I found a similar post: Wiki: Best Practices ,which was two years ago. I think it is worth to ask the question again.
I just wonder there would be nice a place/wiki to put the best practice in general and different subtopics. Is this a taboo for an open-source community? I would like to see the community as a model of a learning network. I have a number of parameters put it, then I will get the answer/output at the last layers. How do we apply a universal approximation theorem, regression and pendulum of different learning rate and other NN analogy in our learning/training model?
The newbie may have some practical questions:
Which config has been working well for what? See the post, read it first, GCP(or whatever) platform FastAi forum. Running GPU with your local machine is in general not recommended.
GCP availability and pricing can change from time to time.
GPU recommended P100, if not available how about T4?
FastAI or FastAi2(I would like to contribute, what is the prerequisite to contribute?)
I think this depends on what kind of contribution.
FastAi2 is rewritten in high-level API and – good for a beginner. As a beginner should I wait for fastai2 or go ahead with fastai1?
As now in April 2020, in my humble opinion, I would suggest going ahead with Fastai2.
Is version 4, only working with fastai2?
That’s why I suggested fastai2 not just for version 4, but fastbook as well.
GPU/CPU? GPU, as the new trend moving for cloud computing and probably the hardware for your future assignment or implication, will scale up, learning the just in time technology would be good. Thus, it is time to learn a bit more Google cloud computing, especially if you can get some free quota.
What platform, pros and cons?
I am using Colab, got some problems with widget in some notebook. Switching back and forth between colab. Colabe shortcut keys are different from Jupyter notebook.
Things keep change so quickly, what can we do?
If you are impatient and can’t wait, then you may expect to more-time to try different things out during the transition, or else you may want to wait until a stable version released. Course version 4 and Fastbook will be released in July 2020/
Linux/Window, Conda or not Conda?
I have been using, so far I am quite happy with that.
How does Virtual environment matter?
I am using a ubuntu 19 in virtual box hosted by Win10 and I am waiting for ubuntu 20, which should be released in April 2020.
Rule of thumb path of resources – wiki, github, document, fastai book, courses, videos, nb, forum, any suggestion of path for badges.
Roadmap/Path: Suggested prerequisite of one year of python coding, I think that is a relevant and fair request for the course. I didn’t have a one-year coding experience. I learned from kaggle micro-courses, start with python. I think that can be a good start. I also found spacy101 is quite helpful. Fastai pushes me to learn different things and revise my mathematics. I found a book of “Model Thinker” by Scott Page. I took that free course in Coursera. You may want to build your missing skills for the prerequisite when you are waiting for the stable release/Fastai2, course v4 or Fastbook, etc …
A thread about using Anki system to learn in FastAI Forum. I will try to find it our later.
ROI and regression evaluation for loss and accuracy for the application in quantitative or quantitative measures.
FastAi seems to be a bit different from other ML approaches, do not require a Ph.D. degree(to start??). I guess many of the good learners/achievers in this community are in very high percentile of persistence, hands-on competence, problem-solving skill and strength in other domains?
I started thinking and working on FastAi learning methodology. Why and how the learning strategy in AI/ML be different from other disciplines?
Someone post a thread about how to spend your time cost-effectively, while you are training data.
I started a journal about what I learned and log the result of my experiment as journal in wiki of my github repository.
I also took a course on “how to learn” in Coursera. I would like to summary of what I learned and shared here or in my GitHub in the future.
I would appreciate it if anyone can point me in that direction.
I came across “Machines learn Together and we should too” April 19, 2020.
How can machines and humans learn from each other?
While FastAi and the community here, always challenge yourself and ourselves how can we do better and learned from others who are different from us. How can we maximize the benefit of our diversity/variety, value, velocity(we are never fast enough), veracity and value(ethics)? Looking forwards to further collaboration and synergy.