plimai 0.1.3

Vision LLMs with LoRA fine-tuning.

PlimAI: Vision LLMs with Efficient LoRA Fine-Tuning

---

PlimAI is a modular, research-friendly framework for building and fine-tuning Vision Large Language Models (LLMs) with efficient Low-Rank Adaptation (LoRA) support. It is designed for:

• Researchers exploring new vision transformer architectures or fine-tuning strategies
• Practitioners who want to adapt large vision models to custom datasets with limited compute
• Developers looking for a clean, extensible codebase for vision-language AI

PlimAI provides a plug-and-play interface for LoRA, making it easy to experiment with parameter-efficient fine-tuning. The codebase is modular, so you can swap out or extend components like patch embedding, attention, or MLP heads.

---

🏗️ Architecture Overview

PlimAI is built around a modular Vision Transformer (ViT) backbone, with LoRA adapters injected into attention and MLP layers for efficient fine-tuning. The main components are:

graph TD
    A[Input Image] --> B[Patch Embedding]
    B --> C[+CLS Token & Positional Encoding]
    C --> D[Transformer Encoder]
    D --> E[LayerNorm]
    E --> F[MLP Head]
    F --> G[Output (e.g., Class logits)]
    subgraph LoRA Adapters
        D
    end

Main Modules

• PatchEmbedding: Splits the image into patches and projects them into embedding space.
• TransformerEncoder: Stack of transformer layers, each with multi-head self-attention and MLP blocks. LoRA adapters can be injected here.
• LoRALinear: Low-rank adapters for efficient fine-tuning, only a small number of parameters are updated.
• MLPHead: Final classification or regression head.
• Config & Utils: Easy configuration and preprocessing utilities.

---

📦 Installation

pip install plimai

Or, for the latest version from source:

git clone https://github.com/plim-ai/plim.git
cd plim
pip install .

---

🧑‍💻 Quick Start

import torch
from plimai.models.vision_transformer import VisionTransformer
from plimai.utils.config import default_config

# Dummy image batch: batch_size=2, channels=3, height=224, width=224
x = torch.randn(2, 3, 224, 224)
model = VisionTransformer(
    img_size=default_config['img_size'],
    patch_size=default_config['patch_size'],
    in_chans=default_config['in_chans'],
    num_classes=default_config['num_classes'],
    embed_dim=default_config['embed_dim'],
    depth=default_config['depth'],
    num_heads=default_config['num_heads'],
    mlp_ratio=default_config['mlp_ratio'],
    lora_config=default_config['lora'],
)
out = model(x)
print('Output shape:', out.shape)

---

📚 Documentation

• API Reference
• Vision Transformer with LoRA: Paper
• LoRA for Vision Models: HuggingFace PEFT

---

🧩 Module Breakdown

Module	Description
PatchEmbedding	Converts images to patch embeddings for transformer input
TransformerEncoder	Stack of transformer layers with optional LoRA adapters
LoRALinear	Low-rank adapters for parameter-efficient fine-tuning
MLPHead	Output head for classification or regression
data.py	Preprocessing and augmentation utilities
config.py	Centralized configuration for model/training hyperparameters

---

🧪 Running Tests

pytest tests/

---

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

• Open issues for bugs or feature requests
• Submit pull requests for improvements
• Star ⭐ the repo if you find it useful!

---

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

---

🌟 Acknowledgements

• PyTorch
• HuggingFace
• LoRA: Low-Rank Adaptation of Large Language Models

Directory Structure

plimai/
  models/
    vision_transformer.py
    lora.py
  components/
    patch_embedding.py
    attention.py
    mlp.py
  utils/
    data.py
    config.py
  example.py