3rd place winning solution for Inclusive Images Challenge (NIPS Competition track 2018)

This repo documents the code to reproduce team WorldWideIncludive winning model. The model scored 0.37102 (3rd place) on final stage 2 Leader Board.

For inference, please refer to Stage 2 Inference Pipeline below.

For training, please refer to Training pipeline part 1 and Training pipeline part 2 below.

For NIPS presentation, please refer to Inclusive Image Challenge NIPS.pdf in this repo.

dependencies:

• python 3.6
• tf: 1.8.0
• keras: 2.2.2
• cv2: 3.4.3
• sklearn: 0.19.1
• spicy: 1.1.0
• pyvips: 2.1.4

Stage 2 Inference Pipeline - for generating final submission:

• Change constants in a00_common_functions.py:

  • TEST_IMAGES_PATH - path to stage 2 test images

  • IS_TEST = 1 - for inference

• Run script: python3 r40_final_inference_submit.py

Training pipeline part 1 (zfturbo's part):

• Change constants in a00_common_functions.py:

  • DATASET_PATH - path to OID images

  • IS_TRAIN - 1 - for training

• Run each script one by one:

  • python3 a02_common_training_structures.py

  • python3 inception_resnet_v2/r40_inception_resnet_v2_training_model.py

  • python3 inception_resnet_v2/r50_inception_resnet_v2_validation.py

  • python3 inception_resnet_v2/r52_validation_with_thr.py

  • python3 resnet50_336/r40_resnet50_sh_336_training_model.py

  • python3 resnet50_336/r50_resnet50_sh_336_validation.py

  • python3 resnet50_336/r52_resnet50_validation_with_thr.py

Training pipeline part 2 (weimin's part):

Please note that all scripts related to weimin's model training are under directory ../weimin_model_training/*:

1. Change constants in a00_common_functions.py:

• DATASET_PATH: path to OID images - must be a directory that contains one and only one subdir (any name), that contains all training images
• ROOT_PATH: root directory that has the input folder where all competition data csv files stay
• TUNING_IMAGE_PATH: absolute path pattern that finds all 1000 tuning label images

2. Run each script one by one:

• python weimin_main_train_inception_resnet_0.1.py --train_new_model=True

• python weimin_main_train_inception_resnet_0.08.py --train_new_model=True

• python weimin_main_train_inception_resnet_0.05.py --train_new_model=True

• python weimin_main_train_inception_resnet_0.15.py --train_new_model=True

• python weimin_main_train_xception.py --train_new_model=True

• python weimin_validation_data_gen_ext_0.py

• python weimin_validation_data_gen_ext_1.py

• python weimin_validation_data_gen_ext_2.py

• python weimin_validation_data_gen_ext_3.py

• python weimin_validation_data_gen_ext_4.py

• python weimin_validation_data_gen_ext_10.py

3. Collapse - At the end of training you will have models and relevant threshold arrays.

• weimin's models and threshold paths are stored below:

  • ROOT_PATH + 'model_0/inception_resnet_v2_latest.h5'

  • ROOT_PATH + 'model_1/inception_resnet_v2_latest.h5'

  • ROOT_PATH + 'model_2/inception_resnet_v2_latest.h5'

  • ROOT_PATH + 'model_3/inception_resnet_v2_latest.h5'

  • ROOT_PATH + 'model_4/new_xception_latest.h5'

  • ROOT_PATH + 'modified_data/thr_arr_inception_resnet_version_1_sp_0.1_ep_0.9_min_1_def_0.9.pklz'

  • ROOT_PATH + 'modified_data/thr_arr_inception_resnet_version_2_sp_0.1_ep_0.9_min_3_def_0.9.pklz'

  • ROOT_PATH + 'modified_data/thr_arr_inception_resnet_weimin_version_3_sp_0.01_ep_0.99_min_1_def_0.99.pklz'

  • ROOT_PATH + 'modified_data/thr_arr_inception_resnet_weimin_version_4_sp_0.01_ep_0.99_min_1_def_0.99.pklz'

  • ROOT_PATH + 'modified_data/thr_arr_xception_sp_0.01_ep_0.99_min_1_def_0.99.pklz'

  • ROOT_PATH + 'modified_data/thr_arr_xception_sp_0.01_ep_0.99_min_1_def_0.9999.pklz' # used for indexing only