Pytorch

I use PyTorch to teach deep learning to my university students and professionally train and deploy models. Sometimes, it is done directly or through other libraries like Transformers from Hugging Face. PyTorch is very flexible and easy to write with a battery included. It offers a nice tradeoff between helpfulness and flexibility.

• Fast prototyping.
• It's not platform-dependent.
• The pre-processing function is already present.

TensorFlow, JAX and Keras

Pytorch is an awesome way of coding Deep Learning and Reinforcement Learning Algorithms with great ease. Since it is mostly pythonic, converting your derived equations/algorithms and using your favourite optimizers to test is so great. Further, it has great extensions to use like weights and biases where you can see how weights change in your neural network. It is an ideal tool for experimentation in Deep learning domain.

• The ability to use it to replicate historic algorithms
• Derive and test new DL flows

• It has a positive aspect as the ease of development results in publishing more papers for the community

TensorFlow

TensorFlow, Google Cloud AI, Slack

We use Pytorch for Data Science related projects; it is a very advanced framework for doing Machine/Deep Learning for people who are already familiar with python. It has a lot of datasets and models integrated that can be used just with a few lines of code to create a quick POC. It's very easy to write our neural networks with Pytorch.

• Most popular datasets like mnist, etc are integrated.
• Fine-tuning models is easy.
• Community support is good.

• It helped us creating quick POCs for customers.
• We can do customisation as we need.
• There is a learning curve so people need to spend some time for getting used to it.

TensorFlow and Keras

TensorFlow, Keras, Python IDLE, PyCharm, Jupyter Notebook, Microsoft Visual Studio Code, Visual Studio IDE

We use Pytorch as the main framework for building ML models and writing data loaders. Being an AI company, we have to train a lot of deep learning models, which involves writing data loaders for our dataset, making networks, or using the existing networks from the torchvision library. Being an AI-first company, ML Scientists are supposed to experiment with the models, and that requires writing very robust and modular code.

• Loss functions
• Base dataloaders
• Torchvision models
• Neural Network module
• Inbuilt optimisers, initialisers

• Less time wasted on handling the library version issues
• Small learning curve as very similar to Python
• Compatibility with other popular Python libraries makes it easy to build a lot of things on it

TensorFlow and Keras

Slack, Brave, Calendly, Google Calendar

We are using Pytorch to construct computer vision Deep Learning models for a battery of projects in the Data Platform project pipeline. Pytorch delivers a very Pythonic way of dealing with Deep Learning models that, from our point of view, make it easier for us to put the code in production, work in teams and be able to improve those different models in an iterative way. The business problems that we are solving are the generation of models to predict different biomarkers in both 2D and 3D images to improve the selection of patients in clinical trials. Both the training and the prediction models in Pytorch are very friendly and with a lot of support from the community.

• Clean code
• Dynamic Graphic memory
• Pre generated docker images for cloud environments

• The ability to make models as never before
• Being able to control the bias of models was not done before the arrival of Pytorch in our company

TensorFlow

TensorFlow, Google Cloud AI, Vertex