Quick Draw
My efforts to try to get a bronzwe medal in QuickDraw Kaggle Competition
Also this is kind of a project to get the structure of a good pytorch project.
It has some tools inside src/tools to create the images and the validation set
To execute the code:
python src/main.py config/some_config.json
You can also create a dry run test, so it's kind of faster... and also uses only 1% of the data
Installation
- Install conda
- Create a new conda environment
conda create --name quickdraw --file requirements.conda
- activate the environment
source activate quickdraw
- install apex
git submodule init
git submodule update
cd apex
python setup.py install
- Install requirments using pip
pip install -r requirements.txt
Enviromnent variables
- DRY_RUN:
Will not write anything into disk (only logs), and will set the % of the data to 1%, the idea is to be able to test that config works properly doing small experiments
- USE_VALID
Will use valid dataset as train, this is just a hack, because it takes like 3 min to list all the training files in quickdraw dataset, and for testing purposes it's useful to make this time go down
Tensor board
It's nice to have the some metrics... so you can use the tensorboard
tensorboard --logdir=./experiments/ &
Remote development with emacs
- Use tramp C-x C-f /ssh:kozko@....
- use anaconda-mode
- M-x pythonic-activate and select the $HOME/anaconda3/env/ environment you want
with that emacs should work perfectly :)
Things to improve
- Multiple valid during epoch:
Since the dataset is SOOO HUGE, if training with the full dataset takes a lot of time to get a valid and metric, like 12 hours per epoch
One thing could be do one valid each 10% or each 25% of epoch
- Loss functions
Test other loos functions like:
- Center Loss
- AM Softmax
- SoftMarginLoss
- l-softmax
- Stochastic Weight average
Try it? maybe... I don't know how much gain could come from here
- Other schedulers
I'm using LRonPlateau but the problem is that the epoch is so inmense it takes forever to detect the change...
- Denset efficient
- SOTA in CIFAR