Spaces:
Running
on
A10G
Running
on
A10G
File size: 6,287 Bytes
45d16e9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 |
"""
Copyright (c) 2022, salesforce.com, inc.
All rights reserved.
SPDX-License-Identifier: BSD-3-Clause
For full license text, see the LICENSE_Lavis file in the repo root or https://opensource.org/licenses/BSD-3-Clause
"""
import gzip
import logging
import os
import random as rnd
import tarfile
import zipfile
import random
from typing import List
from tqdm import tqdm
import decord
from decord import VideoReader
import webdataset as wds
import numpy as np
import torch
from torch.utils.data.dataset import IterableDataset
from video_llama.common.registry import registry
from video_llama.datasets.datasets.base_dataset import ConcatDataset
decord.bridge.set_bridge("torch")
MAX_INT = registry.get("MAX_INT")
class ChainDataset(wds.DataPipeline):
r"""Dataset for chaining multiple :class:`DataPipeline` s.
This class is useful to assemble different existing dataset streams. The
chaining operation is done on-the-fly, so concatenating large-scale
datasets with this class will be efficient.
Args:
datasets (iterable of IterableDataset): datasets to be chained together
"""
def __init__(self, datasets: List[wds.DataPipeline]) -> None:
super().__init__()
self.datasets = datasets
self.prob = []
self.names = []
for dataset in self.datasets:
if hasattr(dataset, 'name'):
self.names.append(dataset.name)
else:
self.names.append('Unknown')
if hasattr(dataset, 'sample_ratio'):
self.prob.append(dataset.sample_ratio)
else:
self.prob.append(1)
logging.info("One of the datapipeline doesn't define ratio and set to 1 automatically.")
def __iter__(self):
datastreams = [iter(dataset) for dataset in self.datasets]
while True:
select_datastream = random.choices(datastreams, weights=self.prob, k=1)[0]
yield next(select_datastream)
def apply_to_sample(f, sample):
if len(sample) == 0:
return {}
def _apply(x):
if torch.is_tensor(x):
return f(x)
elif isinstance(x, dict):
return {key: _apply(value) for key, value in x.items()}
elif isinstance(x, list):
return [_apply(x) for x in x]
else:
return x
return _apply(sample)
def move_to_cuda(sample):
def _move_to_cuda(tensor):
return tensor.cuda()
return apply_to_sample(_move_to_cuda, sample)
def prepare_sample(samples, cuda_enabled=True):
if cuda_enabled:
samples = move_to_cuda(samples)
# TODO fp16 support
return samples
def reorg_datasets_by_split(datasets):
"""
Organizes datasets by split.
Args:
datasets: dict of torch.utils.data.Dataset objects by name.
Returns:
Dict of datasets by split {split_name: List[Datasets]}.
"""
# if len(datasets) == 1:
# return datasets[list(datasets.keys())[0]]
# else:
reorg_datasets = dict()
# reorganize by split
for _, dataset in datasets.items():
for split_name, dataset_split in dataset.items():
if split_name not in reorg_datasets:
reorg_datasets[split_name] = [dataset_split]
else:
reorg_datasets[split_name].append(dataset_split)
return reorg_datasets
def concat_datasets(datasets):
"""
Concatenates multiple datasets into a single dataset.
It supports may-style datasets and DataPipeline from WebDataset. Currently, does not support
generic IterableDataset because it requires creating separate samplers.
Now only supports conctenating training datasets and assuming validation and testing
have only a single dataset. This is because metrics should not be computed on the concatenated
datasets.
Args:
datasets: dict of torch.utils.data.Dataset objects by split.
Returns:
Dict of concatenated datasets by split, "train" is the concatenation of multiple datasets,
"val" and "test" remain the same.
If the input training datasets contain both map-style and DataPipeline datasets, returns
a tuple, where the first element is a concatenated map-style dataset and the second
element is a chained DataPipeline dataset.
"""
# concatenate datasets in the same split
for split_name in datasets:
if split_name != "train":
assert (
len(datasets[split_name]) == 1
), "Do not support multiple {} datasets.".format(split_name)
datasets[split_name] = datasets[split_name][0]
else:
iterable_datasets, map_datasets = [], []
for dataset in datasets[split_name]:
if isinstance(dataset, wds.DataPipeline):
logging.info(
"Dataset {} is IterableDataset, can't be concatenated.".format(
dataset
)
)
iterable_datasets.append(dataset)
elif isinstance(dataset, IterableDataset):
raise NotImplementedError(
"Do not support concatenation of generic IterableDataset."
)
else:
map_datasets.append(dataset)
# if len(iterable_datasets) > 0:
# concatenate map-style datasets and iterable-style datasets separately
if len(iterable_datasets) > 1:
chained_datasets = (
ChainDataset(iterable_datasets)
)
elif len(iterable_datasets) == 1:
chained_datasets = iterable_datasets[0]
else:
chained_datasets = None
concat_datasets = (
ConcatDataset(map_datasets) if len(map_datasets) > 0 else None
)
train_datasets = concat_datasets, chained_datasets
train_datasets = tuple([x for x in train_datasets if x is not None])
train_datasets = (
train_datasets[0] if len(train_datasets) == 1 else train_datasets
)
datasets[split_name] = train_datasets
return datasets
|