Hello, I’m Andrei from Russia.
@yasyrev_andrei
For a long time i worked in business (distribution, construction and others). Couple years before i step side to be a full-time Dad, and now only join short time projects, as my kids take all my time for their school and exercises.
At free time (actually it only instead of sleeping) i learning DL.
It started almost occasionally about 5 years before. I started experiments with micro controllers for home automation and suddenly meet python on raspberry Pi. I find Python very interesting and i start learn it. I started code at school, it was Fortran and we use dark board and chalk for coding. Later we used punched cards and now real computer - some IBM mainframe. It was so interesting so i start to learn more. Then it was Pascal (again BIG IBM comp) and even C at at286. At Uni we don’t learn to code - i studied as engineer mechanic constructions. Once it was short experience with analog coding, it was curios but almost useless. After Uni i gone to business and for a very long time didn’t write any line of code. For business i used exel and intuition and often it was more accuracy then ours well paid analitics with Phd in math.
So i started learn Python. I like it very much. I find what a lot of things and tasks i can now do myself without “special people” and so on. Then i found what a lot of articles in my news line is about ML, DL with Python and other strange but interesting things. In 17-s i take Andrew Ng course. It was on Matlab and it was a lot of math - it scary me a lot but I did it! I understand what my brain is not rusted yet! After course i even rewrite a lot of staff to python, but still did not understand what to do with this. Later i found fastai, it looked very promising, but a cant find time for it - no time after work and small kids. And only late 18 i started part 1 3 edition course. Top - down method is rely great! After this i started understand how it works. And later at part 2 i continued dive to details. It was a lot of low level staff - i understood what i have to learn a lot! I stacked there a little because i like understand details of what i do. I do this as my hobby (yet!) so cant spent to much time for it. But now i feel very comfortable with python, pytorch and fastai (steel v1).
To better understand i started refactor and write from scratch resnet (and xresnet) and use it for imagenette/imagewooff training. Guys from [How we beat the 5 epoch ImageWoof
…] did great work, i tried to do it parallel, but they are SO FAST and active! Anyway i found some interesting things, now testing it, hope i can share it soon. And right now i just show some small trick to resnet model, that help beat current score on Woof leader board Imagenette/ Imagewoof Leaderboards. And thank to nbdev i can easy share my constructor for model https://github.com/ayasyrev/model_constructor, what i use in my experiments for easy change activations, pools and other staff. It not so powerful as xresnet in fastai v2, but hope it can be helpful for study and understanding.
Big thank whole fastai team (including forum guru) for you job and for invitation!
9 Likes