alphapose-feedback/loop_alphapose_training.py

"""
TODO this script runs Alphapose's train.py, the created model is used to re-annotate the training-images, which is then fed back into the system

For now the only thing it does is that it merges alphapose-results.json with the coco input dataset.
"""

import argparse
import datetime
from io import TextIOWrapper
import json
from pathlib import Path
import subprocess
import yaml
import logging

logging.basicConfig()
logger = logging.getLogger('loop_alphapose_training')
logger.setLevel(logging.INFO)

class Iteration:
    def __init__(self, i: int):
        assert i > 0, "Iteration cannot be < 1"
        self.i = i
        self.nr = f"{i:04d}"
        self.name = f"feedback_iteration{self.nr}"

    def training_config_path(self):
        return Path("./data/coco") / "_iterations" / self.nr / "256x192_res50_lr1e-3_1x.yaml"
    
    def model_path(self):
        assert self.i > 0, "Iteration 0 only used at training time"
        return Path(f'exp/{self.name}-256x192_res50_lr1e-3_1x.yaml/final_DPG.pth')
    
    def merged_results_path(self, stage: str):
        assert stage in ['val', 'train']
        return Path(f"./data/coco/_iterations/{self.nr}/alphapose-results-{stage}2017.json")
    
    def interference_results_dir(self, stage: str, for_docker=False):
        assert stage in ['val', 'train']
        relative = "." if not for_docker else ""
        return Path(f"{relative}/out/_iterations/{self.nr}_{stage}2017/")
    
    def prev_iteration(self):
        return Iteration(self.i - 1)
    
    def next_iteration(self):
        return Iteration(self.i+1)

    @classmethod
    def from_str(cls, input):
        raise NotImplemented

def wrap_docker_cmd(cmd: list, container: str = 'alphapose'):
    pwd = Path(__file__).parent.absolute()
    return [
        'docker', 'run',
            '--rm',
            '--gpus', 'all',
            '--shm-size=10g',
            '-v', str(pwd) + '/exp:/build/AlphaPose/exp', 
            '-v', str(pwd) + '/data:/build/AlphaPose/data', 
            '-v', str(pwd) + '/out:/out', 
            '-v', str(pwd) + '/alphapose-docker/detector/yolox/data:/build/AlphaPose/detector/yolox/data', 
            '-v', str(pwd) + '/alphapose-docker/detector/yolo/data:/build/AlphaPose/detector/yolo/data', 
            '-v', str(pwd) + '/alphapose-docker/pretrained_models:/build/AlphaPose/pretrained_models', 
            container,
            *cmd
    ]

def create_config(iteration: Iteration):
    """
    build config e.g. configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml 
    """
    base_config = Path("../AlphaPose/configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml")
    with base_config.open('r') as fp:
        config = yaml.safe_load(fp)
    
    # the first iteration simply copies config
    if iteration.i > 1:
        prev_iteration = iteration.prev_iteration()
        config['DATASET']['TRAIN']['ANN'] = str(prev_iteration.merged_results_path('train').relative_to(config['DATASET']['TRAIN']['ROOT']))
        config['DATASET']['VAL']['ANN'] = str(prev_iteration.merged_results_path('val').relative_to(config['DATASET']['TRAIN']['ROOT']))
        config['DATASET']['TEST']['DET_FILE'] = f'./exp/json/{iteration.name}_test_det_yolo.json'
        config['DATASET']['TEST']['ANN'] = config['DATASET']['VAL']['ANN']

    new_config = iteration.training_config_path()

    if not new_config.parent.exists():
        logger.info(f"Make directory for config: {new_config.parent}")
        new_config.parent.mkdir()

    with new_config.open('w') as fp:
        yaml.dump(config, fp)
    
    return new_config

def run_cmd(cmd, in_docker):
    if in_docker:
        cmd = wrap_docker_cmd(cmd)

    logger.info(f"Run {cmd=}")
    proc = subprocess.Popen (cmd, shell=False)
    proc.communicate()

def create_and_run_training(iteration: Iteration):
    '''
    Basically just runs
        python scripts/train.py 
            --cfg exp/config_first.yaml 
            --exp-id coco_test2
    '''

    create_config(iteration)
    cmd = [
        'python', 'scripts/train.py',
            '--cfg', str(iteration.training_config_path()),
            '--exp-id', iteration.name
    ]

    run_cmd(cmd, in_docker = True)

def run_inferences(iteration: Iteration):
    '''
    create new train & validation datasets by basically running:
    python scripts/demo_inference.py
        --cfg configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml
        --checkpoint exp/coco_test1-256x192_res50_lr1e-3_1x.yaml/final_DPG.pth
        --gpus 0
        --indir data/coco/train2017 
        --outdir /out/first_train2017/ 
        --format coco
        --eval
    python scripts/demo_inference.py
        --cfg configs/coco/resnet/256x192_res50_lr1e-3_1x.yaml
        --checkpoint exp/coco_test1-256x192_res50_lr1e-3_1x.yaml/final_DPG.pth
        --gpus 0
        --indir data/coco/val2017 
        --outdir /out/first_val2017/ 
        --format coco
        --eval
    '''

    base_cmd = [
        'python', 'scripts/demo_inference.py',
        '--cfg', str(iteration.training_config_path()),
        '--checkpoint', str(iteration.model_path()),
        '--gpus', '0',
        '--format', 'coco',
        '--eval',
    ]

    cmd_infer_train = base_cmd.copy()
    cmd_infer_train.extend([
        '--indir', 'data/coco/train2017',
        '--outdir', iteration.interference_results_dir('train', for_docker=True),
    ])

    cmd_infer_val = base_cmd.copy()
    cmd_infer_val.extend([
        '--indir', 'data/coco/val2017',
        '--outdir', iteration.interference_results_dir('val', for_docker=True),
    ])

    run_cmd(cmd_infer_train, in_docker=True)

    run_cmd(cmd_infer_val, in_docker=True)

def coco_alphapose_merge_results_for_iteration(iteration: Iteration):
    for stage in ['val', 'train']:
        coco_alphapose_merge_results(
            Path(f"data/coco/annotations/person_keypoints_{stage}2017.json"),
            iteration.interference_results_dir(stage) / "alphapose-results.json",
            iteration.merged_results_path(stage)
        )

def done_training(iteration: Iteration) -> bool:
    return iteration.model_path().exists()

def done_interference(iteration: Iteration) -> bool:
    return all([
        (iteration.interference_results_dir(stage) / "alphapose-results.json").exists() for stage in ['train','val']
    ])

def done_merge_results(iteration: Iteration) -> bool:
    return all([
        iteration.merged_results_path(stage).exists() for stage in ['train','val']
    ])

def run_iteration(iteration: Iteration):
    if not done_training(iteration):
        create_and_run_training(iteration)
    else:
        logger.info(f"Training exists {iteration.name}")
    if not done_interference(iteration):
        run_inferences(iteration)
    else:
        logger.info(f"Inference results exist {iteration.name}")
    if not done_merge_results(iteration):
        coco_alphapose_merge_results_for_iteration(iteration)
    else:
        logger.info(f"Merged annotations exist {iteration.name}")

def coco_alphapose_merge_results(annotations_file: Path, results_file: Path, out_file: Path):
    today = datetime.datetime.now().strftime("%Y/%m/%d")
    info = {"description": "COCO 2017 Dataset, modified by Ruben van de Ven","url": "http://cocodataset.org","version": "0.1","year": 2023,"contributor": "COCO Consortium, Ruben van de Ven","date_created": today}
    annotations = json.loads(annotations_file.read_text())
    results = json.loads(results_file.read_text())
    # annotations_ann:list = annotations['annotations']

    # id_counts = {}


    for i, result in enumerate(results):
        if type(result['image_id']) == str:
            result['image_id'] = int(result['image_id'][:-4])
        
        result['id'] = i
        result['iscrowd'] = 0 # TODO make sure this is a right terminology/assumption (what is this crowd here anyway individaul/crowd?)
        result['bbox'] = result['box'] # TODO result.pop('box') to rename instead of copy
        result['area'] = 1 # TODO : for now to bypass ignore in alphapose/datasets/mscoco.py:87
        result['num_keypoints'] = 17 # TODO : verify that this is indeed always all points
        
        # There can be multiple annotations per image. Try to match the originals by keeping track
        # of their order of occurence
        # if result['image_id'] not in id_counts:
        #     id_counts[result['image_id']] = 0

        # # find matching annotations in original
        # origs = list(filter(lambda ann: ann['image_id'] == result['image_id'], annotations_ann))
        # assert len(origs) > id_counts[result['image_id']], f"Len should be one, found {len(origs)} for {result['image_id']}: {origs=}"
        # orig = origs[id_counts[result['image_id']]]
        # id_counts[result['image_id']] += 1

        # result['id'] = orig['id'] # we keep track of the original id
        
    annotations['annotations'] = results
    annotations['info'] = info

    with out_file.open('w') as fp:
        json.dump(annotations, fp)

    logger.info(f'wrote to {out_file.as_posix()}')


if __name__ == "__main__":

    i = 1
    while True:
        iteration = Iteration(i)
        logger.info(f"Run iteration {iteration.name}")
        run_iteration(iteration)
        i+=1

    # parser = argparse.ArgumentParser(description='Merge alphapose-results.json with an input dataset')
    # parser.add_argument('--annotations-file', required=True, type=argparse.FileType('r'),
    #                     help='an annotations file from the COCO dataset (eg. person_keypoints_train2017.json)')
    # parser.add_argument('--results-file', required=True, type=argparse.FileType('r'),
    #                     help='path to the alphapose-results.json')
    # parser.add_argument('--out-file', required=True, type=argparse.FileType('w'),
    #                     help='the filename of the merged result')
    

    # args = parser.parse_args()
    
    # coco_alphapose_merge_results(args.annotations_file, args.results_file, args.out_file)