PyTorch Developer Day 2020
2020-11-12
-- online --
who it's for
PyTorch core developers, machine learning engineers, research engineers, infrastructure engineers, mobile engineers, natural language processing engineers, audio engineers, computer vision engineers
topics
complex64 and complex128 with FFT; NumPy compatibility and operator changes; autograd jacobian vjp jvp hessian APIs; distributed RPC and remote references; distributed data parallel communication hooks; torchaudio datasets and JIT modules; torchtext C++ transforms and attention container; TorchVision TorchScript and ONNX export; profiler runtime and memory tracing; TensorBoard SummaryWriter integration; automatic mixed precision and TF32 optimizations; PyTorch on Windows multi GPU DDP
| title | author | length |
|---|---|---|
| Complex Numbers in PyTorch | Anjali Chourdia | 07m |
PyTorch adds native complex dtypes complex64 and complex128 and exposes spectral routines via torch.fft; linear algebra and elementwise kernels now support… | ||
| Making PyTorch More "NumPy Compatible" | Mike Ruberry | 06m |
PyTorch 1.7 implements many NumPy functions and changes specific behaviors to match NumPy (for example division now does true division), adds torch.quantile… | ||
| High-Level API for Autograd | Alban Desmaison | 09m |
New high-level API lives under torch.autograd.fun and exposes concrete functions: jacobian, vector-jacobian product (vjp) and jacobian-vector product (jvp) for… | ||
| PyTorch Distributed RPC | Shen Li | 11m |
Covers the PyTorch RPC package with concrete APIs: rpc.sync, rpc.async, remote for running callables remotely, remote references for zero-copy pointers between… | ||
| PyTorch Distributed Data Parallel (DDP) | Pritam Damania | 10m |
Overview of concrete DDP and c10d changes: DDP communication hook lets Python callables replace the synchronized all-reduce (useful for fp16 compression or… | ||
| TorchAudio | Vincent Quenneville-Belair | 06m |
Library that reads mp3 wav flac sphere files, loads datasets like LibriSpeech and LibriTTS, and exposes tensors via torchaudio.load for GPU/JIT use. It… | ||
| TorchText | George Zhang | 08m |
TorchText provides reusable NLP building blocks, new datasets (beta/nightly), and C++-backed transforms that convert raw text to tensors while supporting… | ||
| TorchVision | Francisco Massa | 06m |
TorchVision now lets teams export complete pipelines to C++: TorchScript and ONNX export for all pre-trained models and mixed-precision inference work out of… | ||
| Latest Profiler APIs and Best Practices | Ilia Cherniasvskii | 03m |
PyTorch Profiler measures runtimes and tensor memory per operator and now supports record_shapes true to group by input shapes, profile_memory true to report… | ||
| Visualization with TensorBoard | Siqi Yan | 05m |
PyTorch exposes native TensorBoard support since PyTorch 1.2 and uses a SummaryWriter that logs to local or remote directories (S3) so training curves, QPS and… | ||
| PyTorch Performance | Natalia Gimelshein | 10m |
PyTorch reports concrete speedups and tools: tf32 on Ampere uses 32-bit storage with 10-bit mantissa reads and shows about 3x speedup on huge phase roberto… | ||
| PyTorch on Windows | Maxim Lukiyanov | 12m |
PyTorch 1.7 on Windows adds distributed data parallel training with multi‑GPU support (demo runs on an Azure VM with two NVIDIA V100s and shows multiple Python… | ||
| PyTorch Mobile | David Reiss | 08m |
PyTorch Mobile reports concrete wins: a MobileNetV2 FP model drops from ~250 ms to under 15 ms (over 10×) since 1.3, a caching allocator yields 5–20% speedups… | ||
| PyTorch Mobile and Android Neural Networks API | Oli Gaymond | 06m |
PyTorch mobile adds direct offload to the Android Neural Networks API, supporting over 100 ops on Android 11 (including LSTM, float and quantized types, and… | ||
| nothing matches | ||