Install requirements

Tested on python3.6. Preferable to start with clean virtual env.

pip install -r requirements.txt

Preprocess dataset

I used resized dataset provided by ppleskov

https://community.drivendata.org/t/resized-dataset-is-now-available/3874/21

Setup vaiable DATASET_PATH

For example:

DATASET_PATH='/home/ivan/projects/datasets/wildlife'

Download and unzip dataset to a DATASET_PATH folder.

Expected structure of data:

ll $DATASET_PATH | grep S | head -3

• drwxr-xr-x 170 ivan ivan 4096 1 21:43 S1/
• drwxr-xr-x 192 ivan ivan 4096 12 20:28 S10/
• drwxr-xr-x 176 ivan ivan 4096 1 21:42 S2/

ll $DATASET_PATH"/S10" | grep S | head -3

• drwxr-xr-x 5 ivan ivan 4096 12 19:48 B08/
• drwxr-xr-x 4 ivan ivan 4096 29 22:34 B09/
• drwxr-xr-x 4 ivan ivan 4096 12 19:48 B010/

DATASET_PATH also contains train_metadata.csv and train_labels.csv.

ll $DATASET_PATH | grep csv

• -rw-r--r-- 1 ivan ivan 3020365 8 20:13 test_metadata.csv
• -rw-r--r-- 1 ivan ivan 307504477 8 20:13 train_labels.csv
• -rw-r--r-- 1 ivan ivan 495022349 8 20:14 train_metadata.csv

All preprocessed data will be stored to:

$DATASET_PATH/preprocessed

Create the dir:

mkdir $DATASET_PATH/preprocessed

Open https://github.com/BraginIvan/Hakuna-Ma-data/blob/master/wildlife/preprocessing/remove_exif.ipynb and run all. (Takes several hours on SSD)

Run script to create background and mean images/.

python wildlife/preprocessing/sequences_to_images.py $DATASET_PATH

It will create DATASET_PATH/preprocessed/background and DATASET_PATH/preprocessed/mean folders with preprocessed images

You can look at images using https://github.com/BraginIvan/Hakuna-Ma-data/blob/master/wildlife/preprocessing/view.ipynb notebook.

Train DNN by original images

arguments:

• path to dataset ($DATASET_PATH)
• mode (0 - fast mode to be sure everything works) (1 and 2 slightly different pipelines for ensembling)

check if pipeline works:

python wildlife/training/original_images.py $DATASET_PATH 0

ll | grep insres

• rw-r--r-- 1 ivan ivan 655902600 6 07:47 insres_224_v0.h5
• rw-r--r-- 1 ivan ivan 655902600 6 00:00 insres_299_v0.h5
• rw-r--r-- 1 ivan ivan 655902600 6 00:03 insres_360_v0.h5
• rw-r--r-- 1 ivan ivan 655902600 6 00:06 insres_384_v0.h5

Run two training pipelines:

python wildlife/training/original_images.py $DATASET_PATH 1

python wildlife/training/original_images.py $DATASET_PATH 2

One pipeline on a single RTX 2080 Ti + 64 RAM + SSD takes ~ 48 hours.

Pre-trained models can be found by the link https://yadi.sk/d/a1HwAVbvKiIozg.

Train DNN by background images

arguments:

• path to dataset ($DATASET_PATH)
• mode (0 - fast mode to be sure everything works) (1 training mode)

check mode (fast training):

python wildlife/training/background_images.py $DATASET_PATH 0

ll | grep insres | grep background

training mode

python wildlife/training/background_images.py $DATASET_PATH 1

Train DNN by concatination of background and mean images

arguments:

• path to dataset ($DATASET_PATH)
• mode (0 - fast mode to be sure everything works) (1 training mode)

check mode (fast training):

python wildlife/training/concat_back_mean.py $DATASET_PATH 0

ll | grep connected

• rw-r--r-- 1 ivan ivan 1019633752 8 07:10 connected_model_v0.h5

training mode

python wildlife/training/concat_back_mean.py $DATASET_PATH 1

ll | grep connected

• rw-r--r-- 1 ivan ivan 1019633752 8 07:10 connected_model_v1.h5

prepare csv with predictions of seasons 10 and 9 for boosting

Looks like there is a memory leak of predict method in tensorflow-gpu==2.1.0. In case of memory error please downgrade to tensorflow-gpu==2.0.1

python wildlife/inference/eval_everything_fullhd.py $DATASET_PATH

ll | grep predictions

• rw-r--r-- 1 ivan ivan 154140986 мар 8 17:28 predictions_SER_S10.csv
• rw-r--r-- 1 ivan ivan 154130133 мар 8 16:33 predictions_SER_S9.csv

train boosting

open in jupyter wildlife/boosing/final_boosting-v12.ipynb, setup DATASET_PATH and run all .