Skip to content

Latest commit

 

History

History
78 lines (42 loc) · 4.18 KB

File metadata and controls

78 lines (42 loc) · 4.18 KB
graph LR
    Image_Dataset_Loader["Image Dataset Loader"]
    Data_Iterator["Data Iterator"]
    Data_Source_Manager["Data Source Manager"]
    Augmentation_Wrapper["Augmentation Wrapper"]
    Basic_Image_Augmentation["Basic Image Augmentation"]
    Differential_Augmentation_Core["Differential Augmentation Core"]
    Data_Source_Manager -- "initializes" --> Image_Dataset_Loader
    Data_Source_Manager -- "initializes" --> Data_Iterator
    Data_Iterator -- "consumes data from" --> Image_Dataset_Loader
    Augmentation_Wrapper -- "utilizes" --> Basic_Image_Augmentation
    Augmentation_Wrapper -- "utilizes" --> Differential_Augmentation_Core
Loading

CodeBoardingDemoContact

Details

The Data Pipeline subsystem is responsible for the entire lifecycle of image data, from initial loading and basic preprocessing to advanced augmentation, ensuring a continuous and appropriately transformed data stream for the GAN training process.

Image Dataset Loader

Loads raw image data from a specified folder, applies initial transformations (resizing, format conversion to RGB/RGBA/L), and includes basic augmentations like random resized crop and center crop. It serves as the primary source of pre-processed image data.

Related Classes/Methods:

Data Iterator

Creates an infinite iterator over the Image Dataset Loader, ensuring a continuous and cyclical supply of data batches for the training loop, which is crucial for long-running GAN training.

Related Classes/Methods:

Data Source Manager

Orchestrates the setup of the entire data source. This involves initializing the Image Dataset Loader and wrapping it with the Data Iterator to prepare the data stream for consumption by the GAN training process.

Related Classes/Methods:

Augmentation Wrapper

Acts as a central point for applying a collection of advanced image augmentation techniques (e.g., horizontal flipping, differential augmentation) to images. This typically occurs just before the images are fed into the Discriminator during the GAN's forward pass. It also manages whether the augmentation operations are detached from the computation graph.

Related Classes/Methods:

Basic Image Augmentation

Performs fundamental, simple image transformations, specifically random horizontal flipping. This is applied directly within the Augmentation Wrapper's processing flow.

Related Classes/Methods:

Differential Augmentation Core

Implements and applies a suite of advanced differential augmentation techniques (e.g., color, translation, cutout) to images. These techniques are designed to improve GAN training stability and performance by introducing subtle, differentiable changes to the input.

Related Classes/Methods: