNet resolution
output = model_tiny(input_tensor)
print(output.shape)  # Output shape: (1, 1000)