[Question] Does thirdai really support training models of tensorflow and pytorch?

[cheburakshu]

Hi, I came across your claim where you state the code attached in the snippet (3 lines!) beats some of the state of art trainings on GPUs etc. Ref: https://www.thirdai.com/cpu-or-gpu/

However, upon closer examination of the attached screenshot I have the following questions:

1. The bolt engine is just imported never used.
2. There is a model_converter library used but is not part of thirdai distribution itself.
3. Distributed training isn't used, which seems like the performance gain will be in order of magnitudes?
4. None of the demos in this repo include even a sample of tensorflow/keras/pytorch based models, which is central to your claim in the above link.

I find every example using only UDT.

What should a user like me, expect from your library after signing up for trial license? Kindly let me know if there is any plan to include examples from tensorflow etc which you would have obviously tested with, as it is claimed in the website.

Thanks.

[vindows]

Hello,

Thanks for your interest in our project and for taking the time to ask such detailed questions.

Firstly, I want to clarify that the Universal Deep Learning Transformer (UDT) is essentially a wrapper around the BOLT engine. This wrapper is designed to simplify usage and abstract many of the configurations of BOLT for different problem types. Therefore, while you may only see UDT being directly used in the code, BOLT is indeed being utilized underneath.

Secondly, our approach to deep learning is somewhat unique. Instead of relying on established frameworks like TensorFlow, Keras, or PyTorch, we've built our own deep learning architectures from scratch. Our custom architectures are designed for efficiency and performance, which is why they are able to outperform other methods in many cases.

Regarding your question about the comparison to TensorFlow, Keras, and PyTorch models, we run these comparisons independently to showcase BOLT's performance. We understand that you might want to conduct these comparisons on your own, and we welcome that. We can provide some scripts that you can use to run these benchmarks independently, or you can use any public benchmark as well.

As a new user, we encourage you to start by trying out one of the demos available in the repository. These demos are designed to showcase different problem types that we support and give you a sense of how to build and test the models on actual use cases. You can use the public datasets provided in the demos or you can bring your own data to see how our models perform in your specific context.

I hope this answers your questions. We're excited for you to try out our library and see the results for yourself!

Thanks.

[cheburakshu]

Thank you @vindows. I really appreciate you taking the time to respond to me.

I have a simple use-case. I would like to know how to train and use this pytorch quick start model with Thirdai's BOLT.

I have many more pytorch models I would like to use BOLT for, once I am successful with this basic use-case.

I also want to know the process for conversion (if you have pre-existing scripts you can share, highly appreciate it), gotchas for pytorch models, and the level of Out-of-the-box support for conversion for pytorch models, as I wasn't able to find literature and/or documentation regarding the same.

Also, please clarify if this use case is not supported, as I won't be able to rewrite all models using UDT.

Thanks!!