Skip to content

Latest commit

 

History

History
269 lines (184 loc) · 10.2 KB

README.md

File metadata and controls

269 lines (184 loc) · 10.2 KB

Reproduce Character classification model

Overview

validation strategy

Dataset is split train and validation by book titles.

generate 2 patterns of training datasets.

  • validation: book_title=200015779 train: others
  • validation: book_title=200003076 train: others

Characters with few occurrences oversampling at this time.

For details, see gen_csv_denoised_pad_train_val.py

preprocessing and data augmentation

  • Preprocessing
    • Denoising and Ben's preprocessing for train and test images, then crop characters
    • When cropping each character, enlarge the area by 5% vertically and horizontally
    • Resize each character to square ignoring aspect ratio
    • To reduce computational resource, some models training with grayscale image
    • To reduce computational resource, undersampling characters that appeared more than 2000 times have been undersampled to 2000 times
  • Data augmentation
    • brightness, contrast, saturation, hue, random grayscale, rotate, random resize crop
    • mixup + RandomErasing or ICAP + RandomErasing (details of ICAP will be described later)
    • no vertical and horizontal flip
  • Others
    • Use L2-constrained Softmax loss for all models
    • Warmup learning rate (5epochs)
    • SGD + momentum Optimizer
    • MultiStep LR or Cosine Annealing LR
    • Test-Time-Augmentation 7 crop

For more details, see options of model training scripts

models

model arch channel input size data augmentation validation
EfficientNet-B4 Grayscale 190x190 mixup + RandomErasing 200015779
ResNet152 Grayscale 112x112 mixup + RandomErasing 200015779
SE-ResNeXt101 RGB 112x112 mixup + RandomErasing 200015779
SE-ResNeXt101 RGB 112x112 ICAP + RandomErasing 200003076
ResNet152 RGB 112x112 ICAP + RandomErasing 200003076

example of mixup + RandomErasing

mixup_re

example of ICAP + RandomErasing

ICAP is the data augmentation method I have implemented. Images cut from the four images are pasted while keeping the original image position.

icap_re

training flow

  • First training
    • Training 5 models with only training dataset (80~90epoch)
  • Pseudo labelling
    • Generate pseudo label and pseudo cropped images using 5 models with the best validation accuracy from the first training.
  • Second training
    • Training 5 models again with train and pseudo images (80~90epoch)
  • Refinement
    • Choose models with the best validation accuracy from the second training and resume training.
    • Using smoother target (0.9) for pseudo images, but 1.0 for original train images.
    • Traininng each model 3 epochs including pseudo images without data augmentaion.

You can download the refinemented models from here.

Details of training Flow

Training model inside the docker container. Start the docker container with run.sh. And train using shellscripts under train_scripts directory.

First training

# bash train_scripts/01_efficientnet_b4_val15779_l2softmax_mixup_re_normalize_gray190.sh
... wait many hours
# bash train_scripts/02_resnet152_val15779_l2softmax_mixup_re_gray112.sh
... wait many hours
# bash train_scripts/03_seresnext101_val15779_l2softmax_mixup_re_rgb112.sh
... wait many hours
# bash train_scripts/04_seresnext101_val3076_l2softmax_icap_re_rgb112.sh
... wait many hours
# bash train_scripts/05_resnet152_val3076_l2softmax_icap_re_rgb112.sh
... wait many hours

The five models require the following training time.

  • 01 about 44hours at GCP(V100x2)
  • 02 about 29hours at GCP(V100x2)
  • 03 about 32hours at GCP(V100x2)
  • 04 about 30hours at GCP(V100x2)
  • 05 about 27hours at GCP(V100x2)

Pseudo labelling

Generate pseudo labels (pseudo images) with 5 models from the first training.

Ensemble and NMS with First training models

When generating pseudo labels, model ensemble uses soft voting insted of hard voting. Then NMS 2 generated results.

Generate results with single detector and 5 models

$ bash scripts/auto_generate_first_model_results.sh

generates 4 results

val_detector_060000_tta7_first_5models_soft_prob val_detector_100000_tta7_first_5models_soft_prob test_detector_060000_tta7_first_5models_soft_prob test_detector_100000_tta7_first_5models_soft_prob

NMS with 2 ensemble results (validation and test)

$ scripts/results_nms_first_model_val.py
$ scripts/results_nms_first_model_test.py

generates 2 results

val_nms030_tta7_first_5models_soft_prob test_nms030_tta7_first_5models_soft_prob

Postprocessing (FalsePositive predictor)

Generate FalsePositive predictor using validation results.

searching hyper parameter with optuna

$ scripts/optuna_search_for_first_model_false_positive_detector.py
....
{'lambda_l1': 1.5464112458912599e-06, 'lambda_l2': 5.346737781503549e-06, 'num_leaves': 140, 'feature_fraction': 0.8534057661739842, 'bagging_fraction': 0.9376615592819334, 'bagging_freq': 1, 'min_child_samples': 72}

generate FalsePositive predictor for first models

$ scripts/gen_first_model_false_positive_detector.py
...
saved: models/booster_for_val_nms030_tta7_first_5models_soft_prob.pkl

Generate pseudo images

$ python scripts/gen_pseudo_label.py

saved pseudo images under input/pseudo_images

In addition, the submission file at this point is generated as first_submission.csv

The scores at this time were:

  • Private Score: 0.930
  • Public Score: 0.928

Second training

Add pseudo images to training data

If you execute the following command after generating pseudo images, csv files for training data with pseudo images added will be generated.

$ bash scripts/gen_pseudo_csvs.sh

generates 2 csv files

input/denoised_train_pseudo_200015779.csv input/denoised_train_pseudo_200003076.csv

Training 5 models with pseudo images

Training 5 models again with train and pseudo images (80~90epoch)

$ bash train_scripts/pseudo_labeling/01_efficientnet_b4_val15779_coslinear_mixup_re_normalize_gray190.sh
... wait many hours
$ bash train_scripts/pseudo_labeling/02_resnet152_val15779_coslinear_mixup_re_gray112.sh
... wait many hours
$ bash train_scripts/pseudo_labeling/03_seresnext101_val15779_coslinear_mixup_re_rgb112.sh
... wait many hours
$ bash train_scripts/pseudo_labeling/04_seresnext101_val3076_coslinear_icap_re_rgb112.sh
... wait many hours
$ bash train_scripts/pseudo_labeling/05_resnet152_val3076_coslinear_icap_re_rgb112.sh
... wait many hours

The five models require the following training time.

  • 01 about 44hours at GCP(V100x2)
  • 02 about 29hours at GCP(V100x2)
  • 03 about 32hours at GCP(V100x2)
  • 04 about 30hours at GCP(V100x2)
  • 05 about 27hours at GCP(V100x2)

Refinement

  • Choose models with the best validation accuracy from the second training and resume training.
  • Using smoother target (0.9) for pseudo images, but 1.0 for original train images.
  • Traininng each model 3 epochs including pseudo images without data augmentaion.
$ bash train_scripts/pseudo_refine/01_refine_efficientnet_b4_l2softmax_gray190.sh
... wait a few hours
$ bash train_scripts/pseudo_refine/02_refine_resnet152_l2softmax_gray112.sh
... wait a few hours
$ bash train_scripts/pseudo_refine/03_refine_seresnext101_l2softmax_rgb112.sh
... wait a few hours
$ bash train_scripts/pseudo_refine/04_refine_seresnext101_l2softmax_rgb112.sh
... wait a few hours
$ bash train_scripts/pseudo_refine/05_refine_resnet152_l2softmax_rgb112.sh
... wait a few hours

The five models require 2 or 3 hours training time.

You can download the refinemented models from here. (repeat)

References