graph LR
models_convnextv2_ConvNeXtV2["models.convnextv2.ConvNeXtV2"]
models_convnextv2_Block["models.convnextv2.Block"]
models_convnextv2_FactoryFunctions["models.convnextv2.FactoryFunctions"]
models_fcmae_FCMAE["models.fcmae.FCMAE"]
models_convnextv2_ConvNeXtV2 -- "composes" --> models_convnextv2_Block
models_convnextv2_FactoryFunctions -- "instantiates" --> models_convnextv2_ConvNeXtV2
models_fcmae_FCMAE -- "utilizes" --> models_convnextv2_ConvNeXtV2
The ConvNeXtV2 subsystem is centered around the ConvNeXtV2 model, which is fundamentally built from modular Block components. Each Block represents a core processing unit, and ConvNeXtV2 sequences these blocks to form the complete neural network. To facilitate easy model creation, FactoryFunctions provide pre-configured instances of ConvNeXtV2. Furthermore, the FCMAE component demonstrates a practical application of the ConvNeXtV2 architecture, leveraging it as an encoder within a larger autoencoder framework for self-supervised learning. This structure highlights a clear separation of concerns: modular building blocks, a composite model, and higher-level applications.
This class defines the complete neural network structure for the ConvNeXtV2 model, encompassing layer definitions, weight initialization, and the forward pass logic. It orchestrates the data flow through the network by composing and processing data through its internal Block instances.
Related Classes/Methods:
This component encapsulates a specific set of operations, such as convolutions, normalizations, and activations, forming a single, reusable layer or stage within the ConvNeXtV2 network. It processes input data and passes the output to the next Block in the sequence, contributing to the overall forward pass.
Related Classes/Methods:
These functions (convnextv2_atto, convnextv2_tiny, etc.) provide a standardized and abstracted way to create different scaled or configured instances of the ConvNeXtV2 model. They abstract the complex instantiation logic and configuration details, allowing users to easily select and instantiate desired model variants.
Related Classes/Methods:
This class implements the Fully Convolutional Masked Autoencoder (FCMAE) architecture. It leverages a SparseConvNeXtV2 (a variant of ConvNeXtV2) as its encoder and includes a decoder and prediction head for reconstructing masked image patches, primarily for self-supervised pre-training.
Related Classes/Methods: