maxliaops/kaggle-quickdraw

Countdown

• calculate confusion on val set, then use category confidence to "correct" category of noisy samples
• k-fold
• cnn tree
• pseudo labeling
• analyze test predictions which are likely to be wrong
• fine tuning
• consider entropy to remove noisy samples for training or find unlikely predictions
• use more/all data for training
• use soft/hard bootstraping losses

Ideas

• train separate models for small number of categories
  • use to decide upon "ambiguous" results
• break down precision per
  • category
  • recognized vs. non-recognized
• visualize confusion matrix
  • which categories are usually confused?
  • train separate models for categories which are often confused?
• ignore samples with recognized=false
• check how close the top-3 predictions softmax is for wrongly classified samples
• normalize batch
• mmap numpy images
• https://www.kaggle.com/c/quickdraw-doodle-recognition/discussion/68006
• add batch normalization as first layer in resnet/drn models
• train x classifiers on disjoint category groups, then classify with every classifier and pick the most discriminative one
• train first with cce loss, then with smooth_topk loss (tune smooth_topk loss parameters)
• remove the first maxpool layer from resnet/senet models
• use attention
• use hypercolumns
• add swa
• use small epsilon threshold to consider a model improvement a relevant improvement (for early abort etc.)
• add residual connections
• add simple augmentation (fliplr + scaling) which can be implemented while drawing strokes
• use cv2.polylines?
• stratify mini-batches (https://discuss.pytorch.org/t/how-to-enable-the-dataloader-to-sample-from-each-class-with-equal-probability/911)
• combine lr-plateau scheduling with sgdr
• augment the number of feature maps in the cnn model and add se/scse block
• use drop connect
• use fully convolutional model with progressive image resizing
• train on category subset and add "other" category for samples on in complementary category set
• http://cybertron.cg.tu-berlin.de/eitz/projects/classifysketch/
• include statistics in one of the channels used for training: http://rstudio-pubs-static.s3.amazonaws.com/292508_8ef4c9ec5f76421e92803fccba9765df.html
• https://github.com/KaiyangZhou/pytorch-center-loss
• ensembling
  • https://mlwave.com/kaggle-ensembling-guide/
  • https://www.analyticsvidhya.com/blog/2015/09/questions-ensemble-modeling/
• https://github.com/batzner/indrnn
• noise
  • https://arxiv.org/pdf/1705.03419.pdf
  • https://hackernoon.com/cnns-with-noisy-labels-using-deepcognition-research-paper-1bf4094e3bd9
  • https://arxiv.org/pdf/1703.08774.pdf
  • https://arxiv.org/pdf/1412.6596.pdf
  • https://arxiv.org/pdf/1710.09412.pdf

Papers

• https://arxiv.org/pdf/1704.03477.pdf
• https://arxiv.org/pdf/1706.03762.pdf
• https://arxiv.org/pdf/1802.07595.pdf
• https://arxiv.org/pdf/1806.05594.pdf
• http://openaccess.thecvf.com/content_cvpr_2018/CameraReady/2763.pdf
• https://arxiv.org/pdf/1501.07873.pdf
• https://arxiv.org/pdf/1803.00942.pdf
• https://arxiv.org/pdf/1804.00021.pdf
• http://cybertron.cg.tu-berlin.de/eitz/pdf/2012_siggraph_classifysketch.pdf
• https://arxiv.org/pdf/1511.04534.pdf
• https://arxiv.org/pdf/1405.3080.pdf
• https://ydwen.github.io/papers/WenECCV16.pdf
• https://arxiv.org/pdf/1706.02677.pdf