Channel: Data Science Archive
https://code.fb.com/ai-research/floating-point-math/
浮点数计算方法改进,ASIC / FPGA and C++ / PyToch code,来自 Facebook AI。
浮点数计算方法改进,ASIC / FPGA and C++ / PyToch code,来自 Facebook AI。
Engineering at Meta
Making floating point math highly efficient for AI hardware
In recent years, compute-intensive artificial intelligence tasks have prompted creation of a wide variety of custom hardware to run these powerful new systems efficiently. Deep learning models, suc…
"FloWaveNet : A Generative Flow for Raw Audio" from Seoul National University
paper: https://arxiv.org/abs/1811.02155
code: https://github.com/ksw0306/FloWaveNet
paper: https://arxiv.org/abs/1811.02155
code: https://github.com/ksw0306/FloWaveNet
GitHub
GitHub - ksw0306/FloWaveNet: A Pytorch implementation of "FloWaveNet: A Generative Flow for Raw Audio"
A Pytorch implementation of "FloWaveNet: A Generative Flow for Raw Audio" - ksw0306/FloWaveNet
WaveGlow 是一个基于流的生成网络,从 Glow 和 WaveNet借鉴而来,用于语音合成,来自 NVIDIA。https://github.com/NVIDIA/waveglow
GitHub
GitHub - NVIDIA/waveglow: A Flow-based Generative Network for Speech Synthesis
A Flow-based Generative Network for Speech Synthesis - NVIDIA/waveglow
一个历史悠久的 ML 工具库,Shogun(将军)。
http://shogun-toolbox.org/examples/latest/index.html
http://shogun-toolbox.org/examples/latest/index.html
关于机器学习系统线上部署的一些问题,隐患和思考,虽然是 NIPS 2015,但是对现在的大部分问题依旧有很强的借鉴意义。https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdhttp:/martin.zinkevich.org/rules_of_ml/rules_of_ml
动态构建知识图谱,看起来是整合一个 SQuAD 和其他离散状态,这里的离散状态包括了每个entity的representation,比如词性,位置等等。于是机器在做阅读理解的时候,一句一句往下读,entity 的状态就会更新。来自 UMass 和 MSR Montreal。
paper: https://arxiv.org/abs/1810.05682
paper: https://arxiv.org/abs/1810.05682
基于 LSTM 构建语言模型,然后用作输入法,以前有看到过一个韩国人做的,这次作者来自东京大学和 CMU,数据集是日语的 BCCWJ。其实是2016年的工作,但是语言模型放进输入法还是一个挺自然的事情,看起来还是挺有意思。
paper:https://arxiv.org/pdf/1810.09309.pdf
code:https://github.com/yohokuno/neural_ime
paper:https://arxiv.org/pdf/1810.09309.pdf
code:https://github.com/yohokuno/neural_ime
一个对 LSTM 中 autoencoder 的科普介绍,还挺清楚。just another,有关键部分的 Keras code 帮助理解。https://machinelearningmastery.com/lstm-autoencoders
语言模型中的迁移学习进展和总结,对目前State of the Art 的 LM 都有介绍,包括allennlp 的 ELMo,ULMFiT,OpenAI 的 Transformer,以及最近 Google 刷屏的 BERT。https://drive.google.com/file/d/1kmNAwrSlFYo0cN_DcURMOArBwe9FxWxR/view
Google Docs
transfer_learning_with_language_models.pdf
PyTorch 的 BERT 实现,包括 script 来将 TensorFlow 的 pre-trained model 进行转换,作者来自huggingface。https://github.com/huggingface/pytorch-pretrained-BERT
GitHub
GitHub - huggingface/transformers: 🤗 Transformers: State-of-the-art Machine Learning for Pytorch, TensorFlow, and JAX.
🤗 Transformers: State-of-the-art Machine Learning for Pytorch, TensorFlow, and JAX. - huggingface/transformers
HotpotQA:一个 wikipedia-based QA pairs dataset。
paper:https://arxiv.org/abs/1809.09600
code:https://github.com/hotpotqa/hotpot
link:https://hotpotqa.github.io/
paper:https://arxiv.org/abs/1809.09600
code:https://github.com/hotpotqa/hotpot
link:https://hotpotqa.github.io/
GitHub
GitHub - hotpotqa/hotpot
Contribute to hotpotqa/hotpot development by creating an account on GitHub.
ICL 数学系DL课程的一些资料,包括有PyTorch和 TensorFlow 的 Tutorial 以及作业相关,看了一下Tutorial 部分,觉得非常有意思,和其他传统的基础作业不太一样,这里都是流行的落地项目,比如Question Answering、Generative Model with VAEs/GANs,非常值得一看。https://github.com/pukkapies/dl-imperial-maths
GitHub
GitHub - pukkapies/dl-imperial-maths: Code and assignment repository for the Imperial College Mathematics department Deep Learning…
Code and assignment repository for the Imperial College Mathematics department Deep Learning course - GitHub - pukkapies/dl-imperial-maths: Code and assignment repository for the Imperial College M...
一个将 scikit-learn estimator 转化成其他语言的工具,这样线上做 prediction 的时候会更加灵活,暂时还没有需要研究,不过看起来是非常有意义的项目,目前更新也比较活跃。https://github.com/nok/sklearn-porter
GitHub
GitHub - nok/sklearn-porter: Transpile trained scikit-learn estimators to C, Java, JavaScript and others.
Transpile trained scikit-learn estimators to C, Java, JavaScript and others. - nok/sklearn-porter
NIPS 2018 上 MPC solver,用于在强化学习模型中的控制辅助。Specifically, we differentiate through MPC by using the KKT conditions of the convex approximation at a fixed point of the controller. 作者是用在 PyTorch 上,做了一个 PyTorch 的 Lib,不过确实先前的control methods 都有局限。
paper: https://arxiv.org/abs/1810.13400
code: https://github.com/locuslab/mpc.pytorch
link: https://locuslab.github.io/mpc.pytorch/
paper: https://arxiv.org/abs/1810.13400
code: https://github.com/locuslab/mpc.pytorch
link: https://locuslab.github.io/mpc.pytorch/
arXiv.org
Differentiable MPC for End-to-end Planning and Control
We present foundations for using Model Predictive Control (MPC) as a differentiable policy class for reinforcement learning in continuous state and action spaces. This provides one way of...
HTML Embed Code: