Optimized_Einstein_Ops

Improved performance from existing einops repo. Tensor operations made 150 percent faster.

🌌 Custom Rearrange Function

This implementation is inspired from einops but takes it further with optimized performance, smart caching, and user-friendly error handling and designed to simplify tensor manipulation with speed.

---

✨ Benefits

• Pattern-Based Flexibility: Transform tensors using readable patterns like 'h w c -> (h w) c' or 'b c h w -> b (h w) c'.
• Broadcasting Magic: Add new dimensions or expand existing ones without breaking a sweat.
• Performance Boost: Powered by lru_cache, it caches operations for fast repeated calls.
• Helpful Debugging: Custom EinopsError exceptions provide clear, actionable feedback when something goes wrong.
• Unique Design: Improved performance and usability advantages.

---

🛠️ How Does It Work?

The rearrange function is a tensor-transforming wizard that operates in four seamless steps:

1. Pattern Parsing

   • Takes a pattern like 'h w c -> w h c' and splits it into input axes (h, w, c) and output axes (w, h, c).
   • Understands grouped axes like (h w) for advanced reshaping.

2. Custom Axis Size

   • Calculates the size of each axis based on the tensor’s shape.
   • Supports custom sizes via an optional axes_lengths argument (e.g., h=32, w=32).

3. Tensor Transformation

   • Splits: Breaks down grouped axes (e.g., (h w) into a single dimension).
   • Permutes: Reorders axes to match the output pattern.
   • Broadcasts: Expands or adds dimensions as needed.
   • Reshapes: Delivers the final tensor in the desired shape.

4. Caching

   • Stores parsed patterns and operations in memory, so repeated calls are fast.

---

🚀 Get Started

Step 1: Open Colab

Step 2: Choose Tesla T4 GPU

Step 3: Run Cells