PyTorch vs TensorFlow in 2023 FULL OVERVIEW

I'm going to show you a lot of information in this video so not time for Hells and hooks and intros just let's dive into the content of the video first we're going to talk about practical considerations when comparing py torch with tensor flow second we will compare them on model availability third we'll compare pytorch and tensorflow on deployment fourth we'll talk about the differences in pytorch and tensor flow ecosystems fifth we're going to talk a little bit about Jacks I hope I pronounced this correctly and six we're going to answer the question should you use pie torch or tensor flow depending on different cases such as industry research whether you're Hobbies or a total beginner I believe this video is going to give you a lot of information on understanding how both platforms work and it's going to save you a lot of time researching all of this on your own so let's [Music] go in the past tensor flow has been usually seen as industry focused while pytorch has been research focused but both have rapidly evolved making those older distinctions less relevant today with the growth of deep learning using pre-trained models had become increasingly important and that's why we're going to compare P torch with tensor flow on model availability also being able to efficiently deploy is crucial especially in the age of microservices and finally deep learning is now used across different Industries making it very important for the Frameworks to have robust ecosystems including Hardware Integrations overall the debate is not about which platform is the best one but it's about understanding which one aligns more closely with your specific needs and workflow in 2023 you are likely to hop on hugging face because it makes it possible to incorporate trained and tuned stateof art models into your pipelines in just a few lines of code hugging face Hub is a platform with over 300,000 ,000 models and more than 60,000 data sets when it comes to tensor flow it offers over 9,000 models at the time but almost 130,000 models for borch considering that the majority of all available tensor flow and caros models are already available for f torch as well looking at Publications by authors that were using either pytorch or tensor flow in 2018 2019 we see that 55% of authors who had used tensor floor in 2018 migrated to py torch in 2019 by the way this graph is available on assembly AI website which is very informative blog for AI overall the conclusion as far as availability and Trend goes is that P torch is leading despite the release of tensor flow version two one important exception is Jacks if I'm pronouncing this correctly even used by Google brain and we're going to talk about this a bit later in the video but to give you an idea in 2020 Google brain R announced that they were using JS to accelerate their research Deep Mind also created Sonet which is a highlevel API for tensor flow that is tailored towards research and sometimes called the research version of Kos on the other hand open AI tenderized the usage of pytorch internally in 2020 but their older Baseline repository is implemented in tensor flow so tensor flow still hold some Sligh in the reinforcement learning community developing and training largely Which models is half of the work deploying those products into production and managing them is the most difficult part and the short version here is that tensor flow is better than pytorch in deployment so let's say a few words about tensor flow light which is a tool used in most Google projects having been tested on over 1,000 Google projects tensor flow serving is capable of handling millions of requests per second it is used for deploying Mach machine learning models on Specialized grpc servers and provides remote access to them model deployment with tensorflow service is highly flexible and it also integrates perfectly with kubernetes and Docker when using tensorflow serving it's easy to update an already deployed model and also to roll back any earlier version of the deployed model without having to shut down the server the tool is designed for industrial production environments and it's a good choice where performance is a concern while 10% increase in runtime might not mean a lot to a researcher working with those Frameworks it will definitely save a lot of money to Big organizations tensorflow light is tailored for mobile and iio devices IT addresses constraints like latency size and power consumption and it supports a range of programming languages by torch obviously has been working on closing the Gap in deployment introducing torch server and py torch life to facilitate easier deployment torch server is a flexible model deployment tool and with a basic set of features like a model Archer tool a server metrix logging an API endpoint specification and model snapshots torch server attracts most of the business cases the latest version of torch server also supports hugging face Nvidia wave glow AWS cloud formation and others pytorch life Builds on pytorch mobile and focuses on deploying models on mobile devices through JavaScript and react native it aims to provide an endtoend workflow for Android iOS and Linux overall while pytorch has been the two gold deployment framework pytorch is definitely catching up and one thing to mention here is on an X that is great for those who prefer tensor flows deployment capabilities but need access to pytorch only model onx can serve as a way to Port models between the two Frameworks ecosystems F torch offers a variety of specialized libraries and platforms that make it attractive for diverse applications from computer vision text to speech audio processing to running models on tpus you got options let's go over now the 10 most popular libraries and tools py Hub is an official repository platform for pre-trained models optimized for research it hosts a variety of models across domains like audio vision and NLP pytorch xlaa is a package that facilitates the training of pytorch models on Google's Cloud tpus it acts as a bridge between pytorch and Google's xlaa compiler torch elastic is a collaboration between AWS and Facebook and this tool manages distributed train training coordinating worker processes and handling issues like server maintenance to prevent loss of training progress torch text is specialized for natural language processing and this Library includes frequently used data sets and data processing utilities for enp tasks torch audio is py torch official audio library boosting models like deep speech it offers walkthroughs for tasks such as automated speech recognition speech brain is another open-source speech to kit for pytorch it provides features like ASR speaker recognition and even sentiment analysis esnet is an endtoend speeech processing talk kit built on pytorch it aligns with cis's data processing style and supports tasks like speech recognition and translation torch X is an SDK for quick building and deployment of machine learning applications it supports launching distributed pytorch jobs and integrates well with torch elastic torch vision is the official computer vision library for py torch providing model architectures and data sets for CV projects lightning know as the caros of pytorch simplifies model engineering and training making the workflow more object oriented and reusable also something very important to add is that in 2023 pytorch released pytorch version two where one of the significant modifications is the torch compile this a feature that improves py torch performance and one of the most useful techniques here is the torch inductor which is a deep learning compiler that uses open AI Tron as a key building block tensorflow of course has its own suit of tools and repositories making it adaptable for different requirements so let's now talk about those we have the tensorflow Hub which is a treasure throve of pre-trained models that are ready for fine-tuning it has models for various domains including image text audio and video model Garden is a repository containing the source code for various state of art models it's beneficial for those who need to customize models Beyond just fine tuning extended tfx is an endtoend platform geared towards Motel deployment it covers everything from data loading to motel deployment and it's closely integrated with Google cloud clab is a cloud-based notebook for both tensor flow and Pie torch and collab offers free GPU and TPU access for model training it's like Jupiter notebook but for better Google Cloud integration media pipe is an open-source framework designed for building multiplatform machine learning pipelines it supports a variety of applications like phase detection and object tracking as well as multiple languages including Python and C++ Coro is focused on onboard locco AI it's a Google product that offers Hardware and software Solutions Coral aims to address challenges related to privacy and efficiency making AI more accessible for Edge Computing verx AI is Google's Cloud new unified machine learning platform that aims to consolidate various Services into a single ecosystem it's particularly strong in automating monitoring and governing machine learning workflows and by the way there's a free course on how to use verx AI by Deep learning with Andrew ink and I'm going to leave the link to the video I made about this as well tensorflow GS provides the power of tensorflow in JavaScript enabling browser and server side machine learning tasks it's great for those with web development background when it comes to data sets tensorflow has a repository of data sets curated by Google research accessible through data set search but it's not exclusive to tensorflow and it can be used by pie torch as well let's now say a few words about Jack Jax is a Google back numerical Computing library and it's rising in popularity and it might be a strong condenser in the future of deep learning it adopts a functionality pure approach different from both pytorch and tensorflow and this functionality pure approach might be a barrier or a boom if the industry adopts it it could be a game changer in general it seems like the future might not be between pytorch and tensorflow but pytorch version two and Jack finally if you're still right now right here debating between learning more specifically pytorch or tensor flow let's talk about a specific case depending on your context if you're already using tensor flow in an industrial setting it might be wise to stick with it especially given its advantages in deployment and monitoring however keep an eye on pytorch for future projects overall pytorch offers a more Dynamic and userfriendly environment making it ideal for new projects especially in the research domain through tools like on andx it's possible to develop in pytorch and then deploy using tensor flow allowing you to leverage the strength of both pytorch is largely considered the to go framework for research and the release of pytorch version two further solidifies its standing in the research Community dener flow has specific advantages for those working in reinforcement learning thanks to its agent library and deep Minds Acme framework Jax is emerging as a strong candidate for future research perly if you're interested in functioning programming and tpus if you are a hobbyist and if you are focused on implementing deep learning in a project then tensor flow is advised especially for iot embedded systems if your goal is understanding the fundamental of deep learning pytorch is generally better particularly if you are comfortable with python and for total beginners caros is recommended for its high level userfriendly components and then you can either drop down to tensor flow from Caris or switch to pytorch based on your comfort and objectives I hope this video helped you navigate in the different deep learning Frameworks that exist out there thank you for watching and I'm going to see you in the next video [Music] bye-bye

Channel: Team Up With AI

Views: 10,895

Keywords: chatgpt, large language models, how to use ai, artificial intelligence, generative ai, prompt engineering

Id: YZ6q1_kL51k

Length: 13min 29sec (809 seconds)

Published: Sat Sep 23 2023