TL;DR MOOC + Kaggle. There are plenty of options for the course:
- fast.ai – an informal, up-to-date and inclusive course, although somewhat chaotic: long videos, loose structure. It has a VERY supportive community and it often dives into practical applications (i.e. Kaggle).
- cs231n.stanford.edu – Computer Vision course from Stanford. I have finished it myself in an offline setting (we had lectures based on this course) and can recommend it highly. There are several lengthy assignments in Python where you build up and train a neural network from scratch. Good stuff.
- coursera.org/learn/machine-learning – this one is classic. A good overview of everything but the homework assignments that are not very useful in practice. There is a successor to this one, deeplearning.ai – here assignments are good, but the material could be better.
- There is also a MOOC by ODS.ai (stands for "Open Data Science") – the Russian DS community – which is my personal favourite. It is published both in Russian and English, so a lot of folks all over the world were able to successfully finish it: mlcourse.ai
Then, after you went through a couple of lectures, you should grab any of the ongoing competitions on Kaggle, and start hacking: kaggle.com/competitions.
There is a handy online course solely about that: coursera.org/learn/competitive-data-science.
If that wasn't enough, or you want more specifics, write a comment below.