wtf is 2c2t.drt?

2C2T.DRT is a Python framework for training neural networks on a regular computer's processor instead of an expensive graphics card, since GPUs suited for deep learning have become costly and hard to find. It can optionally use an integrated graphics chip through OpenCL for some of the heaviest matrix math, but a dedicated GPU is never required. The README is upfront about the tradeoff: this approach is roughly 15 to 30 times slower than an entry-level GPU, and it recommends people who already own a GPU just use PyTorch instead. It includes real measured benchmarks on a 16-core CPU, showing that training one pass over the well-known MNIST dataset with a small model takes around 47 seconds, and 10 full passes take about 8 minutes. Model size is limited mainly by how much RAM your computer has rather than GPU memory, and the project states it can handle roughly 50 to 200 million parameters on a typical consumer machine. It only depends on NumPy, using its OpenBLAS backend for fast matrix operations, and is installed by cloning the repository and running a Python script. It provides a set of building blocks similar to other deep learning libraries: layers like Dense and Conv2D, a Transformer attention layer, normalization layers, several optimizers such as Adam and SGD, and multiple loss functions. It also includes memory-saving techniques for training larger models on limited RAM, including gradient accumulation to simulate bigger batches, gradient checkpointing that recomputes parts of the network during training to save memory, splitting models across disk when they exceed RAM, and 8 or 16 bit weight compression. Several ready-made model sizes are provided through command line flags, from a small multi-layer model needing very little RAM up to a large convolutional model needing significantly more. The project also ships a benchmark script and command-line options for evaluation-only runs and loading saved model weights. The project is licensed under the MIT License.