asappresearch/sru
Training RNNs as Fast as CNNs (https://arxiv.org/abs/1709.02755)
Implements a parallelizable recurrent unit with CUDA kernels that achieves 10-16x speedup over cuDNN LSTM on GPUs while maintaining accuracy across NLP tasks. Features layer normalization, highway connections, and bidirectional processing with a PyTorch API matching `nn.LSTM`. Also includes SRU++, an improved variant combining recurrence with attention mechanisms for further training efficiency gains.
2,112 stars. No commits in the last 6 months.
Stars
2,112
Forks
305
Language
Python
License
MIT
Category
Last pushed
Jan 04, 2022
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/ml-frameworks/asappresearch/sru"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
miguelvr/dropblock
Implementation of DropBlock: A regularization method for convolutional networks in PyTorch.
OML-Team/open-metric-learning
Metric learning and retrieval pipelines, models and zoo.
kk7nc/RMDL
RMDL: Random Multimodel Deep Learning for Classification
MaximeVandegar/Papers-in-100-Lines-of-Code
Implementation of papers in 100 lines of code.
archinetai/surgeon-pytorch
A library to inspect and extract intermediate layers of PyTorch models.