This is a solution to the Kaggle competition from Google. It's a single model (SE-ResNext50), single fold. My code is very similar to my kernel: https://www.kaggle.com/artyomp/resnet50-baseline. I used 224x224 crops from 256x256 images. I trained only on classes with >= 10 samples (92740 classes). Surprisingly, it wasn't much slower than training on 18k or 35k classes (maybe 20% slower).
How to start training:
./model.py --config models/v1.3.8.seresnext50_92740_classes.yml
Generating a submission:
./model.py --config models/v1.3.8.seresnext50_92740_classes.yml --weights saved_pytorch_model.pth --gen_predict
Cleaning a submission:
./patch_sub.py sub.csv or ./patch_sub2.py sub.csv or ./patch_sub3.py sub.csv. Use them all for the best effect.
This is the config of the best model: https://github.com/artyompal/google_landmark_2019/blob/master/models/v1.3.8.seresnext50_92740_classes.yml So, it adds a bottleneck layer with 1024 neurons, SGD with CyclicLR and rectangular crops from https://arxiv.org/pdf/1812.01187.pdf.
It resulted in 0.128 on public LB after almost 5 days of training on a single 1080Ti.
Obviously, such a low score was caused by both a noisy train set and a noisy test set. I had no time to clean the train set and re-train, so I only cleaned the test set.
The first stage of cleaning was just No Landmark Detector from this kernel: https://www.kaggle.com/rsmits/keras-landmark-or-non-landmark-identification. I just got rid of predictions where any of top3 results were "not landmark".
The second stage. I noticed that there's a lot of portraits, selfies, and pictures of cars in the test set. So I took the pretrained Faster R-CNN from the new TorchVision release (checkout object_detection.py) and got rid of images with too many humans or cars (confidence >0.5, total area > 0.4). Check this out: https://github.com/artyompal/google_landmark_2019/blob/master/patch_sub2.py
The third stage. I noticed that there's a lot of random images: dogs, helicopters, flowers and so on. So I took a pretrained ImageNet classifier. I used ResNet50 (I wanted to use something better, but my GPUs were busy doing something else, so I needed a simple network; ideally, I'd take some smarter model). It's in the image_classifier2.py. So I rejected images with any of those classes and confidence > 0.7 (https://github.com/artyompal/google_landmark_2019/blob/master/patch_sub3.py)
The first stage gave me 0.02 on the public LB, but didn't help at all on the private LB. I wonder if the public/private split was made by class? The private part of the stage-2 test set seems to have cleaner data.
In total, before cleaning I had 0.128 on public LB, 0.144 on private LB. After cleaning I got 0.162 on public LB, 0.148 on private LB.
PS: thanks to @pudae81 for this repo: https://github.com/pudae/kaggle-humpback. I borrowed some code from there.