Table of Contents

Class TensorCompressionMetadata<T>

Namespace
AiDotNet.ModelCompression
Assembly
AiDotNet.dll

Metadata for N-dimensional tensor compression operations that wraps the underlying vector compression metadata.

public class TensorCompressionMetadata<T> : ICompressionMetadata<T>

Type Parameters

T

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

Inheritance
TensorCompressionMetadata<T>
Implements
Inherited Members

Remarks

TensorCompressionMetadata stores the information needed to decompress an N-dimensional weight tensor that was compressed by first flattening it to a vector. It preserves the original tensor shape (dimensions) and delegates the actual compression metadata to an inner ICompressionMetadata instance.

For Beginners: Tensors are multi-dimensional arrays used extensively in deep learning:

  • 1D tensor (vector): [100] - like a bias term with 100 values
  • 2D tensor (matrix): [100, 50] - like fully connected layer weights
  • 3D tensor: [32, 100, 50] - like a batch of 32 matrices
  • 4D tensor: [64, 3, 3, 3] - like 64 convolutional filters with 3 channels, 3x3 kernels

When compressing a tensor, we flatten it to a 1D array, compress it, and need to remember the original shape to reconstruct it. This metadata stores:

  1. The original shape - the dimensions of the tensor (e.g., [64, 3, 3, 3])
  2. The inner compression details - how the flattened data was compressed

Think of it like packing a Rubik's cube for shipping:

  • You disassemble it into individual pieces (flatten)
  • You put the pieces in a compressed bag (apply compression)
  • You include assembly instructions with the dimensions (this metadata)
  • Later, you can perfectly reconstruct the original cube

Constructors

TensorCompressionMetadata(int[], ICompressionMetadata<T>)

Initializes a new instance of the TensorCompressionMetadata class.

public TensorCompressionMetadata(int[] originalShape, ICompressionMetadata<T> innerMetadata)

Parameters

originalShape int[]

The shape (dimensions) of the original tensor.

innerMetadata ICompressionMetadata<T>

The compression metadata from the underlying vector compression.

Remarks

For Beginners: When creating tensor compression metadata, you specify:

  • originalShape: The dimensions of the tensor (e.g., [64, 3, 3, 3] for conv filters)
  • innerMetadata: The details of how the flattened weights were compressed

The shape array is copied to prevent external modifications from affecting the metadata.

Exceptions

ArgumentNullException

Thrown when originalShape or innerMetadata is null.

ArgumentException

Thrown when originalShape is empty or contains non-positive dimensions.

Properties

InnerMetadata

Gets the inner compression metadata from the underlying vector compression algorithm.

public ICompressionMetadata<T> InnerMetadata { get; }

Property Value

ICompressionMetadata<T>

Remarks

For Beginners: This contains the actual compression details from whatever algorithm was used (weight clustering, Huffman encoding, pruning, etc.). The tensor compression is just a wrapper that remembers the shape; the real compression work is tracked here.

OriginalLength

Gets the original total number of elements in the flattened tensor.

public int OriginalLength { get; }

Property Value

int

Remarks

For Beginners: This is the total count of weight values in the original tensor, calculated by multiplying all dimensions together. For example:

  • A [64, 3, 3, 3] tensor has 64 * 3 * 3 * 3 = 1,728 elements
  • A [100, 50] tensor has 100 * 50 = 5,000 elements

This is needed to allocate the right amount of memory when decompressing.

OriginalShape

Gets a copy of the original shape (dimensions) of the tensor.

public int[] OriginalShape { get; }

Property Value

int[]

Remarks

For Beginners: The shape array describes the structure of the original tensor:

  • [100] means a 1D tensor with 100 elements
  • [100, 50] means a 2D tensor (matrix) with 100 rows and 50 columns
  • [64, 3, 3, 3] means a 4D tensor (common for conv filters)

This shape is essential for reshaping the decompressed data back into the correct tensor format. A defensive copy is returned to preserve immutability.

Rank

Gets the number of dimensions in the original tensor.

public int Rank { get; }

Property Value

int

Remarks

For Beginners: This tells you how many dimensions the tensor has:

  • 1 for vectors
  • 2 for matrices
  • 3+ for higher-dimensional tensors (common in deep learning)

Type

Gets the compression type from the underlying compression algorithm.

public CompressionType Type { get; }

Property Value

CompressionType

Remarks

For Beginners: This returns the actual compression algorithm type (like WeightClustering or HuffmanEncoding) that was used to compress the flattened tensor data. The tensor metadata itself is just a shape container - it delegates to the inner metadata for the real compression type.

Methods

GetMetadataSize()

Gets the total size in bytes of this metadata structure, including the inner metadata.

public long GetMetadataSize()

Returns

long

The metadata size in bytes.

Remarks

For Beginners: When calculating the total compressed size, you need to include the metadata overhead. This method calculates:

  • The size of the rank (1 integer = 4 bytes, to know how many dimensions to read)
  • The size of the shape array (rank integers, each 4 bytes)
  • Plus the size of the inner compression metadata

For example, a 4D tensor adds 4 + 4*4 = 20 bytes of shape overhead.

Edit this page