dataset.py

import torch
from torch import Tensor
import torch.nn.functional as F
from torch.utils.data import Dataset
from typing import *
from pathlib import Path
import re
import xarray as xr
import numpy as np
import h5py


IGRA_VARS = ['air_temperature', 'relative_humidity', 'wind_speed', 'geopotential_height', 'air_dewpoint_depression']

CLARA_VARS_D = {
    'CFC': ['cfc'],
    'CTO': ['ctt', 'cth', 'ctp'],
    'IWP': ['iwp', 'cot_ice', 'cre_ice'],
    'LWP': ['lwp', 'cot_liq', 'cre_liq']
}

CLARA_VARS = [value for values_list in CLARA_VARS_D.values() for value in values_list]


class ERA5Dataset(Dataset):
    def __init__(
        self, 
        size: Tuple[int] = (720, 1440),
        window: int = 1,
        flatten: bool = True,
        mode: str = 'train'
    ) -> None:
        
        self.size = size
        self.window = window
        self.flatten = flatten
        self.mode = mode
        self.data_dir = Path('.') / 'era5'
        self.normalization_dir = Path('.') / 'era5' / 'stats_v0'
        
        # Lazily load data
        self.file_paths = list((self.data_dir / self.mode).glob('*.h5'))
        self.file_paths.sort()
        
        self.sample_indices = []
        for file_idx, file_path in enumerate(self.file_paths):
            h5_file = h5py.File(file_path, 'r')
            num_samples = h5_file['fields'].shape[0]
            self.sample_indices.extend([(file_idx, i) for i in range(num_samples)])
            h5_file.close()
    
        
        # Retrieve climatology to normalize
        self.normalization_mean = np.load(self.normalization_dir / 'global_means.npy')
        self.normalization_sigma = np.load(self.normalization_dir / 'global_stds.npy')
        

    def _open_file(self, file_path, sample_idx):
        h5_file = h5py.File(file_path, 'r')
        sample = h5_file['fields'][sample_idx]
        h5_file.close()
            
        return sample[:, :self.size[0], :self.size[1]]
    
    
    def __len__(self):
        return len(self.sample_indices) - self.window + 1

    
    def __getitem__(self, i):
        step_indices = [target_idx for target_idx in range(i, i + self.window)]
        
        x = list()
        for step_idx in step_indices:
            file_idx, sample_idx = self.sample_indices[step_idx]
            x.append(self._open_file(self.file_paths[file_idx], sample_idx))
        
        x = (x - self.normalization_mean) / self.normalization_sigma
        x = torch.tensor(x).float()

        if self.flatten:
            return x.flatten(0, 1), {}
        else:
            return x, {}
        

class AuxDataset(Dataset):
    def __init__(
        self,
        data_var: str = '',
        size: Tuple[int] = (720, 1440),
        window: int = 1,
        flatten: bool = True,
        mode: str = 'train'
    ) -> None:
        
        assert data_var in ['igra', 'clara'], 'Dataset is not implemented, choose one of [igra, clara]'
        
        self.data_var = data_var
        self.size = size
        self.window = window
        self.flatten = flatten
        self.mode = mode
        
        self.data_dir = Path('.') / data_var
        self.normalization_file = Path('.') / 'climatology' / f'climatology_{self.data_var}.zarr'
        
        # Check if years specified are within valid bounds
        if self.mode == 'train':
            years = [str(year) for year in np.arange(2011,2016)]
        elif self.mode == 'val':
            years = [str(year) for year in np.arange(2016,2018)]
        elif self.mode == 'test':
            years = [str(year) for year in np.arange(2018,2019)]
        else:
            raise NotImplementedError('Mode is invalid, choose one of [train, val, test]')
        
        # Subset files that match with patterns (eg. years specified)
        self.EXTENSION = 'nc'
        self.ENGINE = 'netcdf4'
        self.PARAMS = IGRA_VARS if self.data_var == 'igra' else CLARA_VARS
        
        file_paths = list()
        for year in years:
            pattern = rf'.*{year}\d{{4}}\.{self.EXTENSION}$'
            curr_files = list(self.data_dir.glob(f'*{year}*.{self.EXTENSION}'))
            file_paths.extend(
                [f for f in curr_files if re.match(pattern, str(f.name)) and not self._is_leap(f.name)]
            )
            
        self.file_paths = file_paths
        self.file_paths.sort()
        
        # Lazily load data (i.e., igra has multiple (4; 6-hourly) timesteps in a daily file, while clara only has 1)
        self.sample_indices = []
        if data_var == 'igra':
            self.sample_indices = [(file_idx, i) for file_idx in range(len(self.file_paths)) for i in range(4)]
        
        else:
            self.sample_indices = [(file_idx, 0) for file_idx in range(len(self.file_paths)) for i in range(4)]
        
        # Retrieve climatology to normalize
        self.normalization = xr.open_dataset(self.normalization_file, engine='zarr')
        self.normalization = self.normalization.sel(param=self.PARAMS)
        self.normalization_mean = self.normalization['mean'].values[np.newaxis, :, np.newaxis, np.newaxis]
        self.normalization_sigma = self.normalization['sigma'].values[np.newaxis, :, np.newaxis, np.newaxis]
        

    def _is_leap(self, filename):
        """FCN training data for ERA5 ignores leap day"""
        match = re.search(r'(\d{4})(\d{2})(\d{2})', filename)
        if match:
            year, month, day = map(int, match.groups())
            return month == 2 and day == 29
        
        return False
    
    def _open_file(self, file_path, sample_idx):
        sample = xr.open_dataset(file_path, engine=self.ENGINE)
        sample = sample[self.PARAMS]
        sample = sample.sortby('lat', ascending=False) # aligned with FCN ERA5
        sample = sample.roll(lon=len(sample.lon) // 2, roll_coords=True) # aligned with FCN ERA5
        
        
        if self.data_var == 'igra':
            
            # Some daily files do not have some snapshots in time; init an empty np.nan tensor instead
            try:
                sample = sample.isel(time=sample_idx).to_array().values
            
            except:
                sample = np.full((len(IGRA_VARS), self.size[0], self.size[1]), np.nan)


        else:
            sample = sample.to_array().values
            
        return sample
    
    
    def __len__(self):
        return len(self.sample_indices) - self.window + 1

    
    def __getitem__(self, i):
        step_indices = [target_idx for target_idx in range(i, i + self.window)]
        
        x = list()
        for step_idx in step_indices:
            file_idx, sample_idx = self.sample_indices[step_idx]
            x.append(self._open_file(self.file_paths[file_idx], sample_idx))
        
        x = (x - self.normalization_mean) / self.normalization_sigma
        x = torch.tensor(x).float()
        x = torch.nan_to_num(x)

        if self.flatten:
            return x.flatten(0, 1), {}
        else:
            return x, {}


class MultimodalDataset(Dataset):
    def __init__(
        self, 
        datasets, 
        background: bool=False
    ):
        self.datasets = datasets
        self.is_flatten = datasets[0].flatten
        
        # Handle variable future timestepping to generate background state
        self.background = background 

    def __len__(self):
        if self.background:
            return len(self.datasets[0]) - 1  # lead_time = 1
        else:
            return len(self.datasets[0])

    def __getitem__(self, idx):
        dim = 0 if self.is_flatten else 1
        x = [dataset[idx][0] for dataset in self.datasets]
        x = torch.cat(x, dim=dim).float()
        
        if self.background:
            # Also return future lead_time as target y
            y = [dataset[idx+1][0] for dataset in self.datasets]
            y = torch.cat(y, dim=dim).float()
            return x, y, {}
        
        else:
            return x, {}