The new kid on the block: deep learning with GluonNLP
summary
Demonstrates GluonNLP on MXNet with concrete utilities: a sorted bucketing reduces average padding 11.7→3.7 for WikiText-2, built-in pretrained models (a two-layer LSTM on WikiText-2 reports perplexity ≈170.49), end-to-end GNMT and Transformer implementations with reproduced BLEU numbers, and a hybridize path that caches symbolic graphs for speed. Shows ONNX import from PyTorch into MXNet and cites huge training costs (hundreds of TPU-days) for state‑of‑the‑art models.
AI-written summary · about
chapters (10)
- 0:00GluonNLP and MXNet intro
- 2:13Why another NLP framework
- 3:43NLP data challenges
- 8:03Data preprocessing
- 10:15Efficient mini-batching
- 14:09Language model building
- 21:47Machine translation
- 26:54Hybrid execution
- 33:08Framework interoperability
- 34:58Resources and next steps
related talks
30mAnyone can Build Great Deep Learning Applications - Deep Numpy
42mAnima Anadkumar, Principal Scientist, Amazon Web Services, Endowed Professor, CalTech
39mAccelerate large-scale machine learning with NP on MXNet
30mApache MXNet 2.0: Towards Synergistic ML and DL with Standardization
23mBest Practices for Building Production-Grade Deep Learning Systems
01h 16mDeep Dive into Deep Learning with MXNet
01h 25mSolving complex ML problems in a few lines of code with AutoGluon
01h 40mA Deep Dive into NLP with PyTorch