offline_dataset.py

import gzip
from pathlib import Path
import re
from typing import List, Tuple

import numpy as np
from rlpyt.utils.collections import namedarraytuple
import torch
from torch.utils.data import DataLoader, Dataset
import os

from itertools import zip_longest
from .rlpyt_atari_env import AtariEnv
from src.utils import discount_return_n_step


OfflineSamples = namedarraytuple("OfflineSamples", ["all_observation", "all_action", "all_reward",  "return_", "done", "done_n", "init_rnn_state", "is_weights"])

class DQNReplayDataset(Dataset):
  def __init__(self, data_path: Path,
               tmp_data_path: Path,
               game: str,
               checkpoint: int,
               frames: int,
               k_step: int,
               max_size: int,
               full_action_set: bool,
               dataset_on_gpu: bool,
               dataset_on_disk: bool,
               load_reward: bool = False) -> None:
    data = []
    self.dataset_on_disk = dataset_on_disk
    self.load_reward = load_reward
    assert not (dataset_on_disk and dataset_on_gpu)
    filetypes = ['reward', 'action', 'terminal', 'observation']
    if not load_reward:
        filetypes = filetypes[1:]
    for i, filetype in enumerate(filetypes):
      filename = Path(data_path / f'{game}/{filetype}_{checkpoint}.gz')
      print(f'Loading {filename}')

      # There's no point in putting rewards, actions or terminals on disk.
      # They're tiny and it'll just cause more I/O.
      on_disk = dataset_on_disk and filetype == "observation"

      g = gzip.GzipFile(filename=filename)
      data__ = np.load(g)
      if i == 0:
        self.has_parallel_envs = len(data__.shape) > 1
        if self.has_parallel_envs:
            self.n_envs = data__.shape[1]
        else:
            self.n_envs = 1
      if not self.has_parallel_envs:
        data__ = np.expand_dims(data__, 1)

      data___ = np.copy(data__[:max_size])
      print(f'Using {data___.size * data___.itemsize} bytes')
      if not on_disk:
        del data__
        data_ = torch.from_numpy(data___)
      else:
        new_filename = os.path.join(tmp_data_path, Path(os.path.basename(filename)[:-3]+".npy"))
        print("Stored on disk at {}".format(new_filename))
        np.save(new_filename, data___,)
        del data___
        del data__
        data_ = np.load(new_filename, mmap_mode="r+")

      if (filetype == 'action') and full_action_set:
        action_mapping = dict(zip(data_.unique().numpy(),
                                  AtariEnv(re.sub(r'(?<!^)(?=[A-Z])', '_', game).lower()).ale.getMinimalActionSet()))
        data_.apply_(lambda x: action_mapping[x])
      if dataset_on_gpu:
        print("Stored on GPU")
        data_ = data_.cuda(non_blocking=True)
        del data___
      data.append(data_)
      setattr(self, filetype, data_)

    self.game = game
    self.f = frames
    self.k = k_step
    self.size = min(self.action.shape[0], max_size)
    self.effective_size = (self.size - self.f - self.k + 1)

  def __len__(self) -> int:
    return self.effective_size*self.n_envs

  def __getitem__(self, index: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    batch_ind = index // self.effective_size
    time_ind = index % self.effective_size
    sl = slice(time_ind, time_ind+self.f+self.k)
    if self.dataset_on_disk:
        obs = torch.from_numpy(self.observation[sl, batch_ind])
    else:
        obs = (self.observation[sl, batch_ind])

    # Buffer reading code still expects us to return rewards even when not used
    if self.load_reward:
        rewards = self.reward[sl, batch_ind]
    else:
        rewards = torch.zeros_like(self.terminal[sl, batch_ind]).float()
    return tuple([obs,
                  self.action[sl, batch_ind],
                  rewards,
                  self.terminal[sl, batch_ind],
                  ])


class MultiDQNReplayDataset(Dataset):
  def __init__(self, data_path: Path,
               tmp_data_path:
               Path, games: List[str],
               checkpoints: List[int],
               frames: int,
               k_step: int,
               max_size: int,
               full_action_set: bool,
               dataset_on_gpu: bool,
               dataset_on_disk: bool) -> None:
    self.games = [DQNReplayDataset(data_path,
                                   tmp_data_path,
                                   game,
                                   ckpt,
                                   frames,
                                   k_step,
                                   max_size,
                                   full_action_set,
                                   dataset_on_gpu,
                                   dataset_on_disk) for ckpt in checkpoints for game in games]

    self.num_blocks = len(self.games)
    self.block_len = len(self.games[0])

  def __len__(self) -> int:
    return len(self.games) * len(self.games[0])

  def __getitem__(self, index: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    game_index = index % len(self.games)
    index = index // len(self.games)
    return self.games[game_index][index]


def sanitize_batch(batch: OfflineSamples) -> OfflineSamples:
    has_dones, inds = torch.max(batch.done, 0)
    for i, (has_done, ind) in enumerate(zip(has_dones, inds)):
        if not has_done:
            continue
        batch.all_observation[ind+1:, i] = batch.all_observation[ind, i]
        batch.all_reward[ind+1:, i] = 0
        batch.return_[ind+1:, i] = 0
        batch.done_n[ind+1:, i] = True
    return batch


def get_offline_dataloaders(
  *,
  data_path: Path,
  tmp_data_path: Path,
  games: List[str],
  checkpoints: List[int],
  frames: int,
  k_step: int,
  n_step_return: int,
  discount: float,
  samples: int,
  dataset_on_gpu: bool,
  dataset_on_disk: bool,
  batch_size: int,
  full_action_set: bool,
  num_workers: int,
  pin_memory: bool,
  prefetch_factor: int,
  group_read_factor: int=0,
  shuffle_checkpoints: bool=False,
  **kwargs,
) -> Tuple[DataLoader, DataLoader, DataLoader]:
  def collate(batch):
    observation, action, reward, done = torch.utils.data.dataloader.default_collate(batch)
    observation = torch.einsum('bthw->tbhw', observation).unsqueeze(2).repeat(1, 1, frames, 1, 1)
    for i in range(1, frames):
        observation[:, :, i] = observation[:, :, i].roll(-i, 0)
    observation = observation[:-frames].unsqueeze(3) # tbfchw
    action = torch.einsum('bt->tb', action)[frames-2:-2].long()
    reward = torch.einsum('bt->tb', reward)[frames-2:-2]
    reward = torch.nan_to_num(reward).sign()  # Apparently possible, somehow.
    done = torch.einsum('bt->tb', done)[frames:].bool()
    return_, done_n = discount_return_n_step(reward[1:], done, n_step_return, discount)
    is_weights = torch.ones(observation.shape[1]).to(reward)
    return sanitize_batch(OfflineSamples(observation, action, reward, return_, done[:-n_step_return], done_n, None, is_weights))

  dataset = MultiDQNReplayDataset(data_path, tmp_data_path, games, checkpoints, frames, k_step, samples, full_action_set, dataset_on_gpu, dataset_on_disk)

  if shuffle_checkpoints:
      data = get_from_dataloaders(dataset.games)
      shuffled_data = shuffle_batch_dim(*data)
      assign_to_dataloaders(dataset.games, *shuffled_data)

  if group_read_factor != 0:
      sampler = CacheEfficientSampler(dataset.num_blocks, dataset.block_len, group_read_factor)
      dataloader = DataLoader(dataset, batch_size=batch_size,
                              sampler=sampler,
                              num_workers=num_workers,
                              pin_memory=pin_memory,
                              collate_fn=collate,
                              drop_last=True,
                              prefetch_factor=prefetch_factor)
  else:
    dataloader = DataLoader(dataset, batch_size=batch_size,
                            shuffle=True,
                            num_workers=num_workers,
                            pin_memory=pin_memory,
                            collate_fn=collate,
                            drop_last=True,
                            prefetch_factor=prefetch_factor)

  return dataloader, None, None


class CacheEfficientSampler(torch.utils.data.Sampler):
    def __init__(self, num_blocks, block_len, num_repeats=20, generator=None):
        self.num_blocks = num_blocks
        self.block_len = block_len  # For now, assume all have same length
        self.num_repeats = num_repeats
        self.generator = generator
        if self.num_repeats == "all":
            self.num_repeats = block_len

    def num_samples(self) -> int:
        # dataset size might change at runtime
        return self.block_len * self.num_blocks

    def __iter__(self):
        n = self.num_samples()
        if self.generator is None:
            generator = torch.Generator()
            generator.manual_seed(int(torch.empty((), dtype=torch.int64).random_().item()))
        else:
            generator = self.generator

        self.block_ids = [np.arange(self.num_blocks)] * (self.block_len // self.num_repeats)
        blocks = torch.randperm(n // self.num_repeats, generator=generator) % self.num_blocks
        intra_orders = [torch.randperm(self.block_len, generator=generator) + self.block_len * i for i in
                        range(self.num_blocks)]
        intra_orders = [i.tolist() for i in intra_orders]

        indices = []
        block_counts = [0] * self.num_blocks

        for block in blocks:
            indices += intra_orders[block][
                       (block_counts[block] * self.num_repeats):(block_counts[block] + 1) * self.num_repeats]
            block_counts[block] += 1

        return iter(indices)

    def __len__(self):
        return self.num_samples()


def shuffle_by_trajectory():
    raise NotImplementedError


def grouper(iterable, n, fillvalue=None):
    "Collect data into fixed-length chunks or blocks"
    # grouper('ABCDEFG', 3, 'x') --> ABC DEF Gxx"
    args = [iter(iterable)] * n
    return zip_longest(*args, fillvalue=fillvalue)


def shuffle_batch_dim(observations,
                      rewards,
                      actions,
                      dones,
                      obs_on_disk=True,
                      chunk_num=1,
                      ):
    """
    :param observations: (T, B, *) obs tensor, optionally mmap
    :param rewards: (T, B) rewards tensor
    :param actions: (T, B, *) actions tensor
    :param dones: (T, B) termination tensor
    :param obs_on_disk: Store observations on disk.  Generally true if using
    more than ~3M transitions
    :return:
    """
    batch_dim = observations[0].shape[1]
    num_sources = len(observations)
    batch_allocations = [np.sort((np.arange(batch_dim) + i) % num_sources) for i in range(num_sources)]

    shuffled_observations, shuffled_rewards, shuffled_actions, shuffled_dones = [], [], [], []

    checkpoints = list(range(num_sources))
    for sources, shuffled, filetype in zip([observations, rewards, actions, dones],
                                           [shuffled_observations, shuffled_rewards, shuffled_actions, shuffled_dones],
                                           ["observations", "rewards", "actions", "dones"]):
        ind_counters = [0]*num_sources

        for start in checkpoints[::chunk_num]:
            chunk = checkpoints[start:start+chunk_num]
            chunk_arrays = []
            for i in chunk:
                if isinstance(sources[0], torch.Tensor):
                    new_array = torch.zeros_like(sources[0])
                else:
                    new_array = np.zeros(sources[0].shape, dtype=sources[0].dtype)
                chunk_arrays.append(new_array)
            for source, allocation in zip(sources, batch_allocations):
                print(chunk, ind_counters)
                for i, new_array in zip(chunk, chunk_arrays):
                    mapped_to_us = [b for b, dest in enumerate(allocation) if dest == i]
                    new_array[:, ind_counters[i]:ind_counters[i]+len(mapped_to_us)] = source[:, mapped_to_us[0]:mapped_to_us[-1]+1]
                    ind_counters[i] += len(mapped_to_us)

            for i, new_array in zip(chunk, chunk_arrays):
                if filetype == "observations" and obs_on_disk:
                    filename = observations[i].filename.replace(".npy", "_shuffled.npy")
                    print("Stored shuffled obs on disk at {}".format(filename))
                    np.save(filename, new_array)
                    del new_array
                    new_array = np.load(filename, mmap_mode="r+")
                shuffled.append(new_array)

    return shuffled_observations, shuffled_rewards, shuffled_actions, shuffled_dones


def get_from_dataloaders(dataloaders):
    observations = [dataloader.observations for dataloader in dataloaders]
    rewards = [dataloader.rewards for dataloader in dataloaders]
    actions = [dataloader.actions for dataloader in dataloaders]
    dones = [dataloader.terminal for dataloader in dataloaders]

    return observations, rewards, actions, dones


def assign_to_dataloaders(dataloaders, observations, rewards, actions, dones):
    for dl, obs, rew, act, done in zip(dataloaders, observations, rewards, actions, dones):
        dl.observations = obs
        dl.rewards = rew
        dl.actions = act
        dl.terminal = done