Machine-Learning

Where to store your models if you want to access them rarely? Let’s say you want to be able to access them in few months, but you don’t want to pay for the running server all the time. This may not be the regular issue, but we faced this challenge when working on some PoC that was supposed to be shown to clients every few months. We didn’t want to pay for the server all the time, didn’t have to maintain it, but we wanted to have the model ready to be served in a few minutes. The way I solved this was by wrapping the ML model in a FastAPI app with BentoML and deploying it on Google Cloud Run as a docker container. ...

The report of all talks during the fourth day of OxML Health 2023 where I’ve been on behalf of Appsilon.. This day talks were on: Insightful Lecture on ML for Mental Health - Munmun De Chodhury Drug Discovery with ML - Ravi Patel How to make sure the Market wants your solution - Reza Khorshidi ...

The report of all talks during the third day of OxML Health 2023 where I’ve been on behalf of Appsilon.. This day talks were on: Introduction to Neural Networks - Gitta Kutyniok Medical Expert’s View on Impactful and Responsible ML in Health with EHR - Kazem Rahimi Chaotic Lecture on Graphs and ML - Pietro Liò ...

The report of all talks during the second day of OxML Health 2023 where I’ve been on behalf of Appsilon.. This day talks were on: The State of Computer Vision and Many CV Experiments - Cristian Rupprecht ML Challenges in Oncology and Multi-omics data - Mireira Crispin What is a Multimodality and How To Work With It? - Louis-Philippe (LP) Morency ...

The report of all talks during the first day of OxML Health 2023 where I’ve been on behalf of Appsilon. This day talks were on: Casual Inference - Cheng Zhang Computer Vision in medicine - Jorge Cardoso Knowledge Representation - Ali Eslami ...

Long time no see ^^ Today, I’d like to share my experience from OxML Health Summer School that I attended in person this year on behalf of Appsilon. ...

It’s always nice if you can use your technical skills for something good for our planet. I have this opportunity during work at Appsilon where from time to time we work on Data4Good projects. One such project was copepods/plankton project where I was doing segmentation of the oil sac in the Arctic plankton. This time I’d like to share something from the other pole of the globe. Currently we work on detecting the Antarctic shags nests on the Antarctic islands. Together with Jędrzej Świeżewski and Joanna Kaleta we were cooperating with researchers from IBB PAS on this task. Checkout the engagement summary blog post on the Appsilon blog. ...

One thing is to train the model, but another is to serve it. In between those two phases, model serialization and deserialization occurs. In simpler words it’s just model saving and loading. It’s important because different methods result in different: Inference speed Model size Python environment size If you are curious what are the ways to serialize model in PyTorch and how they compare, checkout my new post on Appsilon blog. ...

Finally the recording of the talk StyleGAN w Twoich rękach from SFI 2022 is available online! 🎉 ...

When working with deep learning problems I usually use PyTorch. I like this framework as it gives me a lot of freedom and allows to write code in the pythonic way. Using bare pytorch unfortunately often means writing a lot of boilerplate code, which no-one like. Another problem that arises during longer and larger projects is experiments running and configuration maintenance. I tackle those two problems by using pytorch-lightning and Hydra. Recently I described my approach in a post on the Appsilon blog. ...