Summary of pytorch learning resources
https://pytorchchina.com/2019/05/07/awesome-pytorch-chinese/
PyTorch learning tutorials, manuals
-
PyTorch English version official manual : For students who are good in English, I highly recommend the PyTorch official document, which will take you step by step from entry to proficiency. This document introduces in detail from basic knowledge to how to use PyTorch to build deep neural networks, as well as PyTorch syntax and some high-quality cases.
-
PyTorch Chinese official document : It does not matter if you have difficulty reading the above-mentioned English documents. We have prepared a more official PyTorch Chinese document for everyone. The document introduces each function in great detail, which can be used as a quick reference book of PyTorch.
-
PyTorch code tutorial that is more algorithmic and practical : There is a very high star on github. It is recommended that you study the above two PyTorch basic tutorials before reading this document.
-
Open source book : This is an open source book, the goal is to help those who want to use PyTorch for deep learning development and research friends to get started quickly. However, the content of this document is not complete, and it is still being updated.
-
PyTorch Chinese document that is easy to use : very suitable for novices to learn. This document starts from introducing what is PyTorch, to the installation of neural networks and PyTorch, to image classifiers and data parallel processing. It introduces the knowledge system of PyTorch in great detail, which is suitable for beginners to learn. The official website of the document: http://pytorchchina.com .
PyTorch video tutorial
- PyTorch video tutorial at station B : The first PyTorch video tutorial at station B has a very high hit rate recently. Although there are only eight episodes of the video content, it is very exciting to explain the profound things in a simple way. It’s just that there are no Chinese subtitles. It’s time for my friends to practice English...
- Foreign video tutorials : Another video tutorial of a foreign boss has a high click-through rate on YouTube, and it is also a pure English video. Do you think that foreign teaching videos can be very vivid and simple no matter how complicated the problem is? ?
- Mofan : I believe that Mr. Mofan should be very familiar. His series of videos on Python and deep learning have a high click-through rate on station B and YouTube. The PyTorch video tutorial was just released last year and I recommend it to newbies. .
- 101 Academy : The PyTorch series of video courses of the 101 Academy of Artificial Intelligence, which are more detailed and cover a wide range of knowledge points. Interested friends can listen to it.
- July Online : Finally, I recommend the leading artificial intelligence education platform in China-PyTorch introductory and actual combat series of July Online. Although the course is a paid course, the course includes PyTorch grammar, deep learning foundation, word vector foundation, NLP and CV project application, practical combat, etc. The combination of theory and practical combat is indeed more detailed than other courses, and I recommend it to everyone.
NLP&PyTorch combat
- Pytorch text : Torchtext is a very useful library that can help us solve the problem of text preprocessing. This github repository contains two parts:
- torchText.data: general data loader, abstraction and iterator for text (including vocabulary and word vector)
- torchText.datasets: Pre-training loader for general NLP datasets. We only need to install torchtext through pip install torchtext, and then we can begin to experience the conveniences of Torchtext.
- Pytorch-Seq2seq : Seq2seq is a rapidly developing field, and new technologies and frameworks are often released here. This library is the framework of the Seq2seq model implemented in PyTorch. The framework provides modular and extensible components for the training and prediction of the Seq2seq model. This github project is a basic version and the goal is to promote these technologies and applications Development.
- BERT NER : BERT is a pre-training language model proposed by Google in 2018. Since its birth, it has broken a series of NLP tasks, so it has always had a very important influence in the field of nlp. The github library is the PyTorch version of BERT, with many powerful pre-training models built-in, which is very convenient and easy to use.
- Fairseq : Fairseq is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling, and other text generation tasks. It also provides reference implementations of various Seq2seq models. The github repository contains instructions on getting started, training new models, using new models and tasks to extend Fairseq. Friends who are interested in the model can click on the link above to learn.
- Quick-nlp : Quick-nlp is a deep learning Nlp library inspired by the fast.ai library. It follows the same API as Fastai and has been extended to allow quick and easy running of NLP models.
- OpenNMT-py : This is a PyTorch implementation of OpenNMT, an open source neural network machine translation system. It is designed to facilitate research, try new ideas, and try new ideas in many fields such as translation, summary, image to text, morphology, etc. Some companies have proven that the code can be used in actual industrial projects. For more detailed information about this github, please refer to the link above.
CV&PyTorch actual combat
- pytorch vision : Torchvision is a library of convenient image manipulation tools independent of pytorch. Mainly include: vision.datasets, vision.models, vision.transforms, vision.utils several packages, installation and use are very simple, interested friends can refer to the above link.
- OpenFacePytorch : This github library is the implementation of OpenFace in Pytorch. The code requires the input image to be aligned and cropped in the same way as the original OpenFace.
- TorchCV : TorchCV is a PyTorch-based computer vision deep learning framework that supports most visual task training and deployment. This github library provides source code for most CV problems based on deep learning. Those interested in the CV direction are still waiting what?
- Pytorch-cnn-finetune : This github library uses pytorch to fine-tune pre-trained convolutional neural networks. Supported architectures and models include: ResNet, DenseNet, Inception v3, VGG, SqueezeNet, AlexNet, etc.
Pt-styletransfer: This github project is a neural style transfer in Pytorch. Specifically, there are several points to note:- StyleTransferNet as a class that can be imported by other scripts;
- Support VGG (this is before the pre-trained VGG model is provided in PyTorch)
- The ability to save intermediate styles and content targets for display
- Can be used as a function of the image checker matrix
- Automatic style, content and product image saving
- Matplotlib graphs and hyperparameter records lost over a period of time to track favorable results
- Face-alignment : Face-alignment is a 2D and 3D face alignment library implemented with pytorch. It uses the world's most accurate face alignment network to detect facial landmarks from Python and can detect points in 2D and 3D coordinates. This github library introduces the basic process of face alignment using Face-alignment in detail. Interested students are welcome to learn.
PyTorch paper recommendation
- Google_evolution : This paper realizes the result network that realizes the large-scale evolution of the image classifier proposed by Esteban Real et al. Before the experiment, we need to install PyTorch, Scikit-learn and download the CIFAR10 dataset .
PyTorch-value-iteration-networks: This paper is based on the author's original Theano implementation and Abhishek Kumar's Tensoflow implementation, including the realization of value iteration networks (VIN) in PyTorch. Vin won the Best Paper Award in NIPS 2016. - ** Pytorch Highway* *: Highway Netowrks allows information to pass through various layers at high speed and without hindrance. It is inspired by the gate mechanism in Long Short Term Memory (LSTM) recurrent networks, which allows information to pass through many layers without hindrance. The effect of training the deep neural network makes the deep neural network no longer just have the effect of the shallow neural network. This paper is the implementation of Highway network based on Pytorch.
- Pyscatwave : Cupy/Pythorn scattering realization. Scattering network is a kind of convolutional network, its filter is pre-defined as wavelet, no need to learn, can be used for visual tasks such as image classification. Scattering transformation can significantly reduce the spatial resolution of the input (for example, 224×224->14×14), and the pun power loss is obviously negative.
- Pytorch_NEG_loss : This paper is a Pytorch implementation of Negative Sampling Loss. Negative Sampling is a method for solving the word2vec model. It abandons the Huffman tree and uses the Negative Sampling (negative sampling) method to solve the problem. This paper is a research on the loss function of Negative Sampling. Interested friends can Click on the paper link above to learn.
- Pytorch_TDNN : This paper is a Pytorch implementation of Time Delayed NN. The paper described the principle and realization process of TDNN in detail.
PyTorch book recommendations
Compared with the current situation where Tensorflow-type books are already rotten, there are not so many PyTorch books published. The author recommends four PyTorch books that I think are not bad.
- "Getting the deep learning PyTorch" , Electronic Industry Press, author: Liao Xingyu. This "Introduction to Deep Learning PyTorch" is a relatively early one published among all PyTorch books. The author explains the grammar, principles, and actual combat of PyTorch in a simple way with his own way of introductory deep learning. It is suitable for Introductory learning for novices. But the downside is that there are many inaccurate and mechanical aspects in the book that need to be carefully identified by readers. Recommendation index: ★★★
- "PyTorch deep learning" People's Posts and Telecommunications Press author: Wang Hailing, Liu Jiangfeng. This book is an English translation book. The original author is two Indian tycoons. In addition to the basic grammar and functions of PyTorch, the book also covers advanced neural network architectures such as ResNET, Inception, DenseNet, and their application cases. . This book is suitable for data analysts, data scientists and other readers who have relatively some theoretical foundation and practical experience. It is not recommended as an entry choice for novices. Recommendation index: ★★★
- "Deep learning framework PyTorch entry and practice" , Electronic Industry Press, author: Chen Yun. This is a PyTorch book that went on the market in 2018. It contains two parts: introduction to theory and practical projects. Compared with other books of the same type, the book case is very informative, including: classic projects in Kaggle competitions, GAN generated anime avatars, AI Filters, RNN writing poetry, image description tasks, etc. The content setting of theory + actual combat is also more suitable for deep learning beginners and practitioners. Recommendation index: ★★★★
- "PyTorch machine learning from entry to actual combat" , Machinery Industry Press, author: Bao online school, Sun Lin and so on. This book is also a Pytorch tutorial that combines theory with actual combat. Compared with the previous introductory + actual combat tutorial, the feature of this book is that the theoretical part of deep learning is very detailed, and the actual combat projects behind it are more comprehensive. Overall, this book is also a good PyTorch introductory book for novices. Recommendation index: ★★★