Recursion Cellular Image Classification - Winning Solution

October 9, 2019 ยท View on GitHub

This repository presents an outline of my approach for the Recursion Cellular Image Classification competition.

The pipeline of this solution is shown as bellows Pipeline

There are 3 main parts:

  • I. Pretrained from control images which has 31 siRNAs
  • II. Continue fintuning models with image dataset which has 1108 siRNAs.
  • III. Continue fintuning models with image dataset and pseudo labels.

The writeup can be found in here

If you run into any trouble with the setup/code or have any questions please contact me at ngxbac.dt@gmail.com

Hardware

DGX Workstation: 4 x V100 (16G)

Software

Please check the docker/Dockerfile.
Besides, you can check requirement.txt

Getting started

Thing you should know about the project.

  • We run experiments via bash files which are located in bin folder.

  • The config files (yml) are located in configs folder which are corresponding to each bash files.

    Ex: train_control.sh should go with config_control.yml

  • The yml config file allows changing either via bash scripts for the flexible settings or directly modification for the fixed settings.
    Ex: stages/data_params/train_csv can be ./csv/train_0.csv, ./csv/train_2.csv,... etc. So when training K-Fold we make a for loop for the convinent.

Common settings

The common settings in yml config file.

  1. Define the model
model_params:
  model: cell_senet
  n_channels: 5
  num_classes: 1108
  model_name: "se_resnext50_32x4d"
  • model: Model function (callable) which returns model for the training. It can be found in src/models/ package. All the settings bellow model_params/model are considered as parameters of the function.
    Ex: cell_senet has default paramters as model_name='se_resnext50_32x4d', num_classes=1108, n_channels=6, weight=None. Those parameters can be set/overried as the config above.
  1. Metric monitoring
    We use MAP@3 for monitoring. 
    state_params:
  main_metric: &reduce_metric accuracy03
  minimize_metric: False
  2. Loss
    LabelSmoothingCrossEntropy is used. 
    criterion_params:
  criterion: LabelSmoothingCrossEntropy
  3. Data settings 
      batch_size: 64
  num_workers: 8
  drop_last: False

  image_size: &image_size 512
  train_csv: "./csv/train_0.csv"
  valid_csv: "./csv/valid_0.csv"
  dataset: "non_pseudo"
  root: "/data/"
  sites: [1]
  channels: [1,2,3,4,5,6]
  • train_csv: path to train csv.
  • valid_csv: path to valid csv.
  • dataset: can be control, non_pseudo, pseudo. control is used to train with control images (Part I), non_pseudo is used to train non-pseudo dataset (Part II) and pseudo is used to train pseudo dataset (Part III).
  • root: path to data root. Default is: /data
  • channels: a list of combination channels. Ex: [1,2,3], [4,5,6], etc.

  1. Optimizer and Learning rate

      optimizer_params:
    optimizer: Nadam
    lr: 0.001

  2. Scheduler

    OneCycleLR.

    scheduler_params:
  scheduler: OneCycleLR
  num_steps: &num_epochs 40
  lr_range: [0.0005, 0.00001]
  warmup_steps: 5
  momentum_range: [0.85, 0.95]

Build docker

cd docker 
docker build . -t ngxbac/pytorch_cv:kaggle_cell

Run container

In Makefile, change:

  • DATA_DIR: path to the data from kaggle.
|-- pixel_stats.csv
|-- pixel_stats.csv.zip
|-- recursion_dataset_license.pdf
|-- sample_submission.csv
|-- test
|-- test.csv
|-- test.zip
|-- test_controls.csv
|-- train
|-- train.csv
|-- train.csv.zip
|-- train.zip
`-- train_controls.csv
  • OUT_DIR: path to the folder which contains log, checkpoints.

Run the commands:

make run 
make exec 
cd /kaggle-cell/

Part I. Train with from control images

bash bin/train_control.sh

This part, we use all the control images from train and test.

  • Input:

    • model_name: name of model.
      In our solution, we train:
      • se_resnext50_32x4d, se_resnext101_32x4d for cell_senet.
      • densenet121 for cell_densenet.

  • Output: Default output folder is: /logs/pretrained_controls/ where stores the models trained by control images. Here is an example we train se_resnext50_32x4d with 6 combinations of channels.

/logs/pretrained_controls/
|-- [1,2,3,4,5]
|   `-- se_resnext50_32x4d
|-- [1,2,3,4,6]
|   `-- se_resnext50_32x4d
|-- [1,2,3,5,6]
|   `-- se_resnext50_32x4d
|-- [1,2,4,5,6]
|   `-- se_resnext50_32x4d
|-- [1,3,4,5,6]
|   `-- se_resnext50_32x4d
`-- [2,3,4,5,6]
    `-- se_resnext50_32x4d

Part II. Finetuning without pseudo label

bash bin/train.sh

  • Input:

    • PRETRAINED_CONTROL: The folder where stores the model trained with control images. Default: /logs/pretrained_controls/
    • model_name: name of model.
    • TRAIN_CSV/VALID_CSV: train and valid csv file for each fold. They are automaticaly changed each fold.

  • Output:
    Default output folder is: /logs/non_pseudo/. Here is an example we train K-Fold se_resnext50_32x4d with 6 combinations of channels.

    /logs/non_pseudo/
|-- [1,2,3,4,5]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
|-- [1,2,3,4,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
|-- [1,2,3,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
|-- [1,2,4,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
|-- [1,3,4,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
`-- [2,3,4,5,6]
    |-- fold_0
    |   `-- se_resnext50_32x4d
    |-- fold_1
    |   `-- se_resnext50_32x4d
    |-- fold_2
    |   `-- se_resnext50_32x4d
    |-- fold_3
    |   `-- se_resnext50_32x4d
    `-- fold_4
        `-- se_resnext50_32x4d

Part III. Finetuning pseudo labels

The different between Part III and Part II is only train/valid csv input files.

bash bin/train_pseudo.sh

  • Input:

    • PRETRAINED_CONTROL: The folder where stores the model trained with control images. Default: /logs/pretrained_controls/
    • model_name: name of model.
    • TRAIN_CSV/VALID_CSV: train and valid csv file for each fold. They are automaticaly changed each fold.

  • Output:
    Default output folder is: /logs/pseudo/. Here is an example we train K-Fold se_resnext50_32x4d with 6 combinations of channels.

    /logs/pseudo/
|-- [1,2,3,4,5]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
|-- [1,2,3,4,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
|-- [1,2,3,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
|-- [1,2,4,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
|-- [1,3,4,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   `-- fold_4
|       `-- se_resnext50_32x4d
`-- [2,3,4,5,6]
    |-- fold_0
    |   `-- se_resnext50_32x4d
    |-- fold_1
    |   `-- se_resnext50_32x4d
    |-- fold_2
    |   `-- se_resnext50_32x4d
    |-- fold_3
    |   `-- se_resnext50_32x4d
    `-- fold_4
        `-- se_resnext50_32x4d

Predict

export LC_ALL=C.UTF-8
export LANG=C.UTF-8

CUDA_VISIBLE_DEVICES=2,3 python src/inference.py predict-all --data_root=/data/ --model_root=/logs/pseudo/ --model_name=se_resnext50_32x4d --out_dir /predictions/pseudo/

Where:

  • data_root: path to the data from kaggle.
  • model_root: path to the log folders (Ex: /logs/pseudo/, /log/non_pseudo/)
  • model_name: can be se_resnext50_32x4d, se_resnext101_32x4d or densenet121.
  • out_dir: folder where stores the logit files.

The out_dir will have the structure as follows:

/predictions/pseudo/
|-- [1,2,3,4,5]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   `-- fold_4
|       `-- se_resnext50_32x4d
|           `-- pred_test.npy
|-- [1,2,3,4,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   `-- fold_4
|       `-- se_resnext50_32x4d
|           `-- pred_test.npy
|-- [1,2,3,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   `-- fold_4
|       `-- se_resnext50_32x4d
|           `-- pred_test.npy
|-- [1,2,4,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   `-- fold_4
|       `-- se_resnext50_32x4d
|           `-- pred_test.npy
|-- [1,3,4,5,6]
|   |-- fold_0
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_1
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_2
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   |-- fold_3
|   |   `-- se_resnext50_32x4d
|   |       `-- pred_test.npy
|   `-- fold_4
|       `-- se_resnext50_32x4d
|           `-- pred_test.npy
`-- [2,3,4,5,6]
    |-- fold_0
    |   `-- se_resnext50_32x4d
    |       `-- pred_test.npy
    |-- fold_1
    |   `-- se_resnext50_32x4d
    |       `-- pred_test.npy
    |-- fold_2
    |   `-- se_resnext50_32x4d
    |       `-- pred_test.npy
    |-- fold_3
    |   `-- se_resnext50_32x4d
    |       `-- pred_test.npy
    `-- fold_4
        `-- se_resnext50_32x4d
            `-- pred_test.npy

Ensemble

Please note that: logits are the number of last FC layer which is not applied softmax.

In src/ensemble.py, model_names is the list of model that be used for ensemble.

Ex: model_names=['se_resnext50_32x4d', 'se_resnext101_32x4d', 'densenet121']

export LC_ALL=C.UTF-8
export LANG=C.UTF-8

python src/ensemble.py ensemble --data_root /data/ --predict_root /predictions/pseudo/ --group_json group.json

Ensemble with other logits

In our solution, we ensemble with other memeber. Following changes will make it works.

In src/ensemble.py,
ensemble_preds = (ensemble_preds + other_logits) / 121
Where: other_logits = np.load(<logit_path>).

export LC_ALL=C.UTF-8
export LANG=C.UTF-8

python src/ensemble.py ensemble --data_root /data/ --predict_root /predictions/pseudo/ --group_json group.json

Where:

  • data_root: path to the data from kaggle.
  • predict_root: folder where stores the logit files.
  • group_json: JSON file stores the plate groups of test set.

Output:
The submission.csv will be located at ${predict_root}/submission.csv.