Requirement

Main requirement:

• Pytorch 0.4.1

If you miss any package, please install it by: pip install missing_package

Data split

Change following parameters:

• skiprows: Number of row you want to skip
  Ex: If you want to skip 30k data, skiprows=(1,30000)
• nrows: Number of data / class
  Default: 50000 data / class
• root_csv: Directory of your train_simplified folder
  Ex: /media/ngxbac/Bac/competition/kaggle/competition_data/quickdraw/data/csv/train_simplified/
• split_csv: Directory where you want to save splited data into
  Ex:
  /media/ngxbac/Bac/competition/kaggle/competition_data/quickdraw/data/50k/

Run:
python split_data_top.py
Output:
There are 340 csv files of train and valid are saved at your split_csv. Each csv file has nrows data.

Run model

• Configure train.yml
  In this file, please change the main parameters as following:

  • train_split
    Path to train folder: {split_csv}/train

  • train_token
    Dont care, but it is same as train_split

  • valid_split
    Path to valid folder: {split_csv}/valid

  • valid_token
    Dont care, but it is same as valid_split

  You can change other parameters workers, batch_size, ... to be suiatable for your environement

• Run

  bash run_model.sh

  Log and checkpoints will be saved to ./logs/se_resnext101_50k. Change it as you want

Predict model

• Configure inference.yml
  In this file, please change:

  • infer_csv
    Path to your test_simplified.csv file

• Run

  bash predict_5best.sh

  We use multiple checkpoints (snapshot) during training. Ensembling 5 best checkpoints will give free 0.0005 boost.
  Outputs are the logits will be saved into your log folder that you defined above

Predict dataset for cleaning

• Configure inference.yml
  In this file, please change:

  • infer_csv
    Comment this line
  • data_clean_train
    Path to train data you want to clean
  • data_clean_valid
    Path to valid data you want to clean

• Run

  bash predict_data_for_clean.sh

  Please change to the best checkpoint of your model you use for clean data
  Ex: LOGDIR=$(pwd)/logs/clean_model_2_resnet34/

Clean data

In this file, change following parameter correct to your environment

• data_clean_train
  Path to train data you want to clean.
  Ex: /media/ngxbac/Bac/competition/kaggle/competition_data/quickdraw/data/30k/data_2/train/

• data_clean_valid
  Path to valid data you want to clean.
  Ex: /media/ngxbac/Bac/competition/kaggle/competition_data/quickdraw/data/30k/data_2/valid/

• data_clean_train_out
  Output of train data after clean.
  Ex: /media/ngxbac/Bac/competition/kaggle/competition_data/quickdraw/data/30k/data_2/train/

• data_clean_valid_out
  Output of valid data after clean.
  Ex: /media/ngxbac/Bac/competition/kaggle/competition_data/quickdraw/data/30k/data_2_cleannn/valid/

• data_train_predict
  Logit prediction of data_clean_train when using a model to predict.
  Ex: ./logs/clean_model_1_resnet34/dataset.predictions.data_2_train.logits.satge1.5.npy

• data_valid_predict
  Logit prediction of data_clean_valid when using a model to predict.
  Ex: ./logs/clean_model_1_resnet34/dataset.predictions.data_2_valid.logits.satge1.5.npy

Make submission

python make_submission.py

Make sure you change correct log_dir in make_submission.py

How to resume

Define the resume in train.yml and Run model again. Usually, we will resume from checkpoint.best.pth.tar in the logs folder.

Supported architectures and models

From torchvision package:

• ResNet (resnet18, resnet34, resnet50, resnet101, resnet152)
• DenseNet (densenet121, densenet169, densenet201, densenet161)
• Inception v3 (inception_v3)
• VGG (vgg11, vgg11_bn, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19, vgg19_bn)
• SqueezeNet (squeezenet1_0, squeezenet1_1)
• AlexNet (alexnet)

From Pretrained models for PyTorch package:

• ResNeXt (resnext101_32x4d, resnext101_64x4d)
• NASNet-A Large (nasnetalarge)
• NASNet-A Mobile (nasnetamobile)
• Inception-ResNet v2 (inceptionresnetv2)
• Dual Path Networks (dpn68, dpn68b, dpn92, dpn98, dpn131, dpn107)
• Inception v4 (inception_v4)
• Xception (xception)
• Squeeze-and-Excitation Networks (senet154, se_resnet50, se_resnet101, se_resnet152, se_resnext50_32x4d, se_resnext101_32x4d)
• PNASNet-5-Large (pnasnet5large)
• PolyNet (polynet)