Diffusion code for Landmark

This repository is re-implementation of ducha-aiki/manifold-diffusion for Google Landmark Retrieval 2019 competition.

This program performs reranking index samples for each test sample in submission.csv .

Requirements

• Ubuntu 16.04
• Python >= 3.7
• Conda 4.6.1

You also needs to prepare the input directory as below.

/path/to/ids-and-features/
|-- test_features.npy # np.ndarray of test features. The shape is (num-samples, num-features)
|-- index_features.npy # np.ndarray of index features. The shape is (num-samples, num-features) |-- test_ids.pickle # pd.Series of test ids. `-- index_ids.pickle # pd.Series of index ids. /path/to/similarity
`-- submission.csv # pre-caluculated submission file. this file is reranking target.

Example Usage

Install dependencies

conda install -c pytorch faiss-gpu
pip isntall -r requirements.txt

Calculate similarity before diffusion

export PYTHONPATH=.
# CPU-only 
python3 ./landmark_diffusion/pre_diffusion.py \
-i /path/to/ids-and-features/ \
-o /path/to/similarity/ \
--nogpu \
--norm \
--K 50 \
--QUERYKNN 10
# Use GPU
python3 ./landmark_diffusion/pre_diffusion.py \
-i /path/to/ids-and-features/ \
-o /path/to/similarity/ \
--gpu \
--norm \
--K 50 \
--QUERYKNN 10

Generate submission.csv with diffusion

python3 ./landmark_diffusion/run_diffusion_with_similarity_for_subset.py \
-i /path/to/ids-and-features/ \
-s /path/to/similarity/ \
-o /path/to/result/submission.csv \ # Result file of this program
-R 10