PyTorch Developer Conference 2019
2019-10-08..2019-10-11
San Francisco, United States
| title | author | length |
|---|---|---|
| PyTorch at Tesla | Andrej Karpathy | 11m |
Tesla uses PyTorch to train vision-first networks on raw video from eight cameras, running large multihead “HydraNets” with shared ResNet-50 backbones and a… | ||
| Detectron2 - Next Gen Object Detection Library | Yuxin Wu | 10m |
Detectron2 is a ground-up PyTorch rewrite of Detectron that implements Mask R-CNN, panoptic and instance segmentation, TensorMask and other models, claims 2–3x… | ||
| Privacy Preserving AI | ANDREW TRASK | 16m |
Claims it is possible to answer questions using data you cannot see by combining PySyft, PyTorch remote execution and cryptography: pointer tensors let code… | ||
| Dataloader Design for PyTorch | Tongzhou Wang | 07m |
PyTorch DataLoader now lets datasets return whole batches and supports iterable streams so data loading moves from one-index-per-call to either batch-style… | ||
| PyTorch Developer Conference 2019 | Full Livestream | Mike Schroepfer, Lin Qiao, Gregory Chanan, Michael Suo, DMYTRO DZHULGAKOV, Yisong Yue, Yuhao Zhang, Robert Stojnic, Sasha Rush, ANDREW TRASK, Laurens van der Maaten, Sidney Zhang, Andrej Karpathy, Saurabh Tiwary, Vivek Kumar | 04h 52m |
Watch the full set of talks from the 2019 PyTorch Developer Conference. Deep dive on PyTorch 1.3 and new tools and libraries including PyTorch Mobile, CrypTen… | ||
| Research to Production: PyTorch JIT/TorchScript Updates | Michael Suo | 10m |
TorchScript serializes PyTorch models to run without Python and the JIT compiles and optimizes them for production, released in PyTorch 1.3. A single call… | ||
| Quantization | DMYTRO DZHULGAKOV | 10m |
PyTorch 1.3 adds eager-mode 8-bit quantization API with three workflows: dynamic one-line quantization (weights → int8, activations float), post-training… | ||
| Model Interpretability with Captum | Narine Kokhilkyan | 08m |
Library called Captum provides PyTorch gradient‑based attributions including integrated gradients, DeepLIFT, SmoothGrad and VarGrad and computes attributions… | ||
| PyTorch ONNX Export Support | Lara Haidar | 05m |
PyTorch ONNX exporter converts traced torch models by translating torch IR to an ONNX IR and adds concrete features: missing operator export and custom… | ||
| PyTorch Mobile | David Reiss | 09m |
PyTorch Mobile provides an end-to-end path from Python to iOS and Android by running TorchScript interpreter models without a Python runtime, supporting traced… | ||
| PyTorch in Robotics | Yisong Yue | 10m |
Caltech researchers build a neural lander that learns the missing ground‑effect force as a residual dynamics term using least‑squares regression and a small… | ||
| PyTorch at Uber | Sidney Zhang | 13m |
Uber ATG describes a single, joint perception–prediction neural network that consumes multi‑sensor time series and HD maps and outputs multiple future… | ||
| Speech Extensions to Fairseq | Dmytro Okhonko | 07m |
Fairseq now includes extensions that add end-to-end speech and audio recognition pipelines: dataset readers and feature extraction, encoder/decoder modules… | ||
| What’s new in PyTorch 1.3 | Lin Qiao | 12m |
PyTorch 1.3 adds explicit named tensors to attach semantics to tensor dims, a TorchScript-based mobile workflow that enables 60 FPS MobileNet on iPhone for… | ||
| Keynote | Mike Schroepfer | 13m |
PyTorch 1.3 announces roughly 1,200 contributors (about 50% YOY growth) and a push from research into production at Facebook where most models now train on… | ||
| PyTorch at Microsoft | Saurabh Tiwary | 07m |
PyTorch becomes the production DL stack for training at scale with native Windows support, ONNX for runtime acceleration, and model/data parallelism to handle… | ||
| PyTorch on Google Cloud TPUs | Brian McCann | 23m |
PyTorch gains official Google Cloud TPU support in version 1.3 via torch_xla, using a lazy tensor architecture that records graphs and compiles them with XLA… | ||
| PyTorch at Dolby Labs | Vivek Kumar | 10m |
Dolby uses PyTorch to build audio-first models (patches get merged in days) and focuses on architectures designed for audio rather than spectrogram shortcuts… | ||
| StanfordNLP | Yuhao Zhang | 11m |
StanfordNLP is a native Python library built on PyTorch that replaces rule‑based tokenizers with a lightweight character LSTM tokenizer and an end‑to‑end… | ||
| Collaborative Natural Language Inference | Sasha Rush | 11m |
Proposes building models that natively support human–model collaboration and counterfactual feedback, combining probabilistic control with pretrained… | ||
| Panel Discussion | Soumith Chintala, Lisha Li, Michela Paganini, Clément Delangue, Rachel Thomas, Maria Popova | 37m |
Panelists report concrete workflows and pain points: teams move development and testing into Jupyter notebooks for development, production runs use Python and… | ||
| Linear Algebra in PyTorch | Vishwak Srinivasan | 04m |
PyTorch 1.2–1.3 adds native batching so batched QR, LU and other decompositions run as single ops instead of Python loops, and most linear-algebra ops are… | ||
| Apex | MICHAEL CARILLI | 05m |
Apex is an open-source PyTorch extension that uses mixed precision (FP16 for compute, FP32 for accumulators) to exploit NVIDIA tensor cores and deliver large… | ||
| PyTorch Front-End Features: Named Tensors and Type Promotion | Gregory Chanan | 08m |
PyTorch 1.3 introduces named tensors that attach dimension names (e.g., N C H W) so code can sum and align dimensions by name instead of position, enabling a… | ||
| CrypTen | Laurens van der Maaten | 11m |
CrypTen is an open-source PyTorch extension that implements secure multi-party computation with encrypted tensors that behave like regular tensors, real… | ||
| PyTorch Summer Hackathon Winners | Joe Spisak, Sebastien Arnold, Tristan Deleu | 11m |
Menlo Park hosts a 28‑hour hack with 74 participants; the global event runs five weeks and attracts about 1,500 submissions. First‑place build is a usable… | ||
| Recap | 03m | |
Conference recap highlights concrete releases and uses: torchscript runs full prediction models in production at Microsoft and powers an Uber deployment that… | ||
| Sotabench for Reproducible Research | Robert Stojnic | 11m |
Sotabench benchmarks open source models by connecting repositories to continuous integration with GitHub so each commit can run reproducible tests; add one… | ||
| nothing matches | ||