ProDiff/dataset/data_util.py

import torch
import numpy as np
import pandas as pd
from torch.utils.data import Dataset
from tqdm import tqdm
import h5py

class TrajectoryDataset(Dataset):
    """Dataset for loading trajectory data from HDF5 files."""
    def __init__(self, file_paths, traj_length):
        self.samples = []
        self.load_samples(file_paths, traj_length)
    
    def load_samples(self, file_paths, traj_length):
        for file_path in tqdm(file_paths, desc="Loading files", unit="file"):
            with h5py.File(file_path, 'r') as h5_file:
                for user_id in h5_file.keys(): # Iterate over users in the HDF5 file
                    user_group = h5_file[user_id]
                    latitudes = user_group['latitudes'][:]
                    longitudes = user_group['longitudes'][:]
                    hours = user_group['hours'][:]
                    
                    # Create samples by sliding a window of traj_length over the user's trajectory
                    if len(latitudes) > traj_length:
                        for j in range(len(latitudes) - traj_length + 1):
                            self.samples.append((hours[j:j+traj_length], latitudes[j:j+traj_length], longitudes[j:j+traj_length]))
                    elif len(latitudes) == traj_length:
                        self.samples.append((hours[:], latitudes[:], longitudes[:]))
    
    def __len__(self):
        return len(self.samples)

    def __getitem__(self, idx):
        hours, latitudes, longitudes = self.samples[idx]
        return torch.tensor(hours, dtype=torch.float32), torch.tensor(latitudes, dtype=torch.float32), torch.tensor(longitudes, dtype=torch.float32)


class PatternDataset:
    """Dataset for loading trajectory patterns, possibly for prototype learning."""
    def __init__(self, file_paths):
        self.trajectories = []
        self.load_samples(file_paths)

    def load_samples(self, file_paths):
        for file_path in tqdm(file_paths, desc="Loading files", unit="file"):
            with pd.HDFStore(file_path, 'r') as store: # Using pandas HDFStore
                data = store['data']
                for i in range(len(data)):
                    abs_time_list = np.array(data['ABS_TIME'][i])
                    lat_list = np.array(data['LAT'][i])
                    lng_list = np.array(data['LNG'][i])
                    trajectory = list(zip(abs_time_list, lat_list, lng_list))
                    self.trajectories.append(trajectory)

    def get_all_trajectories(self):
        return self.trajectories

    def pad_trajectories(self):
        max_length = max(len(traj) for traj in self.trajectories)
        padded_samples = []

        for traj in self.trajectories:
            if len(traj) < max_length:
                # Pad shorter trajectories with their last point
                last_point = traj[-1]
                padding = [last_point] * (max_length - len(traj))
                padded_traj = traj + padding
            else:
                padded_traj = traj
            padded_samples.append(padded_traj)

        return padded_samples


class MinMaxScaler:
    """Min-Max Scaler for trajectory data with global normalization support."""
    def __init__(self, global_params_file=None):
        self.min_val = None
        self.max_val = None
        self.global_params_file = global_params_file
        self.is_global = global_params_file is not None
        
        if self.is_global:
            self._load_global_params()

    def _load_global_params(self):
        """Load global normalization parameters from file."""
        import json
        import torch
        
        with open(self.global_params_file, 'r') as f:
            params = json.load(f)
        
        # Convert to tensor format: [hours, latitudes, longitudes]
        min_vals = [
            params['hours']['min'],
            params['latitudes']['min'], 
            params['longitudes']['min']
        ]
        max_vals = [
            params['hours']['max'],
            params['latitudes']['max'],
            params['longitudes']['max']
        ]
        
        # Shape: (1, 1, 3) to match data format (batch_size, traj_length, 3)
        self.min_val = torch.tensor(min_vals, dtype=torch.float32).view(1, 1, 3)
        self.max_val = torch.tensor(max_vals, dtype=torch.float32).view(1, 1, 3)

    def fit(self, data):
        """Fit scaler to data. If using global params, this does nothing."""
        if not self.is_global:
            self.min_val = data.amin(dim=(0, 1), keepdim=True)
            self.max_val = data.amax(dim=(0, 1), keepdim=True)

    def transform(self, data):
        """Transform data to [0, 1] range."""
        if self.min_val is None or self.max_val is None:
            raise ValueError("Scaler not fitted. Call fit() first.")
        
        # Move tensors to same device as data
        min_val = self.min_val.to(data.device)
        max_val = self.max_val.to(data.device)
        
        # Avoid division by zero
        range_val = max_val - min_val
        range_val = torch.where(range_val == 0, torch.ones_like(range_val), range_val)
        
        # Clamp input data to avoid extreme values
        if self.is_global:
            data_clamped = torch.clamp(data, min_val, max_val)
        else:
            data_clamped = data
            
        return (data_clamped - min_val) / range_val

    def inverse_transform(self, data):
        """Transform data back to original scale."""
        if self.min_val is None or self.max_val is None:
            raise ValueError("Scaler not fitted. Call fit() first.")
            
        # Move tensors to same device as data
        min_val = self.min_val.to(data.device)
        max_val = self.max_val.to(data.device)
        
        return data * (max_val - min_val) + min_val