I work within the field of NLP and data logistics, but most of the information is stored in the description. I mostly use NLP algorithms to extract the sentiment or the similarity if they have a similarity score calculation. For example, something like the windshield cracking, or something extremely severe like a garage burning, or multiple car damage.
The NLP tasks don't deal with the financial data directly but use extra information to get some insights. It's not really a good practice in India to use deep learning models directly on financial data because of audit reasons.
I like that PyTorch actually follows the pythonic way, and I feel that it's quite easy. It's easy to find compared to others who require us to type a long paragraph of code. It's one of the easiest tools and takes the hint on its own. I was quite happy about its existence.
I have been using PyTorch for about three months.
I think PyTorch is a stable solution. I had to upgrade the box it was running on several times to handle the load from many models I was running parallely, but that's not specific to PyTorch. But I think that was more of an environmental issue and not a problem with PyTorch. There was a transition period where the initial models weren't production-ready, and I had slated Caffe2. But I think the current version is pretty good and pretty stable.
Initially, I was using the Keras addon, and these are like high-level wrappers for TensorFlow. I started to look at PyTorch, and it was quite easy to learn and easy to implement.
PyTorch is a new insert, and I'm the only one using it. Most of my teammates are on TensorFlow because that's something that has been stable for a long time. I'm just showing them that it's easy to use PyTorch.
I was also experimenting with MXNet, which is part of Amazon Deep Learning for three months. I've been providing proof of concepts for existing models using these different frameworks that are available in the marketplace.
The initial setup is quite easy as it is just a library install. The model building itself is fairly simple because it's pythonic. It was more natural and followed how Python should be used. I would say that it helps the structure.
We just followed all these open source guys out there. They had like a lab for data science and a data lab in our company, which is basically an incident technology solutions provider. Just like bringing an analytics practice into the mix. But it's still a work in progress, and we just use a Python server that has all the latest policies installed. You just query whatever you want. The Ops and the support team take care of it.
PyTorch is an open-source solution.
I would recommend this solution because PyTorch is like a godsend.
On a scale from one to ten, I would give PyTorch a ten.
