MiniGPTv2<T>

Models & Types · AiDotNet.VisionLanguage.InstructionTuned

MiniGPT-v2: unified interface for multi-task VL learning with task-specific tokens.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → MiniGPTv2<T>

Implements: IInstructionTunedVLM<T>

For Beginners

MiniGPT-v2 upgrades MiniGPT-4 to handle multiple visual tasks in a single model using a clever trick: task-specific identifier tokens. Instead of needing separate models for different tasks, you prepend a special token that tells the model what type of task you want — like “[vqa]” for visual question answering, “[caption]” for image captioning, or “[grounding]” for locating objects. The model learns to produce the right kind of output based on which task token it sees. It also upgrades the language backbone from Vicuna to LLaMA-2. Default values follow the original paper settings.

How It Works

MiniGPT-v2 (Chen et al., 2023) upgrades MiniGPT-4 with LLaMA-2 backbone and a unified multi-task learning framework using task-specific identifier tokens. It supports visual question answering, image captioning, visual grounding, and referring expression comprehension through a single model with task-prefixed instructions.

References:

  • Paper: “MiniGPT-v2: Large Language Model as a Unified Interface for Vision-Language Multi-task Learning” (Chen et al., 2023)

Example

var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
var trainModel = new MiniGPTv2<double>(architecture, new MiniGPTv2Options());

Constructors

Constructor Summary
MiniGPTv2(NeuralNetworkArchitecture<T> architecture, MiniGPTv2Options options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
MiniGPTv2(NeuralNetworkArchitecture<T> architecture, string modelPath, MiniGPTv2Options options)

Properties

Property Summary
int DecoderEmbeddingDim { get; }
int EmbeddingDimension { get; }
int MaxGenerationLength { get; }
string LanguageModelName { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Chat(Tensor<T> image, IEnumerable<ValueTuple<string, string>> conversationHistory, string userMessage)
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> GenerateFromImage(Tensor<T> image, string prompt) Generates text using MiniGPT-v2’s unified multi-task Q-Former architecture.