In reality it shouldn’t continue improving forever. At particular learning rate, the rate would be so large that the gradient descent would in-fact diverges from minima.
where are running this?
If you want to be nerdy, you can actually calculate how much memory your network will occupy in the GPU - by calculating all the weights and biases and gradients etc. (don’t do it, there will be millions and millions of them in a serious network)
Practically - you run your program, if it reports OOM error, reduce batch size. Good thing is, you will get to know if your batch size is big or not in the first epoch itself.
I have a laptop with a gpu built in.
This might help…
We need to move files to the correct directory…
Is the learning rate plot created from randomized mini-batches?
Is there a link to the new AWS setup? Jeremy mentioned that he’s posting it but not sure if he already posted it.
What is the assignment for this week ? Any updates on setting up AWS ?
To install OpenCV through conda:
conda install --channel https://conda.anaconda.org/menpo opencv3
So trial and error, cool.
Preferably in the multiples of 2.
Will this or other lessons have any homeworks other than going through notebooks ?
Jeremy mentioned things we could do in lecture. I believe homework for this week is getting comfortable in tools we will be using and playing with code.
Talking Machines! It’s so good it ruined me for other DS podcasts, can’t find any others worth sitting through.
I’ve tried, but it keeps showing a error message:
➜ fast_ai conda install --channel https://conda.anaconda.org/menpo opencv3
Fetching package metadata …
Solving package specifications: .
UnsatisfiableError: The following specifications were found to be in conflict:
- opencv3 -> python 2.7*
- python 3.6*
Use "conda info " to see the dependencies for each package.
OpenCV can be a challenge to get installed correctly. This tutorial is the most helpful thing I’ve found to get it working in an anaconda environment:
I suggest just
pip install opencv-python - at least that’s what I use.
Regarding Data Augmentation, I was reading https://petewarden.com/2017/10/29/how-do-cnns-deal-with-position-differences/ blog retweeted by @jeremy on Twitter and then I wonder what difference does data augmentation make?
I think its pandas_summary. (underscore)
You will also require pytorch, torchvision, opencv in addition to the ones @poppingtonic has mentioned.
It works perfectly! Thanks, Jeremy!
The answer lies within this parody song of a friend of mine: