• RandomCropIfNeeded
• batch accumulation
• With SIZE = 320 for SEResNext101 / SENet154 and SIZE = 331 for PNasNet-5
• Filtering predictions by high loss
• Pseudo labeling add the most confident predictions (highest np.mean(np.abs(probabilities - 0.5)) ) to the training dataset.
• Culture and tags separately
• 2nd stage I construct the binary classification dataset: I took 1103 (number of classes) rows per each image and trying to predict that this class relates to this image (0 or 1)
• feature engineering
• probabilities of each models, sum / division / multiplication of each pair / triple / .. of models
• mean / median / std / max / min of each channel
• brightness / colorness of each image (you can say me that NN can easily detect it — yes, but here i can do it without cropping and resizing — it is less noisy)
• Max side size and binary flag — height more than width or no (it is a little bit better for tree boosting than just height + width in case of lower side == 300)
• add all 1000 (number of ImageNet classes) predictions to this dataset
• lightgbm
• postprocess
• (Done)Different threshold for cultures and tags models

• The technique was quite similar to my quickdraw solution. I put top 30 tags and top 20 cultures to the dataset for LGMB

• Tag Relevance Prediction

• (Done)adjust the threshold for each image according to the max probability of that image
• Mixup and RandomErasing

• knowledge Distillation
• train only fresh params

• Small modifications were made to the network, our logit takes input from the last two layers rather than the last one layer

• two a fully connected layers, one for the culture and one for the tag

Make folds

python make_folds.py --n-folds 40

Train se_resnext101 from fold 0 to 9:

python main.py train model_se101_{fold} --model se_resnext101_32x4d --fold {fold} --n-epochs 40 --batch-size 32 --workers 8

Train inceptionresnetv2 from fold 5 to 9:

python main.py train model_inres2_{fold} --model inceptionresnetv2 --fold {fold} --n-epochs 40 --batch-size 32 --workers 8

Train pnas models from fold 0 to 4:

python main.py train model_pnas_{fold} --model pnasnet5large --fold {fold} --n-epochs 40 --batch-size 24 --workers 8

The ensemble of these model is used to predict results in imet-predict-final.ipynb.