#!/usr/bin/env python3 """MultiTask Dataset module compatible with torch.utils.data.Dataset & DataLoader.""" from __future__ import annotations import os from pathlib import Path from typing import Dict, List, Tuple, Iterable from argparse import ArgumentParser from pprint import pprint from natsort import natsorted from loggez import loggez_logger as logger import torch as tr import numpy as np from torch.utils.data import Dataset, DataLoader from lovely_tensors import monkey_patch from tqdm import trange try: from npz_representation import NpzRepresentation except ImportError: from .npz_representation import NpzRepresentation monkey_patch() BuildDatasetTuple = Tuple[Dict[str, List[Path]], List[str]] MultiTaskItem = Tuple[Dict[str, tr.Tensor], str, List[str]] # [{task: data}, stem(name) | list[stem(name)], [tasks]] TaskStatistics = Tuple[tr.Tensor, tr.Tensor, tr.Tensor, tr.Tensor] # (min, max, mean, std) class MultiTaskDataset(Dataset): """ MultiTaskDataset implementation. Reads data from npz files and returns them as a dict. Parameters: - path: Path to the directory containing the npz files. - task_names: List of tasks that are present in the dataset. If set to None, will infer from the files on disk. - task_types: A dictionary of form {task_name: task_type} for the reader to call to read from disk, plot etc. - normalization: The normalization type used in __getitem__. Valid options are: - None: Reads the data as-is using task.read_from_disk(path) - 'min_max': Calls task.normalize(task.read_from_disk(path), mins[task], maxs[task]) - 'standardization': Calls task.standardize(task.read_from_disk(path), means[task], stds[task]) If normalization is not 'none', then task-level statistics will be computed. Environmental variable STATS_PBAR=0/1 enables tqdm during statistics computation. - handle_missing_data: Modes to handle missing data. Valid options are: - 'drop': Drop the data point if any of the representations is missing. - 'fill_{none,zero,nan}': Fill the missing data with Nones, zeros or NaNs. - files_suffix: What suffix to look for when creating the dataset. Valid values: 'npy' or 'npz'. - cache_task_stats: If set to True, the statistics will be cached at '{path}/.task_statistics.npz'. Can be enabled using the environmental variable STATS_CACHE=1. Defaults to False. - batch_size_stats: Controls the batch size during statistics computation. Can be enabled by environmental variable STATS_BATCH_SIZE. Defaults to 1. Expected directory structure: path/ - task_1/0.npz, ..., N.npz - ... - task_n/0.npz, ..., N.npz Names can be in a different format (i.e. 2022-01-01.npz), but must be consistent and equal across all tasks. """ def __init__(self, path: Path, task_names: list[str], task_types: dict[str, type], normalization: str | None | dict[str], handle_missing_data: str = "fill_none", files_suffix: str = "npz", cache_task_stats: bool = (os.getenv("STATS_CACHE", "0") == "1"), batch_size_stats: int = int(os.getenv("STATS_BATCH_SIZE", "1")), statistics: dict[str, TaskStatistics] | None = None, ): assert Path(path).exists(), f"Provided path '{path}' doesn't exist!" assert handle_missing_data in ("drop", "fill_none", "fill_zero", "fill_nan"), \ f"Invalid handle_missing_data mode: {handle_missing_data}" assert isinstance(task_names, Iterable), type(task_names) self.path = Path(path).absolute() self.handle_missing_data = handle_missing_data self.suffix = files_suffix self.files_per_repr, self.file_names = self._build_dataset(task_types, task_names) # + handle_missing_data self.cache_task_stats = cache_task_stats self.batch_size_stats = batch_size_stats assert all(isinstance(x, str) for x in task_names), tuple(zip(task_names, (type(x) for x in task_names))) assert (diff := set(self.files_per_repr).difference(task_names)) == set(), f"Not all tasks in files: {diff}" self.task_types = {k: v for k, v in task_types.items() if k in task_names} # all task_types must be provided! self.task_names = sorted(task_names) logger.info(f"Tasks used in this dataset: {self.task_names}") if normalization is not None: if isinstance(normalization, str): logger.info(f"Normalization provided as a string ({normalization}). Setting all tasks to this") normalization: dict[str, str] = {task: normalization for task in self.task_names} if "*" in normalization.keys(): # for the lazy, we can put {"*": "standardization", "depth": "min_max"} value = normalization.pop("*") for missing_task in set(self.task_names).difference(normalization.keys()): normalization[missing_task] = value assert all(n in ("min_max", "standardization") for n in normalization.values()), normalization assert all(k in task_names for k in normalization.keys()), set(normalization).difference(task_names) self.normalization: dict[str, str] | None = normalization self._data_shape: tuple[int, ...] | None = None self._tasks: list[NpzRepresentation] | None = None self._default_vals: dict[str, tr.Tensor] | None = None if statistics is not None: self._statistics = self._load_external_statistics(statistics) else: self._statistics = None if normalization is None else self._compute_statistics() if self._statistics is not None: for task_name, task in self.name_to_task.items(): if not task.is_classification: task.set_normalization(self.normalization[task_name], self._statistics[task_name]) # Public methods and properties @property def name_to_task(self) -> dict[str, NpzRepresentation]: return {task.name: task for task in self.tasks} @property def default_vals(self) -> dict[str, tr.Tensor]: """default values for __getitem__ if item is not on disk but we retrieve a full batch anyway""" _default_val = float("nan") if self.handle_missing_data == "fill_nan" else 0 return {task: None if self.handle_missing_data == "fill_none" else tr.full(self.data_shape[task], _default_val) for task in self.task_names} @property def data_shape(self) -> dict[str, tuple[int, ...]]: """Returns a {task: shape_tuple} for all representations. At least one npz file must exist for each.""" first_npz = {task: [_v for _v in files if _v is not None][0] for task, files in self.files_per_repr.items()} data_shape = {task: self.name_to_task[task].load_from_disk(first_npz[task]).shape for task in self.task_names} return {task: tuple(shape) for task, shape in data_shape.items()} @property def mins(self) -> dict[str, tr.Tensor]: """returns a dict {task: mins[task]} for all the tasks if self.statistics exists""" assert self.normalization is not None, "No statistics for normalization is None" return {k: v[0] for k, v in self._statistics.items() if k in self.task_names} @property def maxs(self) -> dict[str, tr.Tensor]: """returns a dict {task: mins[task]} for all the tasks if self.statistics exists""" assert self.normalization is not None, "No statistics for normalization is None" return {k: v[1] for k, v in self._statistics.items() if k in self.task_names} @property def means(self) -> dict[str, tr.Tensor]: """returns a dict {task: mins[task]} for all the tasks if self.statistics exists""" assert self.normalization is not None, "No statistics for normalization is None" return {k: v[2] for k, v in self._statistics.items() if k in self.task_names} @property def stds(self) -> dict[str, tr.Tensor]: """returns a dict {task: mins[task]} for all the tasks if self.statistics exists""" assert self.normalization is not None, "No statistics for normalization is None" return {k: v[3] for k, v in self._statistics.items() if k in self.task_names} @property def tasks(self) -> list[NpzRepresentation]: """ Returns a list of instantiated tasks in the same order as self.task_names. Overwrite this to add new tasks and semantics (i.e. plot_fn or doing some preprocessing after loading from disk in some tasks. """ if self._tasks is None: self._tasks = [] for task_name in self.task_names: t = self.task_types[task_name] try: t = t(task_name) # hack for not isinstance(self.task_types, NpzRepresentation) but callable except Exception: pass self._tasks.append(t) assert all(t.name == t_n for t, t_n in zip(self._tasks, self.task_names)), (self.task_names, self._tasks) return self._tasks def collate_fn(self, items: list[MultiTaskItem]) -> MultiTaskItem: """ given a list of items (i.e. from a reader[n:n+k] call), return the item batched on 1st dimension. Nones (missing data points) are turned into nans as per the data shape of that dim. """ assert all(item[2] == self.task_names for item in items), ([item[2] for item in items], self.task_names) items_name = [item[1] for item in items] res = {k: tr.zeros(len(items), *self.data_shape[k]).float() for k in self.task_names} # float32 always for i in range(len(items)): for k in self.task_names: try: res[k][i][:] = items[i][0][k] if items[i][0][k] is not None else float("nan") except Exception as e: print(k, items) raise e return res, items_name, self.task_names def add_task(self, task: NpzRepresentation): """Safely adds a task to this reader. Most likely can be optimized""" logger.debug(f"Adding a new task: '{task.name}'") assert task.name not in self.task_names, f"Task '{task.name}' already exists: {self.task_names}" self.task_names = sorted([*self.task_names, task.name]) self.task_types[task.name] = task self._tasks = None self.files_per_repr, self.file_names = self._build_dataset(self.task_types, self.task_names) def remove_task(self, task_name: str): """Safely removes a task from this reader""" logger.debug(f"Removing a task: '{task_name}'") assert task_name in self.task_names, f"Task '{task_name}' doesn't exist: {self.task_names}" self.task_names = sorted(name for name in self.task_names if name != task_name) del self.task_types[task_name] self._tasks = None self.files_per_repr, self.file_names = self._build_dataset(self.task_types, self.task_names) # Private methods def _get_all_npz_files(self) -> dict[str, list[Path]]: """returns a dict of form: {"rgb": ["0.npz", "1.npz", ..., "N.npz"]}""" assert self.suffix == "npz", f"Only npz supported right now (though trivial to update): {self.suffix}" in_files = {} all_repr_dirs: list[str] = [x.name for x in self.path.iterdir() if x.is_dir()] for repr_dir_name in all_repr_dirs: dir_name = self.path / repr_dir_name if all(f.is_dir() for f in dir_name.iterdir()): # dataset is stored as repr/part_x/0.npz, ..., part_k/n.npz all_files = [] for part in dir_name.iterdir(): all_files.extend(part.glob(f"*.{self.suffix}")) else: # dataset is stored as repr/0.npz, ..., repr/n.npz all_files = dir_name.glob(f"*.{self.suffix}") in_files[repr_dir_name] = natsorted(all_files, key=lambda x: x.name) # important: use natsorted() here assert not any(len(x) == 0 for x in in_files.values()), f"{ [k for k, v in in_files.items() if len(v) == 0] }" return in_files def _build_dataset(self, task_types: dict[str, NpzRepresentation], task_names: list[str]) -> BuildDatasetTuple: logger.debug(f"Building dataset from: '{self.path}'") all_npz_files = self._get_all_npz_files() all_files: dict[str, dict[str, Path]] = {k: {_v.name: _v for _v in v} for k, v in all_npz_files.items()} relevant_tasks_for_files = set() # hsv requires only rgb, so we look at dependencies later on for task_name in task_names: relevant_tasks_for_files.update(task_types[task_name].dep_names) if (diff := relevant_tasks_for_files.difference(all_files)) != set(): raise FileNotFoundError(f"Missing files for {diff}.\nFound on disk: {[*all_files]}") names_to_tasks: dict[str, list[str]] = {} # {name: [task]} for task_name in relevant_tasks_for_files: # just the relevant tasks for path_name in all_files[task_name].keys(): names_to_tasks.setdefault(path_name, []) names_to_tasks[path_name].append(task_name) if self.handle_missing_data == "drop": b4 = len(names_to_tasks) names_to_tasks = {k: v for k, v in names_to_tasks.items() if len(v) == len(relevant_tasks_for_files)} logger.debug(f"Dropped {b4 - len(names_to_tasks)} files not in all tasks") all_names: list[str] = natsorted(names_to_tasks.keys()) logger.info(f"Total files: {len(names_to_tasks)} per task across {len(task_names)} tasks") files_per_task: dict[str, list[str | None] | list[list[str] | None]] = {task: [] for task in task_names} for name in all_names: for task in task_names: all_deps_exist = set(deps := task_types[task].dep_names).issubset(names_to_tasks[name]) if not all_deps_exist: files_per_task[task].append(None) # if any of the deps don't exist for this task, skip it. else: paths = [all_files[dep][name] for dep in deps] files_per_task[task].append(paths if len(deps) > 1 else paths[0]) return files_per_task, all_names def _compute_statistics(self) -> dict[str, TaskStatistics]: cache_path = self.path / f".task_statistics.npz" res: dict[str, TaskStatistics] = {} if self.cache_task_stats and cache_path.exists(): res = np.load(cache_path, allow_pickle=True)["arr_0"].item() logger.info(f"Loaded task statistics: { {k: tuple(v[0].shape) for k, v in res.items()} } from {cache_path}") missing_tasks = [t for t in set(self.task_names).difference(res) if not self.name_to_task[t].is_classification] if len(missing_tasks) == 0: return res logger.info(f"Computing global task statistics (dataset len {len(self)}) for {missing_tasks}") res = {**res, **self._compute_channel_level_stats(missing_tasks)} logger.info(f"Computed task statistics: { {k: tuple(v[0].shape) for k, v in res.items()} }") np.savez(cache_path, res) return res def _compute_channel_level_stats(self, missing_tasks: list[str]) -> dict[str, TaskStatistics]: # kinda based on: https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Welford's_online_algorithm def update(counts: tr.Tensor, counts_delta: float, mean: tr.Tensor, M2: tr.Tensor, new_value: tr.Tensor) -> tuple[tr.Tensor, tr.Tensor, tr.Tensor]: new_count = counts + counts_delta batch_mean = new_value.nanmean(0) batch_var = ((new_value - batch_mean) ** 2).nansum(0) delta = batch_mean - mean new_count_no_zero = new_count + (new_count == 0) # add 1 (True) in case new_count is 0 to not divide by 0 new_mean = mean + delta * counts_delta / new_count_no_zero new_M2 = M2 + batch_var + delta**2 * counts * counts_delta / new_count_no_zero assert not new_mean.isnan().any() and not new_M2.isnan().any(), (mean, new_mean, counts, counts_delta) return new_count, new_mean, new_M2 assert not any(mt := [self.name_to_task[t].is_classification for t in missing_tasks]), mt assert len(missing_tasks) > 0, missing_tasks ch = {k: v[-1] if len(v) == 3 else 1 for k, v in self.data_shape.items()} counts = {task_name: tr.zeros(ch[task_name]).long() for task_name in missing_tasks} mins = {task_name: tr.zeros(ch[task_name]).type(tr.float64) + 10**10 for task_name in missing_tasks} maxs = {task_name: tr.zeros(ch[task_name]).type(tr.float64) - 10**10 for task_name in missing_tasks} means_vec = {task_name: tr.zeros(ch[task_name]).type(tr.float64) for task_name in missing_tasks} M2s_vec = {task_name: tr.zeros(ch[task_name]).type(tr.float64) for task_name in missing_tasks} old_names, old_normalization = self.task_names, self.normalization self.task_names, self.normalization = missing_tasks, None # for self[ix] res = {} BS = min(len(self), self.batch_size_stats) n = (len(self) // BS) + (len(self) % BS != 0) logger.debug(f"Global task statistics. Batch size: {BS}. N iterations: {n}.") for ix in trange(n, disable=os.getenv("STATS_PBAR", "0") == "0", desc="Computing stats"): item = self[ix * BS: min(len(self), (ix + 1) * BS)][0] for task in missing_tasks: item_flat_ch = item[task].reshape(-1, ch[task]) item_no_nan = item_flat_ch.nan_to_num(0).type(tr.float64) mins[task] = tr.minimum(mins[task], item_no_nan.min(0)[0]) maxs[task] = tr.maximum(maxs[task], item_no_nan.max(0)[0]) counts_delta = (item_flat_ch == item_flat_ch).long().sum(0) counts[task], means_vec[task], M2s_vec[task] = \ update(counts[task], counts_delta, means_vec[task], M2s_vec[task], item_no_nan) for task in missing_tasks: res[task] = (mins[task], maxs[task], means_vec[task], (M2s_vec[task] / counts[task]).sqrt()) assert not any(x[0].isnan().any() for x in res[task]), (task, res[task]) self.task_names, self.normalization = old_names, old_normalization return res def _load_external_statistics(self, statistics: dict[str, TaskStatistics | list]) -> dict[str, TaskStatistics]: tasks_no_classif = [t for t in set(self.task_names) if not self.name_to_task[t].is_classification] assert (diff := set(tasks_no_classif).difference(statistics)) == set(), f"Missing tasks: {diff}" res: dict[str, TaskStatistics] = {} for k, v in statistics.items(): if k in self.task_names: res[k] = tuple(tr.Tensor(x) for x in v) assert all(_stat.shape == (nd := (self.name_to_task[k].n_channels, )) for _stat in res[k]), (res[k], nd) logger.info(f"External statistics provided: { {k: tuple(v[0].shape) for k, v in res.items()} }") return res # Python magic methods (pretty printing the reader object, reader[0], len(reader) etc.) def __getitem__(self, index: int | slice | list[int] | tuple) -> MultiTaskItem: """Read the data all the desired nodes""" assert isinstance(index, (int, slice, list, tuple, str)), type(index) if isinstance(index, slice): assert index.start is not None and index.stop is not None and index.step is None, "Only reader[l:r] allowed" index = list(range(index.stop)[index]) if isinstance(index, (list, tuple)): return self.collate_fn([self.__getitem__(ix) for ix in index]) if isinstance(index, str): return self.__getitem__(self.file_names.index(index)) res: dict[str, tr.Tensor] = {} for task_name in self.task_names: task = [t for t in self.tasks if t.name == task_name][0] file_path = self.files_per_repr[task_name][index] res[task_name] = self.default_vals[task_name] if file_path is None else task.load_from_disk(file_path) if not task.is_classification: if self.normalization is not None and self.normalization[task_name] == "min_max": res[task_name] = task.normalize(res[task_name]) if self.normalization is not None and self.normalization[task_name] == "standardization": res[task_name] = task.standardize(res[task_name]) return (res, self.file_names[index], self.task_names) def __len__(self) -> int: return len(self.files_per_repr[self.task_names[0]]) # all of them have the same number (filled with None or not) def __str__(self): f_str = f"[{str(type(self)).rsplit('.', maxsplit=1)[-1][0:-2]}]" f_str += f"\n - Path: '{self.path}'" f_str += f"\n - Tasks ({len(self.tasks)}): {self.tasks}" f_str += f"\n - Length: {len(self)}" f_str += f"\n - Handle missing data mode: '{self.handle_missing_data}'" f_str += f"\n - Normalization: '{self.normalization}'" return f_str def __repr__(self): return str(self) def main(): """main fn""" parser = ArgumentParser() parser.add_argument("dataset_path", type=Path) parser.add_argument("--handle_missing_data", choices=("drop", "fill_none"), default="fill_none") args = parser.parse_args() reader = MultiTaskDataset(args.dataset_path, task_names=None, handle_missing_data=args.handle_missing_data) print(reader) print(f"Shape: {reader.data_shape}") rand_ix = np.random.randint(len(reader)) data, name, repr_names = reader[rand_ix] # get a random single data point print(f"Name: {name}. Nodes: {repr_names}") pprint({k: v for k, v in data.items()}) data, name, repr_names = reader[rand_ix: min(len(reader), rand_ix + 5)] # get a random batch print(f"Name: {name}. Nodes: {repr_names}") pprint({k: v for k, v in data.items()}) # Nones are converted to 0s automagically loader = DataLoader(reader, collate_fn=reader.collate_fn, batch_size=5, shuffle=True) data, name, repr_names = next(iter(loader)) # get a random batch using torch DataLoader print(f"Name: {name}. Nodes: {repr_names}") pprint({k: v for k, v in data.items()}) # Nones are converted to 0s automagically if __name__ == "__main__": main()