{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "ebe0faa7",
   "metadata": {},
   "source": [
    "This notebook prepares the datasets for training of the turtle detection model. First, it goes through the SeaTurtleID2022 dataset and converts the existing masks into the YOLO format needed by Ultralytics. Then it goes through the TurtlesOfSMSRC dataset, loads the masks created in the smsrc_prepare notebook and again, converts the masks to the YOLO format. Finally, the metadata are merged together and are ready to use the segmentation_train script, which first trains on SeaTurtleID2022 (photos below water) and then finetunes on the combined dataset (photos above water were added)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a2e66c17",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import json\n",
    "import shutil\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "from tqdm import tqdm\n",
    "from wildlife_datasets.datasets import SeaTurtleID2022, TurtlesOfSMSRC\n",
    "from wildlife_datasets.datasets.utils import find_images, parse_bbox_mask\n",
    "from wildlife_datasets.splits import ClosedSetSplit\n",
    "from turtle_detector import get_index, rle_to_yolo, uncompressed_rle_to_yolo"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "93be7212",
   "metadata": {},
   "outputs": [],
   "source": [
    "root_out = f'/data/wildlife_datasets/turtle-detector'\n",
    "\n",
    "for addition in ['images/train', 'images/val', 'labels/train', 'labels/val']:\n",
    "    for dataset_name in ['SeaTurtleID2022', 'TurtlesOfSMSRC']:\n",
    "        os.makedirs(os.path.join(root_out, addition, dataset_name), exist_ok=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "14b3e193",
   "metadata": {},
   "source": [
    "# SeaTurtleID2022"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4c664fa1",
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset_name = 'SeaTurtleID2022'\n",
    "root = '/data/wildlife_datasets/data/SeaTurtleID2022'\n",
    "\n",
    "dataset = SeaTurtleID2022(root)\n",
    "if dataset.df['path'].nunique() != len(dataset):\n",
    "    raise ValueError('path is not unique')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "64391cd5",
   "metadata": {},
   "outputs": [],
   "source": [
    "splitter = ClosedSetSplit(0.8)\n",
    "idx_train, idx_test = splitter.split(dataset.df)[0]\n",
    "idx_train += 1\n",
    "idx_test += 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "057b036e",
   "metadata": {},
   "outputs": [],
   "source": [
    "flipper_categories = {\n",
    "    '': 0,\n",
    "    'front_left': 2,\n",
    "    'front_right': 3,\n",
    "    'rear_left': 4,\n",
    "    'rear_right': 5,\n",
    "}\n",
    "\n",
    "root_ann = f'{root}/turtles-data/data'\n",
    "with open(os.path.join(root_ann, 'annotations.json')) as file:\n",
    "    annotations = json.load(file)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7ff91b8a",
   "metadata": {},
   "outputs": [],
   "source": [
    "for ann_img in tqdm(annotations['images']):\n",
    "    file_name = os.path.join(root_ann, ann_img['file_name'])\n",
    "    if ann_img['id'] in idx_train:\n",
    "        shutil.copy(file_name, f'{root_out}/images/train/{dataset_name}')\n",
    "    elif ann_img['id'] in idx_test:\n",
    "        shutil.copy(file_name, f'{root_out}/images/val/{dataset_name}')\n",
    "    else:\n",
    "        raise ValueError('Split wrong')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a5a19846",
   "metadata": {},
   "outputs": [],
   "source": [
    "for ann_ann in tqdm(annotations['annotations']):\n",
    "    if ann_ann['category_id'] == 1:\n",
    "        category_id = 0\n",
    "    elif ann_ann['category_id'] == 3:\n",
    "        category_id = 1\n",
    "    else:\n",
    "        location = ann_ann['attributes'].get('location', '')\n",
    "        category_id = flipper_categories[location]\n",
    "\n",
    "    image_id = ann_ann['image_id']\n",
    "    rle = ann_ann['segmentation']    \n",
    "    yolo_segments = uncompressed_rle_to_yolo(rle, class_id=category_id)\n",
    "    ann_img = annotations['images'][image_id - 1]\n",
    "    base_name = os.path.basename(ann_img['file_name'])\n",
    "    base_name = os.path.splitext(base_name)[0] + '.txt'\n",
    "\n",
    "    if image_id != ann_img['id']:\n",
    "        raise ValueError('Image ids are not ordered')\n",
    "    if ann_img['id'] in idx_train:\n",
    "        file_name = f'{root_out}/labels/train/{dataset_name}/{base_name}'\n",
    "    elif ann_img['id'] in idx_test:\n",
    "        file_name = f'{root_out}/labels/val/{dataset_name}/{base_name}'\n",
    "    else:\n",
    "        raise ValueError('Split wrong')\n",
    "\n",
    "    with open(file_name, 'a') as myfile:\n",
    "        for yolo_segment in yolo_segments:\n",
    "            myfile.write(yolo_segment + '\\n')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "65a8b7ce",
   "metadata": {},
   "source": [
    "# TurtlesOfSMSRC"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "33d88ac1",
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset_name = 'TurtlesOfSMSRC'\n",
    "root = '/data/wildlife_datasets/TurtlesOfSMSRC'\n",
    "\n",
    "dataset = TurtlesOfSMSRC(root)\n",
    "masks = pd.read_csv(f'{root}/masks.csv')\n",
    "masks['mask'] = masks['mask'].apply(parse_bbox_mask)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e632972b",
   "metadata": {},
   "outputs": [],
   "source": [
    "splitter = ClosedSetSplit(0.8)\n",
    "idx_train, idx_test = splitter.split(dataset.df)[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e6acc2bd",
   "metadata": {},
   "outputs": [],
   "source": [
    "annotation_categories = {\n",
    "    'turtle': 0,\n",
    "    'head': 1,\n",
    "    'flipper_fl': 2,\n",
    "    'flipper_fr': 3,\n",
    "    'flipper_rl': 4,\n",
    "    'flipper_rr': 5,\n",
    "}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b3377df4",
   "metadata": {},
   "outputs": [],
   "source": [
    "for image_id in tqdm(masks['image_id'].unique()):\n",
    "    i = get_index(dataset, image_id)\n",
    "    file_name = os.path.join(root, dataset.metadata.loc[i, 'path'])\n",
    "    if i in idx_train:\n",
    "        shutil.copy(file_name, f'{root_out}/images/train/{dataset_name}')\n",
    "    elif i in idx_test:\n",
    "        shutil.copy(file_name, f'{root_out}/images/val/{dataset_name}')\n",
    "    else:\n",
    "        raise ValueError('Split wrong')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d1eba0f2",
   "metadata": {},
   "outputs": [],
   "source": [
    "for _, mask in tqdm(masks.iterrows(), total=len(masks)):\n",
    "    category_id = annotation_categories[mask['label_side']]\n",
    "    image_id = mask['image_id']\n",
    "    rle = mask['mask']    \n",
    "    yolo_segments = rle_to_yolo(rle, class_id=category_id)\n",
    "    i = get_index(dataset, image_id)\n",
    "\n",
    "    base_name = os.path.basename(dataset.metadata.loc[i, 'path'])\n",
    "    base_name = os.path.splitext(base_name)[0] + '.txt'\n",
    "\n",
    "    if i in idx_train:\n",
    "        file_name = f'{root_out}/labels/train/{dataset_name}/{base_name}'\n",
    "    elif i in idx_test:\n",
    "        file_name = f'{root_out}/labels/val/{dataset_name}/{base_name}'\n",
    "    else:\n",
    "        raise ValueError('Split wrong')\n",
    "\n",
    "    with open(file_name, 'a') as myfile:\n",
    "        for yolo_segment in yolo_segments:\n",
    "            myfile.write(yolo_segment + '\\n')"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0b6f6683",
   "metadata": {},
   "source": [
    "# Create metadata"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d5908b52",
   "metadata": {},
   "outputs": [],
   "source": [
    "n_repeat = {\n",
    "    'SeaTurtleID2022': 1,\n",
    "    'TurtlesOfSMSRC': 30,\n",
    "}\n",
    "\n",
    "# First split and only then oversample to prevent train-test leak\n",
    "images = find_images(root_out)\n",
    "images = root_out + '/' + images['path'] + '/' + images['file']\n",
    "images_train = images[images.str.contains('/train/')]\n",
    "images_test = images[images.str.contains('/val/')]\n",
    "if len(images_train) + len(images_test) != len(images):\n",
    "    raise ValueError('The split into train and test images failed.')\n",
    "\n",
    "# Oversample (even the test set)\n",
    "idx_train = []\n",
    "idx_test = []\n",
    "for dataset_name in ['SeaTurtleID2022', 'TurtlesOfSMSRC']:\n",
    "    idx_part = list(images_train[images_train.str.contains(dataset_name)].index)\n",
    "    idx_train += n_repeat[dataset_name] * idx_part\n",
    "    idx_part = list(images_test[images_test.str.contains(dataset_name)].index)\n",
    "    idx_test += n_repeat[dataset_name] * idx_part\n",
    "images_train = images_train.loc[idx_train]\n",
    "images_test = images_test.loc[idx_test]\n",
    "\n",
    "# Save the oversampled splits\n",
    "images_train.to_csv(f'{root_out}/train.txt', header=False, index=False)\n",
    "images_test.to_csv(f'{root_out}/val.txt', header=False, index=False)\n",
    "for dataset_name in ['SeaTurtleID2022', 'TurtlesOfSMSRC']:\n",
    "    subset_train = images_train[images_train.str.contains(dataset_name)]\n",
    "    subset_train.to_csv(f'{root_out}/train_{dataset_name}.txt', header=False, index=False)\n",
    "    subset_test = images_test[images_test.str.contains(dataset_name)]\n",
    "    subset_test.to_csv(f'{root_out}/val_{dataset_name}.txt', header=False, index=False)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "sam3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.12.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}