Description

This PR adds a comprehensive Vision Transformer (ViT) implementation to the computer_vision folder for image classification tasks.

Implementation Details

Implementation of the Vision Transformer architecture from "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale" (Dosovitskiy et al., 2020).

Core Components:

• ✅ Patch Embedding: Splits images into non-overlapping patches
• ✅ Linear Projection: Projects flattened patches to embedding dimension
• ✅ Positional Encoding: Adds learnable positional embeddings with CLS token
• ✅ Attention Mechanism: Scaled dot-product attention
• ✅ Layer Normalization: Normalizes layer outputs
• ✅ Feed-Forward Network: Position-wise FFN with GELU activation
• ✅ Transformer Encoder Block: Complete encoder with multi-head attention
• ✅ Vision Transformer Pipeline: Full ViT for image classification

Code Quality:

• ✅ Comprehensive docstrings with explanations
• ✅ Type hints for all parameters and return values
• ✅ Doctests for validation of each function
• ✅ Example usage in __main__ block
• ✅ Configurable parameters (patch size, embedding dim, layers, etc.)
• ✅ Educational comments explaining the architecture
• ✅ Follows repository coding standards

Example Usage:

from computer_vision.vision_transformer import vision_transformer
import numpy as np

image = np.random.rand(224, 224, 3)
logits = vision_transformer(image, num_classes=1000)
predicted_class = np.argmax(logits)