WhisperX: Batched Whisper, wav2vec2 Alignment and Diarization From the Source Up
WhisperX is one of those repositories that looks small from the outside — a Python package and a CLI called whisperx — but underneath it coordinates three very different machine learning models against the same audio buffer. A faster-whisper CTranslate2...
TensorFlow Serving: A Source Tour of Google's Model Server
TensorFlow Serving is Google’s open-source model server: the piece of infrastructure that takes a trained TensorFlow model off a researcher’s laptop and makes it answer real network traffic, reliably, at low latency. You export a SavedModel into a versioned directory,...
Swin Transformer: A Source Tour of Microsoft's Hierarchical Vision Backbone
When the Swin Transformer landed in 2021, most vision transformers were still computing attention over every pair of tokens in the image — quadratic cost, single-scale feature maps, and an awkward fit for dense prediction tasks that expect pyramid-shaped backbones....
pyannote.audio source tour: the PyTorch core behind speaker diarization
If you have ever asked “who spoke when” about a meeting recording, there is a good chance pyannote.audio did the heavy lifting. It is the open-source Python toolkit behind many production diarization systems, a long-running entry in the Hugging Face...
Piper: A Tiny C++ Core That Speaks for Itself
Piper is a fast, fully local neural text-to-speech system from the Rhasspy project, the same community behind popular offline voice tooling for home automation. You hand it a sentence on standard input, it hands you a WAV file a fraction...
MediaPipe: A Source Tour of Google's On-Device ML Framework
If you have ever pointed a phone camera at your own hand and watched a wireframe skeleton snap to your knuckles in real time, you have almost certainly met MediaPipe. The framework started life inside Google as a C++ engine...
GPT-2 at the Source: OpenAI's Original Next-Token Predictor, Line by Line
In February 2019, OpenAI published a short paper with a long shadow: “Language Models are Unsupervised Multitask Learners.” The model it described, GPT-2, was trained on WebText, a large scrape of outbound Reddit links, and it demonstrated something uncanny for...
Detectron2: A Source Tour of Meta's Detection Platform
Detectron2 is Meta AI Research’s next-generation library for object detection and segmentation, and it is the direct successor of the original Detectron and of maskrcnn-benchmark. Written in PyTorch, it has carried a whole generation of computer vision research: Faster and...
Hugging Face Datasets Source Tour: From load_dataset to Arrow Shards
If you have trained a model on a public corpus in the last five years, the odds are very high that the data reached you through huggingface/datasets. The library is the de facto data layer of the open-source model ecosystem:...
MosaicML Composer: A Source Tour of Event-Driven Speedup Training
MosaicML composer is an open-source PyTorch training library built around one idea: if you can observe training as a sequence of well-defined moments, you can intervene at those moments to make models train faster without changing the math of your...
YOLOv5: A Source Tour of Ultralytics' Real-Time Object Detector
When YOLOv5 appeared in mid-2020, it was slightly scandalous: a “YOLOv5” with no paper behind it, built directly in PyTorch while YOLOv4 was still a Darknet project. Four years of production use later, the Ultralytics/yolov5 repository has become one of...
timm: A Source Tour of PyTorch Image Models
PyTorch Image Models, known to everyone simply as timm, is the library you reach for when you need a vision backbone that actually works. Started and still curated by Ross Wightman, now hosted under the Hugging Face organization, it packs...
Stable Diffusion: A Source Tour of the Original Latent Diffusion Code
Every image generator you have used in the last three years — the browser apps, the commercial APIs, the open-source forks — descends from one small research repository: CompVis/stable-diffusion. In August 2022 it shipped the weights and the reference code...
SentencePiece: A Source Tour of Google's Language-Agnostic Tokenizer
Ask a Llama checkpoint what “tokenization” means and it will answer in the dialect of SentencePiece. Google’s tokenizer library — a compact C++ core with Python bindings — has quietly become the substrate under an enormous slice of modern NLP:...
Segment Anything: A Source Tour of Meta's Promptable Vision Model
The Segment Anything Model (SAM) is Meta AI’s answer to a simple but stubborn question: can a vision model segment any object in any image, given only a hint about what to look for? Published by the facebookresearch team in...
PaddleOCR: A Source Tour of the Multilingual OCR Toolkit
Optical character recognition looks deceptively simple from the outside — an image goes in, a string comes out — but the systems that make it work reliably are chains of specialized neural networks. PaddleOCR, the OCR toolkit from Baidu’s PaddlePaddle...
MNN: A Source Tour of Alibaba's Mobile Inference Engine
If you have installed a shopping app from Alibaba in the last decade, odds are good that MNN ran a piece of it. MNN is the lightweight inference engine behind Taobao, Tmall, Youku, DingTalk, and dozens of other apps, covering...
Megatron-LM: A Source Tour of NVIDIA's Large-Scale Training Framework
Megatron-LM is NVIDIA’s reference implementation for training large language models on clusters of GPUs, and it is the codebase behind many of the record-setting scaling results you see in model announcements. The repository actually ships two things: Megatron-LM, a set...
Llama 2: A Source Tour of Meta's Reference LLM Inference Code
When Meta released Llama 2 in July 2023, the weights grabbed the headlines, but the small Python package that shipped alongside them — meta-llama/llama — quietly became one of the most consequential pieces of inference code in modern machine learning....
BLIP: A Source Tour of Salesforce's Vision-Language Pretraining Model
Most multimodal papers describe their models with a diagram and leave the wiring to your imagination. BLIP — Bootstrapping Language-Image Pre-training, released by Salesforce AI Research in 2022 — is nicer than average: the whole system lives in a handful...
PyTorch/XLA: Inside the Bridge from Eager PyTorch to TPU Hardware
Most deep learning engineers meet PyTorch as an eager framework: every op fires a kernel the moment you call it. On a TPU, that model of execution does not work. TPUs are fed by a compiler — XLA, the Accelerated...
TFLite Micro: How Neural Network Inference Fits in 256 KB Without malloc
Every inference runtime we have toured in this series — TensorFlow, ONNX Runtime, llama.cpp — assumes a fat host: gigabytes of RAM, an allocator that can hand out memory at will, maybe a GPU on the side. The repo under...
TensorRT Source Tour: From ONNX Graph to Serialized CUDA Engine
Most deep-learning frameworks you meet day to day are executors: they take a graph, schedule its ops, and hope the kernels underneath are fast. TensorRT is a different animal. It is a compiler: it ingests a trained network, picks concrete...
TensorRT Model Optimizer: A Source Tour of NVIDIA's Model Compression Toolkit
Most inference stacks spend their lives downstream of a model. NVIDIA’s Model Optimizer — ModelOpt for short — lives upstream, where the model itself is made smaller. The repository, NVIDIA/TensorRT-Model-Optimizer on GitHub, is the source behind the nvidia-modelopt PyPI package:...