Recap on ML

A log of errors and solutions encountered building a SegFormer PoC notebook, covering label-count mismatches, CUDA out-of-memory errors, device-map issues, and collator bugs during training.

Describes the SegFormer setup for semantic segmentation fine-tuning on the MIT ADE20k scene-parsing dataset, detailing the Trainer execution order from on-the-fly transforms through metric computation.

Introduction to a series on quantizing the pre-trained SegFormer vision transformer, outlining the goal of reducing model size via 8-bit/4-bit and custom quantization using HuggingFace, PyTorch, and bitsandbytes.

A reference post on Python Futures — the concurrent.futures abstraction for managing asynchronous execution of callables across threads and processes.

A structured look at Transformer trade-offs, examining quadratic time/space complexity, O(1) path length, parallelization advantages, and the role of transfer learning and pre-training.

A comparison of Python's matmul operator and NumPy/PyTorch einsum for tensor contraction, covering theoretical background and practical case studies relevant to ML workloads.