Table of Contents

Interface IGradientCache<T>

Namespace
AiDotNet.Interfaces
Assembly
AiDotNet.dll

Defines an interface for storing and retrieving pre-computed gradients to improve performance in machine learning models.

public interface IGradientCache<T>

Type Parameters

T

The numeric type used for calculations (e.g., double, float).

Remarks

For Beginners: This interface defines methods for saving and reusing calculations to make your AI models run faster.

In machine learning, models often need to calculate "gradients" - mathematical directions that show how to adjust the model to make better predictions. These calculations can be time-consuming, especially for complex models.

Think of a gradient cache like a notebook where you write down answers to difficult math problems:

  • When you solve a problem, you write down the answer in your notebook
  • Later, if you need the same answer again, you can just look it up instead of re-solving the problem
  • This saves you time and effort

The gradient cache works the same way:

  • When a gradient is calculated, it's stored with a unique name (the "key")
  • If the same gradient is needed again, it can be retrieved using that name
  • This avoids repeating expensive calculations

This is especially useful for:

  • Complex models with many parameters
  • Models that use the same calculations repeatedly
  • Training scenarios where speed is important

Methods

CacheGradient(string, IGradientModel<T>)

Stores a gradient in the cache with a unique key for later retrieval.

void CacheGradient(string key, IGradientModel<T> gradient)

Parameters

key string

The unique identifier to associate with this gradient.

gradient IGradientModel<T>

The gradient to cache.

Remarks

For Beginners: This method saves a calculation result so you can reuse it later.

The parameters:

  • key: A unique name or identifier you choose for this gradient (like naming a file or labeling a page in your notebook)
  • gradient: The actual gradient calculation result you want to save

What this method does:

  1. Takes the gradient you've calculated
  2. Stores it in memory using the key as its identifier
  3. Makes it available for quick retrieval later

It's important to use meaningful, consistent keys so you can easily find your cached gradients later. For example, you might use keys like:

  • "layer1_weights_gradient"
  • "output_bias_gradient"

If you cache a gradient with a key that already exists in the cache, the new gradient typically replaces the old one.

ClearCache()

Removes all cached gradients, freeing up memory.

void ClearCache()

Remarks

For Beginners: This method erases all saved calculations from the cache.

What this method does:

  1. Removes all gradients that have been stored in the cache
  2. Frees up the memory they were using
  3. Essentially gives you a "clean slate"

You might want to clear the cache when:

  • You've made significant changes to your model
  • You're starting a new phase of training
  • You suspect the cached gradients might be outdated
  • You need to free up memory
  • You want to measure performance without caching

After clearing the cache, any attempt to retrieve a gradient will return null until new gradients are cached.

GetCachedGradient(string)

Retrieves a previously cached gradient using its unique key.

IGradientModel<T>? GetCachedGradient(string key)

Parameters

key string

The unique identifier for the gradient.

Returns

IGradientModel<T>

The cached gradient if found; otherwise, null.

Remarks

For Beginners: This method looks up a saved calculation in the cache.

The parameter:

  • key: A unique name or identifier for the gradient you want to retrieve (like a file name or a label in your notebook)

What this method does:

  1. Checks if a gradient with the given key exists in the cache
  2. If found, returns the pre-computed gradient
  3. If not found, returns null (meaning you'll need to calculate it from scratch)

The returned ISymbolicModel represents the gradient in a form that can be:

  • Evaluated with different input values
  • Combined with other mathematical expressions
  • Used directly in optimization algorithms

Using cached gradients can significantly speed up training because you avoid repeating the same complex calculations multiple times.

Edit this page