Kuzushiji Recognition Kaggle 2019. Build a DL model to transcribe ancient Kuzushiji into contemporary Japanese characters. Opening the door to a thousand years of Japanese culture.
Google Landmark Retrieval 2019 Competition fast.ai Starter Pack
The code here is all you need to do the first submission to the Google Landmark Retrieval 2019 Competition. It is based on FastAi library release 1.0.47 and borrows helpser code from great cnnimageretrieval-pytorch library. The latter gives much better results than code in the repo, but not ready-to-make submission and takes 3 days to converge compared to 45 min here.
Making first submission
Install the fastai library, specifically version 1.0.47.
Install the faiss library.
conda install faiss-gpu cudatoolkit=9.0 -c pytorch-y
Clone this repository.
Start the download process for the data. It would take a lot, so in mean time you can run the code.
Because the code here does not depend on competition data for training, only for submission.
validation-no-training - playing with pretrained networks and setting up validation procedure
training-validate-bad - training DenseNet121 on created micro-train in 45 min and playing with post-processing. It works as described, but just because of pure luck: lots of different "subclusters" == labels are depicting the same landmark. So, do not use it for training of all 19k subclusters
submission-trained - creating a first submission. Warning, this could take a lot (~4-12 hours) because of the dataset size
Note that the project description data, including the texts, logos, images, and/or trademarks,
for each open source project belongs to its rightful owner.
If you wish to add or remove any projects, please contact us at [email protected].