1664 - 了解为什么 PyTorch 是机器学习领域事实上的开源框架.pdf

1、Xu ZhaoResearch ScientistFacebook(Meta)Sahdev ZalaOpen Source Developer,STSMIBMPyTorch:PyTorch:Intro and Deep diveIntro and Deep dive2About SpeakersAbout SpeakersXu ZhaoXu Zhao PyTorch Developer PyTorch Benchmarking Expert Author of TorchBench/TritonBenchSahdev ZalaSahdev Zala Open Source developer

2、for 12+years PyTorch contributor PyTorch Ambassador etcd(Kubernetes data store)maintainer Kubernetes IBM Cloud project co-lead OpenStack core contributor OASIS TOSCA Tech Committee Member3AgendaAgendaPyTorch BasicOverview of deep learning and PyTorch workflowOverview of PyTorch tensors,data,and devi

3、ce supportDive into model-build,train,validate,predictAdvanced topicsDistributed trainingPyTorch Benchmarking and ProfilingTorch compiler and OpenAI Triton integration into PyTorchContributing to PyTorch4Some BasicsSome BasicsDeep LearningTensorModel parametersForward passBackward pass5What is PyTor

4、chWhat is PyTorch A leading framework for machine learning,in particular deep learning Provides building blocks to create and train model Provides support for CPU,GPU,parallel and distributed training Open Source project under the PyTorch FoundationPart of the Linux FoundationOpen governance,dynamic

5、 communityIBM is a premier member of the foundation Easy to use,rapid prototyping Eager mode by default Graph mode for compiler optimizations Primarily a Python interfaceSupports C+frontend as well Provides tools for benchmarking and profiling Wide extensibility and customization options Libraries/s

6、olutions for many domainsTorchVision,TorchAudio-provides popular public datasets,model architectures,etc.ExecuTorch for mobile/edge devices6InstallationInstallationCondaconda install pytorch torchvision-c pytorchPip pip install torch torchvision Build from source codeFor development or testing PyTor