PyTorch Reviews

I have been using PyTorch for the last three years, mostly through my thesis work and model deployment projects.

I use PyTorch for pretty much everything in my deep learning work, specifically to build model architecture from scratch. My main use case was my thesis on deepfake audio detection, where I worked on spectrogram-based models and self-supervised learning. I also used it to build computer vision projects at work.

For my thesis, I built a deepfake detection model using PyTorch. I took audio, converted it, and used self-supervised learning and foundational models to detect whether the audio was real or fake. I designed the architecture, trained the model, and ran all my experiments. That work was actually accepted for a poster presentation at a conference, making PyTorch central to the project through peer review.

Besides the audio work, I also used PyTorch for a computer vision project at work where I built and deployed a model for a real production system. I have used PyTorch on both research and production. I have also worked with different types of models through PyTorch, from conventional ones such as normal convolutional neural networks for vision to contrastive learning setups for my audio detection work.

The best features PyTorch offers for me is how flexible and natural it feels to build models, as the dynamic computation graph lets me design and change my architecture easily, which is a big help when I am doing research and experimenting a lot. I also appreciate how I can run multiple experiments using multiple GPUs and distributed computing directly with PyTorch. I appreciate how Pythonic it is, especially in comparison with TensorFlow, as the code feels like normal Python, making it easy to write, read, and debug. The GPU acceleration is very smooth, and the ecosystem around it, such as TorchAudio for my audio work, is strong, so it fits well. Overall, PyTorch just gets out of my way and lets me focus on the actual model.

There are a few areas for improvement. The biggest one for me is version and dependency management, as I have to ensure my Torch, TorchAudio, and CUDA versions are all aligned. If one is off, things break in ways that are not always obvious, wasting time before I even start training.

The deployment side could also be smoother. Building and training a model is very clean, but taking it to production takes more effort than it should, with most of my time on the work project going into deployment rather than training. Additionally, the documentation falls behind for newer features, so sometimes I have to dig through GitHub issues or understand how something works.

Another area needing improvement is the debugging experience for GPU memory issues, which could be better. When I hit an out-of-memory error, it is not always clear what is causing it or how to fix it, leading me to guess and trial-run. On the community and support side, I do not have many complaints. The community is large and active, so help is usually easy to find. However, the official support could do more to keep newer features documented alongside the core ones. Overall, the PyTorch system is one of its biggest strengths, not a pain point.

PyTorch is very accurate and reliable in my experience, as the results I get are consistent. When I train a model, it behaves the way I expect. For my thesis on deepfake audio detection, the accuracy was strong enough to hold up under peer review, so I trust what it produces. Regarding reliability, the framework itself is stable, as long as my data and architecture are also stable. PyTorch does its part without surprises, and the output quality ultimately depends on how I build and train. PyTorch gives me the control to get it right.

I used multiple GPUs on Great Lakes, which is a distributed system provided by my university where I can request a session for some GPUs. The thing is they have different GPUs, so sometimes I get an A40 with 48GB of RAM, while at other times I might get a Tesla GPU with only 16GB. PyTorch really helped me, especially the distributed computing, as everything allowed me to arrange my code to work on different GPUs. For example, when computing loss, it assigns different parts of the model to different GPUs or can break the model and assign it to different GPUs while computing the gradient from all the GPUs. This process requires extensive coding, but with PyTorch, I can implement it very easily with just a function, which was a huge help for me.

My main advice for others looking into using PyTorch is to start building. PyTorch feels like normal Python, so the best way to learn it is to pick a small project and work through it hands-on. A few practical tips include getting your environment right from the start, as most early headaches come from mismatched Torch, TorchAudio, and CUDA versions. Lean on the official trials, tutorials, and documentation since the core resources are excellent. Make good use of the ecosystem such as TorchAudio and TorchVision instead of building everything yourself and relying on other libraries that might not be as efficient with PyTorch. If you are coming from TensorFlow, especially, give it a little time. Once the flexibility clicks, it is hard to go back. Overall, PyTorch is a tool that rewards experimentation, so do not be afraid to try things and break.

PyTorch has been a reliable tool across everything I have done. From my thesis research to a real production deployment, it is flexible, easy to work with, and backed by a strong community ecosystem. I rate this product an eight out of ten.

I use the solution to manage libraries, develop code, and push it into GitLab.

PySearch is the best option for developing any project in the AIML domain. It's a top framework for running and managing projects efficiently in the current market scenario. It allows you to attach databases and manage all aspects of your project. It’s reliable, secure and user-friendly. It allows you to develop any AIML project efficiently. If any bugs or issues arise, you can manage them effectively. There are many services available for handling various tasks.

I would like to see better learning documents for PySearch.

If I'm working on a task and encounter an issue that can't be resolved promptly, I will connect with the development team through email, including screenshots of the problem. I also raise a ticket through ServiceNow and follow up to see how much time they need to provide a solution.

The product is easy to install. Once you understand it, you can attach and integrate libraries into your project.

The solution is affordable.

I recommend the solution. Suppose you have a cloud-based project. You can develop and migrate it using PySearch. You develop your project in the solution, using Python and its libraries, then push the code to GitLab. From GitLab, you can deploy it to the production level. It's user-friendly, secure, and suitable for developing any project. It facilitates easy deployment, and its runtime is the best. Overall, I rate it an eight out of ten.

I have been using PyTorch for research purposes. I have implemented high-end projects like image captioning, style transfer projects, and a query chatbot using PyTorch. I was learning by reading research papers, and all of them were implemented in PyTorch.

PyTorch allows me to build my projects from scratch. It provides a platform where I have control over the model parameters, and I can tweak the results. This feature is beneficial when working on complex and custom models. It offers more control for research and investigation purposes, unlike TensorFlow, which is more suited for speed and efficiency.

I do not have any complaints.

TensorFlow is better for large-scale implementation and provides speed and efficiency, whereas PyTorch is preferred for research and complex models.

I initially learned TensorFlow. However, during my internship, we experimented with PyTorch. It became popular in the market, and they switched because most research and development were using PyTorch.

During my internship, I worked with the back-end team using PyTorch. My involvement included checking the local server aspects while the team transitioned from TensorFlow to PyTorch.

I have learned and used TensorFlow before switching to PyTorch.

If you are developing something new and need more control for research purposes, you should use PyTorch. For business models requiring pre-existing implementations at a high scale in a shorter time, TensorFlow is better.

I'd rate the solution eight out of ten.

I use PyTorch for developing my own projects, such as artificial intelligence and machine learning projects.

I like PyTorch's scalability. I can use a very large scale of models that I can easily train in. OpenCV is also there, however, compared to OpenCV, PyTorch is better for converting actual text data to visual data. I'm actually developing my own tool for the application that I'm building for myself. It's a crypto exchange.

The analyzing and latency of compiling could be improved to provide enhanced results.

It's been four years since I started learning and using it.

The stability is fine for me. I haven't trained a huge level of models, so it's not been a hassle as of now.

PyTorch offers good scalability. It's very easy to use, and its flexibility with multiple languages makes it very good to use.

I did not have to contact their support team. I didn't encounter any issues like that.

Positive

Before using PyTorch, I did not use anything else. The first tool I used was PyTorch, then OpenCV. These two are the main tools that I use for LLM models.

I haven't gone for a paid plan yet. I've just been using the free trial or open-source version.

Overall, I would give PyTorch a ten out of ten. I would definitely recommend PyTorch to anyone trying to get into artificial intelligence or data science.

There have been a couple of use cases. Firstly, we use it for generative AI, specifically a style transfer kind of use case. 

The other use involves building classifiers over video streams. So, it's mostly classification as well as severity-related use cases.

The best feature is that it allows us to do batched mode, distributed data parallelism, and model parallelism. This enables distributed training, and the overall API interface is quite simpler compared to a data flow API. 

It makes sense both when using it yourself and when teaching it to more junior team members. It's much easier to discuss.  

Additionally, high-level APIs like Hugging Face simplify building on top of PyTorch. If you want to do things from scratch, the relative learning curve is simpler compared to Tensorflow, but beginners still need some help.

PyTorch could make certain things more obvious. Even though it does make things like defining loss functions and calculating gradients in backward propagation clear, these concepts may confuse beginners. We find that it's kind of problematic. Despite having methods called on loss functions during backward passes, the oral documentation for beginners is quite complex.

So, it is somewhat complicated for a beginner to learn how to use PyTorch.

If you refer to the original documentation, there are numerous examples. And as PyTorch has become very popular compared to Tensorflow in the last couple of years, there are many resources available, making it easier for us to adopt. 

I have been using PyTorch for three years. 

It is quite a stable product, and it's consistent across the board.

It's scalable. As far as my experience goes, it's been pretty scalable in terms of using multiple GPUs.

We've been able to take code from open-source, modify it, and tweak it to our liking on the PyTorch side. 

TensorFlow has a lot of baggage... like the technical debt from the transition between TensorFlow 1 and 2. It would be a big challenge if the models we find are built on TensorFlow 1. PyTorch has this consistency, which is quite good. So, it's stable and scalable as well.

We are a team of about 12 data scientists, and at least 95% of us use PyTorch.  

Before PyTorch, my company was using TensorFlow 1.

When we wanted to accelerate our delivery, we found that a lot of research papers and models were built on PyTorch. We could use and replicate them well because researchers were open-sourcing them on PyTorch. 

Also, the interface is simpler compared to TensorFlow. Keras made things simple for TensorFlow 1, but TensorFlow 2 integrated Keras, which has its complexities. PyTorch has been very consistent, so we discussed it internally and decided to go with PyTorch.

PyTorch's framework is quite good, and it's always improving. Recently, I've looked at a few frameworks: JAX, TensorFlow, and PyTorch. Since we are in the services industry, we need to deliver, and we don't always have the time to deeply explore the low-level development of neural networks.

JAX is very low-level, giving you lots of control over vectors and similar, a bit too detailed for us. The other two are from Google, and PyTorch, which has ties to the fast.ai community, has APIs that are much more understandable across the board, especially for PyTorch users.

PyTorch's Dynamic Computation Graph is a bit more nuanced because we don't have an apples-to-apples comparison between developing the same model on PyTorch versus TensorFlow.

However, PyTorch makes things easier to write – code readability and maintainability are big factors in data science. When we transfer data science code into the training server and beyond, PyTorch's dynamic graph and training setup make things easier and scalable.

The initial deployment is difficult, especially setting up a CPU version because my company currently works with Windows systems. We utilize the Windows Subsystem for Linux, and that kind of setup makes things a little bit complicated overall. 

If you have Windows OS and you're not using a GPU, then finding the right installation package will require some workarounds and research.

We deployed it on the cloud, both AWS and Azure. 

PyTorch is open source.

I would recommend using this solution. However, there are two tips I would like to add. 

1. Firstly, get a good book. One of my recommendations would be "Deep Learning with PyTorch". It's a really good book that I enjoyed. It goes into developing neural networks and other advanced analytics use cases from scratch using PyTorch. 
2. The other tip would be to look at PyTorch Lightning. It has a great API and adds nice definitions and code restructuring, allowing you to do things like distributed training much more easily across a GPU cluster.

Overall, I would rate the solution a seven out of ten because things are just getting better. Just recently, the team behind PyTorch Lightning (which is separate from the main PyTorch team) released PyTorch Lightning Studio. 

It makes the transition between training in distributed settings, utilizing client GPUs, and deployment/inference much more seamless. Nothing like that exists for TensorFlow, as far as I understand. 

A lot of researchers also prefer PyTorch, which is great because many new architectures that we might need to implement, study, or tinker with can be found as open-source code. We can try to build them using PyTorch and compare them to our own implementations. 

This gives us a baseline and reduces friction compared to understanding research and open-source code written in TensorFlow and then trying to implement it in PyTorch. 

Given these three things – research support, PyTorch Lightning simplifying things, and the platform I'm excited about, along with consistency and good resources – that's why I give PyTorch a high seven.

We use the solution for reliability engineering. We apply ML techniques to predict product failures and identify the reasons behind those failures.

The most valuable feature would be the solution’s performance. The product is more advanced than the other libraries that I have used. Since every functionality is production-ready, I can easily write code. I don't have to rewrite the code for production. It has production-ready code from the start. The tool is very user-friendly. It took us a week to learn how to use it. It's straightforward to learn.

I've had issues with stability when I use a lot of data and try out different combinations of modeling techniques. I would also like to see some improvement in parallel processing. We can take advantage of the GPU and compute it.

I have been using the solution for more than a year.

The tool is very scalable. When we started, we had only 1 GB of data. Currently, we process more than 10 GB of data. The product is pretty solid so far. I am the only one using the product in my organization. I am testing its performance.

Usually, I check the documents and all the support materials available online, so I never have to contact the support team.

I use the tool within the Databricks environment on AWS. A different team in my company deployed the tool in my organization. The deployment was fairly easy.

PyTorch is open-sourced. It is a versatile tool. We can get everything online. We can get paid support if we need it.

I chose PyTorch because I had done some projects using PyTorch in my previous company. I know its capabilities. I have used other products. Each has its own benefits. PyTorch is quick and easy to learn. Pushing it into production is also very easy.

I haven’t used the computational graph extensively. The accuracy metrics impact the model development. We haven't looked into the computational graphs yet. People who want to use the solution must look at all the options in the market, like TensorFlow and H2O.ai. PyTorch is useful for 95% of use cases without any problem. PyTorch would be a great place to start. Overall, I rate the product a nine out of ten.

We work a lot with text processing, vectorization, and other NLP tasks. Sometimes, we need to process websites, presentations, or optics quickly because they're used in user engines and other applications. We use PyTorch to test our implementations as well.

One of the things I really like about PyTorch is that it doesn't break with every update or deletion. That's why I switched from TensorFlow to PyTorch. I can still run the code I wrote three years ago in PyTorch on the latest version.

It's very backward compatible, and it's also very simple to use. It's not overly technical, and the flow is pretty intuitive.

And now that PyTorch is gaining credibility in the research space, it's becoming easier to find examples of papers that use PyTorch. This is an advantage for someone who uses PyTorch primarily.

On the production side of things, having more production tooling frameworks would be helpful. TensorFlow has a lot of elaborate frameworks e.g. for serving models, and that's one area where PyTorch could improve.

I have been using it for three years now. I use the latest version. 

I would rate stability a ten out of ten. I haven't encountered any stability issues with PyTorch.

I would rate scalability a nine out of ten. I haven't seen any scalability issues, and once you know how to tune the code, it performs very well.

PyTorch is used in a lot of the latest ML Research papers. Also, a moderate number of people use it in my organization use it. 

I switched to PyTorch from TensorFlow about three years ago. TensorFlow was quite hard to work with. The interface was a little bit difficult to handle, ended up using Keras more than Tensorflow itself, and with every update, they kept breaking things. So, it would be very frustrating to have to relearn everything after every update.

The initial setup is fairly easy. The deployment model varies depending on the use case. It can be on-premises, while others use it in the cloud.

The deployment time is very fast. With PyTorch, it depends on the project because it's just a library that's part of our system.

There is definitely an ROI. 

It is free.  There are no additional fees unless you want something very specialized. In that case, you might need to consult with an expert.

If you're looking for something cutting-edge, PyTorch is the best one right now because it's very versatile.

Overall, I would rate the solution a nine out of ten. 

I use the solution in my company primarily for building models. I took it up because I saw others using PyTorch in general. It has a whole graph structure that is automatically maintained behind it, which makes it suitable for building machine-learning models.

The best thing about PyTorch is that documentation is available. A lot of times, information is available in the documentation. In my company, we are aware of the functionalities we want to use, but we would like to know what functionality has been offered by the framework we are using. For instance, if I want to create a graph structure, I would want to have more and more control over what is happening in my model. The documentation is proper. The internal graph structure that the tool maintains and its own way of handling tensors are why I believe I would prefer PyTorch over the other products.

As we know there are newer models coming in with newer functionalities. Such new functionalities can be embedded in PyTorch by itself, so as users, we don't have to update anything. Considering the newer models that are coming into the AI world along with the new functionalities, one should be able to ensure that upgrades are made to the existing structures. When the new upgrades come in, the tool should make it easier for users to use them.

It should be made easier for users to move from one version to another.

I have been using PyTorch for three years. I am a user of the tool.

The product has breakdowns when we change the versions a lot. There are workarounds to deal with it, but when the version is being shifted to another version, we will get other errors owing to the updates. In many of the environments, I believe that there is a solution to deal with the breakdowns in the product. From PyTorch's side, it would be good if the upgrade to the next version is made easier.

I believe PyTorch would go into the hands of those people who are using it to build models. Right now, the tool is only used by my team, which consists of ten people. I don't know if people from some other team use the product or not, but if they do, then it is good for them.

For me, the product's initial setup phase is easy.

The solution is deployed locally and on a cloud service from Google.

Even before I joined the company, I believe that the organization used to use PyTorch. Based on whatever I got from the discussions, I came to know that the company was looking at a lot of different frameworks. With PyTorch, my company found that they had more control over what they were doing when it came to the functionalities since they got direct access to variables and functionalities in the framework in comparison to the other frameworks, where it was sort of they felt that there was less control, and it may be because the updates were not applied, which is what I feel. I also felt that I had more control over my network and my models with PyTorch's framework.

I wouldn't tell you point blank or ask you to use the tool since I use it. To give you proper experience and to help you decide, I would say that you look up the documentation and try all the basic functionalities that you require for your model building separately first. Don't just barge into model building, but rather try to take the whole code and run it. Take whatever basic functionalities you need, try it out separately, and see if you are comfortable with it or not, as it is exactly what I also did, after which I decided to take it up. Going through the process that I followed can make it very easy for you to decide whether you are comfortable enough or whether there are some things that will constantly worry you.

For beginners, it is fairly easy to learn.

I think most of the models that I have done were with the help of PyTorch, and all of them have been doing pretty well. I haven't had that sort of discomfort or anything with the tool.

I rate the tool an eight or nine out of ten.

Our primary use case for this solution is training your mathematical models. We are a data science team that trains mathematical models with this solution. It can spin up VMs, and you can use it up, or in your local machines.

The framework of the solution is valuable.

The training of the models could be faster. However, with PyTorch, modern training becomes a bit slower because it is within the models at Python.

We have been using the solution for approximately two years.

The solution is stable.

The solution is scalable. Approximately seven people are using it in our organization.

We do not have experience with customer service and support.

The initial setup is easy and took approximately ten minutes.

We chose this solution because it is primarily Google's framework, and over 90% of people worldwide use it.

I rate the solution a nine out of ten.