Skip to content

Latest commit

 

History

History
103 lines (57 loc) · 6.87 KB

File metadata and controls

103 lines (57 loc) · 6.87 KB
graph LR
    easy_rec_python_main["easy_rec.python.main"]
    easy_rec_python_model_easy_rec_estimator["easy_rec.python.model.easy_rec_estimator"]
    easy_rec_python_builders_optimizer_builder["easy_rec.python.builders.optimizer_builder"]
    easy_rec_python_builders_loss_builder["easy_rec.python.builders.loss_builder"]
    easy_rec_python_loss["easy_rec.python.loss"]
    easy_rec_python_compat_optimizers["easy_rec.python.compat.optimizers"]
    easy_rec_python_compat_sync_replicas_optimizer["easy_rec.python.compat.sync_replicas_optimizer"]
    easy_rec_python_compat_early_stopping["easy_rec.python.compat.early_stopping"]
    easy_rec_python_main -- "invokes" --> easy_rec_python_model_easy_rec_estimator
    easy_rec_python_model_easy_rec_estimator -- "invokes" --> easy_rec_python_builders_optimizer_builder
    easy_rec_python_model_easy_rec_estimator -- "invokes" --> easy_rec_python_builders_loss_builder
    easy_rec_python_model_easy_rec_estimator -- "integrates" --> easy_rec_python_compat_sync_replicas_optimizer
    easy_rec_python_model_easy_rec_estimator -- "utilizes" --> easy_rec_python_compat_early_stopping
    easy_rec_python_builders_optimizer_builder -- "utilizes" --> easy_rec_python_compat_optimizers
    easy_rec_python_builders_loss_builder -- "builds instances from" --> easy_rec_python_loss
Loading

CodeBoardingDemoContact

Details

The easy_rec.python subsystem orchestrates the training and evaluation of machine learning models within the EasyRec framework. The easy_rec.python.main component serves as the primary entry point, initiating the training process by invoking the easy_rec.python.model.easy_rec_estimator. This estimator, a core TensorFlow component, manages the entire training lifecycle, including model training, evaluation, and export. It dynamically configures optimizers and loss functions through interactions with easy_rec.python.builders.optimizer_builder and easy_rec.python.builders.loss_builder, respectively. These builders abstract the creation of various optimizers and loss functions, leveraging compatibility utilities provided by easy_rec.python.compat.optimizers and concrete loss implementations from easy_rec.python.loss. For distributed training, the easy_rec_estimator integrates with easy_rec.python.compat.sync_replicas_optimizer to ensure synchronized gradient updates. Additionally, it utilizes easy_rec.python.compat.early_stopping to prevent overfitting and optimize training duration. This structured interaction ensures a robust and flexible training pipeline.

easy_rec.python.main

Primary entry point for initiating training and evaluation tasks, orchestrating the training environment. It sets up the overall training flow.

Related Classes/Methods:

easy_rec.python.model.easy_rec_estimator

Core TensorFlow Estimator encapsulating the model's training loop, evaluation, and export, managing optimizers, loss, and distributed training. This is the heart of the training process.

Related Classes/Methods:

easy_rec.python.builders.optimizer_builder

Responsible for constructing and configuring various TensorFlow optimizers based on the provided training configuration. It abstracts the creation of optimizers.

Related Classes/Methods:

easy_rec.python.builders.loss_builder

Responsible for constructing and configuring different loss functions based on the training configuration. It abstracts the creation of loss functions.

Related Classes/Methods:

easy_rec.python.loss

Contains concrete implementations of various loss functions used during training. These are the actual mathematical functions for calculating training error.

Related Classes/Methods:

easy_rec.python.compat.optimizers

Provides compatibility utilities and wrappers for TensorFlow optimizers, including gradient manipulation functionalities. It ensures optimizers work correctly across different TensorFlow versions or configurations.

Related Classes/Methods:

easy_rec.python.compat.sync_replicas_optimizer

An optimizer wrapper specifically designed for synchronized distributed training, ensuring proper gradient aggregation across multiple training replicas. Crucial for large-scale distributed training.

Related Classes/Methods:

easy_rec.python.compat.early_stopping

Implements logic for early stopping during training to prevent overfitting and optimize training duration by monitoring performance metrics.

Related Classes/Methods: