load_script.py

import os
import json
import cv2
from datasets import Dataset, DatasetDict, Features, Value, Array2D, Array3D, Sequence
from pathlib import Path
import numpy as np

# Define constants
VIDEO_EXTENSIONS = ['.avi']
JSON_EXTENSIONS = ['.json']
KEYPOINTS = [
    "nose", "left_eye", "right_eye", "left_ear", "right_ear", "left_shoulder", "right_shoulder",
    "left_elbow", "right_elbow", "left_wrist", "right_wrist", "left_hip", "right_hip", 
    "left_knee", "right_knee", "left_ankle", "right_ankle"
]

def load_video(video_path):
    """Reads a video file and returns a list of frames as NumPy arrays."""
    cap = cv2.VideoCapture(video_path)
    frames = []
    
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
        frames.append(frame)
    
    cap.release()
    return np.array(frames)

def load_json(json_path):
    """Loads the JSON keypoint data for each frame."""
    with open(json_path, 'r') as f:
        data = json.load(f)
    return data

def process_frame_data(frame_data):
    """Converts the frame's keypoints into a structured format."""
    detections = []
    
    # Check if 'data' exists in the frame_data
    if 'detections' in frame_data:
        for detection in frame_data['detections']:  # Now using 'data' instead of 'detections'
            if detection:  # Check if there's any valid detection data
                person = {
                    "confidence": detection.get("confidence", 0),
                    "box": detection.get("box", {}),
                    "keypoints": {
                        keypoint['label']: keypoint['coordinates']
                        for keypoint in detection.get('keypoints', [])
                    }
                }
                detections.append(person)
            else:
                print(f"Warning: Empty detection in frame {frame_data['frame_index']}")
    else:
        # Handle the case where 'data' is missing in the frame data
        print(f"Warning: 'data' key missing in frame data: {frame_data}")
    
    return detections



def get_file_paths(base_path, split="train"):
    """Returns video and JSON file paths."""
    video_paths = []
    json_paths = []
    split_path = os.path.join(base_path, split)
    
    for label in ['Fight', 'NonFight']:
        label_path = os.path.join(split_path, label)
        for video_folder in os.listdir(label_path):
            video_folder_path = os.path.join(label_path, video_folder)
            video_file = next((f for f in os.listdir(video_folder_path) if any(f.endswith(ext) for ext in VIDEO_EXTENSIONS)), None)
            json_file = next((f for f in os.listdir(video_folder_path) if any(f.endswith(ext) for ext in JSON_EXTENSIONS)), None)
            
            if video_file and json_file:
                video_paths.append(os.path.join(video_folder_path, video_file))
                json_paths.append(os.path.join(video_folder_path, json_file))
    
    return video_paths, json_paths

def load_data(base_path, split="train"):
    """Loads and processes the data for a given split (train or val)."""
    video_paths, json_paths = get_file_paths(base_path, split)
    dataset = []
    
    for video_path, json_path in zip(video_paths, json_paths):
        # Load video frames
        frames = load_video(video_path)
        
        # Load JSON keypoints
        keypoints_data = load_json(json_path)
        
        # Process the data
        frame_data = [process_frame_data(frame) for frame in keypoints_data]
        
        # Construct the data record
        dataset.append({
            'video': frames,
            'keypoints': frame_data,
            'video_path': video_path,
            'json_path': json_path
        })
    
    return dataset

def main():
    # Path to the dataset directory
    dataset_dir = '.'  # Replace with your actual dataset path
    
    # Load training and validation data
    train_data = load_data(dataset_dir, split="train")
    val_data = load_data(dataset_dir, split="val")
    
    # Convert to Hugging Face Dataset
    train_features = Features({
        'video': Array3D(dtype='int32', shape=(None, None, None)),  # None indicates variable sizes
        'keypoints': Sequence(Features({
            'person_id': Value('int32'),
            'confidence': Value('float32'),
            'box': {
                'x1': Value('float32'),
                'y1': Value('float32'),
                'x2': Value('float32'),
                'y2': Value('float32')
            },
            'keypoints': {key: Array2D(dtype='float32', shape=(2,)) for key in KEYPOINTS}
        })),
        'video_path': Value('string'),
        'json_path': Value('string')
    })
    
    # Create DatasetDict
    dataset_dict = DatasetDict({
        'train': Dataset.from_dict(train_data, features=train_features),
        'val': Dataset.from_dict(val_data, features=train_features)
    })
    
    # Save or push dataset to Hugging Face
    dataset_dict.save_to_disk("keypoints_keyger")
    # Or to upload: dataset_dict.push_to_hub("your_dataset_name")

if __name__ == "__main__":
    main()