Spaces:
Running on Zero
Running on Zero
File size: 7,475 Bytes
d1f1097 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 | # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
import os
import random
import sys
import tempfile
import nltk
import numpy
from megatron.core.datasets.indexed_dataset import IndexedDataset
from tests.unit_tests.data.test_preprocess_data import dummy_jsonl, gpt2_merge, gpt2_vocab
from tools.merge_datasets import main as merge_main
from tools.preprocess_mmdata import Encoder
from tools.preprocess_mmdata import get_args as build_args
from tools.preprocess_mmdata import main as build_main
def dummy_img(odir_txt, odir_img):
for name in os.listdir(odir_txt):
with open(os.path.join(odir_txt, name), "rt") as reader_txt:
length = sum(1 for _ in reader_txt)
os.makedirs(os.path.join(odir_img, os.path.splitext(name)[0]), exist_ok=False)
for i in range(length):
with open(
os.path.join(odir_img, os.path.splitext(name)[0], f"{str(i).zfill(4)}.img"), "wb"
) as writer_img:
# 32 * 32 - 1 to induce preprocessing 0-index padding
writer_img.write(bytes([random.randint(0, 255) for _ in range(32 * 32 - 1)]))
def build_datasets(idir_txt, idir_img, odir, extra_args=[]):
for name in os.listdir(idir_txt):
sys.argv = [
sys.argv[0],
"--input",
os.path.join(idir_txt, name),
"--input-image",
os.path.join(idir_img, os.path.splitext(name)[0]),
"--output-prefix",
os.path.join(odir, os.path.splitext(name)[0]),
] + extra_args
build_main()
def merge_datasets(idir):
sys.argv = [
sys.argv[0],
"--input",
idir,
"--output-prefix",
os.path.join(idir, "merge"),
"--multimodal",
]
merge_main()
def do_test_preprocess_mmdata(temp_dir, extra_args=[]):
# set the default nltk data path
os.environ["NLTK_DATA"] = os.path.join(temp_dir, "nltk_data")
nltk.data.path.append(os.environ["NLTK_DATA"])
path_to_raws_txt = os.path.join(temp_dir, "sample_raws_txt")
path_to_raws_img = os.path.join(temp_dir, "sample_raws_img")
path_to_data = os.path.join(temp_dir, "sample_data")
os.mkdir(path_to_raws_txt)
os.mkdir(path_to_raws_img)
os.mkdir(path_to_data)
# create the dummy text resources
dummy_jsonl(path_to_raws_txt)
# create the dummy image resources
dummy_img(path_to_raws_txt, path_to_raws_img)
# build the datasets
build_datasets(path_to_raws_txt, path_to_raws_img, path_to_data, extra_args=extra_args)
# merge the datasets
merge_datasets(path_to_data)
sys.argv = [
sys.argv[0],
"--input",
None,
"--input-image",
None,
"--output-prefix",
None,
] + extra_args
encoder = Encoder(build_args())
encoder.initializer()
def tokens_to_string(toks):
for option in ["decode", "detokenize"]:
try:
return getattr(encoder.tokenizer, option)(toks)
except AttributeError:
continue
raise RuntimeError(f"{type(encoder.tokenizer)} tokenizer cannot `decode` or `detokenize`.")
merged_index = 0
merged_dataset = IndexedDataset(os.path.join(path_to_data, "merge"), multimodal=True)
# sorted to ensure ordering matches merged dataset
basenames = sorted(
[
name
for name in os.listdir(path_to_data)
if name.endswith(".idx") and not name.startswith("merge")
]
)
# index into the merged document index
merged_doc_index_index = 0
for basename in basenames:
realpath_raw_txt = os.path.join(path_to_raws_txt, f"{os.path.splitext(basename)[0]}.jsonl")
realpath_raw_img = os.path.join(path_to_raws_img, os.path.splitext(basename)[0])
realpath_doc = os.path.join(path_to_data, os.path.splitext(basename)[0])
dataset_index = 0
dataset = IndexedDataset(realpath_doc, multimodal=True)
merged_doc_idx = merged_dataset.document_indices[
merged_doc_index_index : merged_doc_index_index + len(dataset.document_indices)
]
merged_doc_idx = merged_doc_idx - merged_doc_idx[0]
assert (
dataset.document_indices == merged_doc_idx
).all(), f"ERROR: {basename.split('_')[:-2]}: merged dataset document indices mismatch"
merged_doc_index_index += len(dataset.document_indices) - 1
with open(realpath_raw_txt, "rt") as reader:
for json_line, image_path in zip(
reader,
[
os.path.join(realpath_raw_img, basename)
for basename in os.listdir(realpath_raw_img)
],
):
toks, image, length = encoder.encode((json_line, image_path))
raw_text = tokens_to_string(toks)
# reverse to account for preprocessing 0-index padding
raw_image = image[::-1]
processed_toks = dataset[dataset_index][0]
assert dataset[dataset_index][1] == 0
processed_text = tokens_to_string(processed_toks)
processed_image = dataset[dataset_index + 1][0]
assert dataset[dataset_index + 1][1] == 1
# reverse to account for preprocessing 0-index padding
processed_image = processed_image[::-1][0 : raw_image.size]
assert (
raw_text == processed_text
), f"ERROR: {basename.split('_')[:-2]}: raw and processed documents (text) do not match"
assert numpy.allclose(
raw_image, processed_image
), f"ERROR: {basename.split('_')[:-2]}: raw and processed documents (image) do not match"
dataset_index += 2
merged_toks = merged_dataset[merged_index][0]
assert merged_dataset[merged_index][1] == 0
merged_text = tokens_to_string(merged_toks)
merged_image = merged_dataset[merged_index + 1][0]
assert merged_dataset[merged_index + 1][1] == 1
# reverse to account for preprocessing 0-index padding
merged_image = merged_image[::-1][0 : raw_image.size]
assert (
raw_text == merged_text
), f"ERROR: {basename.split('_')[:-2]}: raw and merged documents (text) do not match"
assert numpy.allclose(
raw_image, merged_image
), f"ERROR: {basename.split('_')[:-2]}: raw and merged documents (image) do not match"
merged_index += 2
print(
f"INFO: {''.join(basename.split('_')[:-2])}: raw, processed, and merged documents match!"
)
print("INFO: Success!")
def test_preprocess_mmdata():
with tempfile.TemporaryDirectory() as temp_dir:
# gpt specific args
gpt_args = [
"--pad-length",
"1024",
"--tokenizer-type",
"GPT2BPETokenizer",
"--vocab-file",
gpt2_vocab(temp_dir),
"--merge-file",
gpt2_merge(temp_dir),
"--append-eod",
"--workers",
"10",
"--log-interval",
"1",
]
do_test_preprocess_mmdata(temp_dir, extra_args=gpt_args)
if __name__ == "__main__":
test_preprocess_mmdata()
|