Development on Inference.
In some models, the user must set stop_gradient for some variables which not need to compute gradients, it can lead to poor user experience.
The options of build_strategy should be opened.
Multi-Node GPU Training Benchmark
This project mainly focuses on building a standard benchmark for multi-card GPU training. The purpose are three folds:
- build a multi-card GPU training benchmark scripts that can be run under different released version.
- continuously release benchmark statistics on different standard benchmark models.
- provide the best practice of multi-card training on GPU for various kinds of tasks.
Task we chose:
- VGG16 on Imagenet Dataset
- Resnet50 on Imagenet Dataset
- Transformer on En-Fr/En-De Dataset
- Bert on English wikipedia Dataset
Hardware we chose: V100 with 8 cards on single machine. In general, our benchmark should work on 1 x 8, 2 x 8, 4 x 8, 8 x 8 V100.
No description
mixed precision training
float16 support for training in Volta GPU
The functions of inference, CE or tests related. To ensure the usability of Inference.
No description
For details, please refer to https://github.com/PaddlePaddle/Paddle/issues/10248
No description
No description
Support engines in the inference phase.
- Build a general framework to run sub-graph on the third-party engines
High performance and robust runtime for single-machine-multi-gpu scenario and apply it to 5~10 different production models.
No description
Since embedding layer may include a very large parameter which could not be saved in one trainer's memory, so we need to implement an approach to support lookup remote table instead of lookup local table.
Performance Tuning on the single node and single GPU.
Bootcamp
No description
This project tries to create several deep neural network based language model.
Sequence tagging is a very important task in Natural Language Processing, for example word segmentation, part-of-speech tagging, name entity recognition and so on. In this project, we aim to provide several basic sequence tagging examples to help people start with this task. We also hope to replicate several research paper model in Paddle Fluid.
To start with NLP application, it is helpful to use text classification as hands-on tutorial. In this project, we will create several deep neural networks for text classification problem. We will also replicate algorithms proposed in some research papers.
This project implements the CSP (https://en.wikipedia.org/wiki/Communicating_sequential_processes) model for implementing concurrent programming in Fluid. The goal of this project is to build Concurrent programming structures like Channel, Select, GoRoutines in Fluid.
To manage speech-related tasks
This project will add an web dashboard to the master process. This web dashboard will allow a Paddle cloud users to see detailed information about the current submitted job, including:
- Properties of the master process like task completion progress, number of pservers, ect
- Basic scalar metrics on the job (ie: Cost, accuracy, mean, ect)
- Integrated with Kubernetes dashboard, allowing users to view submitted jobs, and open up a dashboard to a particular job
Enhance Computer Vision on Fluid
Memory optimization on Fluid
No description
Fluid benchmark & Fluid books performance
PaddlePaddle Fluid benchmark & book performance validation
Multi-device Multi-thread Support
No description
Intel MKL and MKL-DNN optimization on fluid.
No description
Refactor distributed training for fluid.