Spaces:
Sleeping
Sleeping
File size: 10,228 Bytes
e45d058 |
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 |
import os
from pathlib import Path
current_dir = Path(__file__).parent.absolute()
import pytest
import torch
import dotenv
from src.datamodules.language_modeling_hf import LMDataModule
# load environment variables from `.env` file if it exists
# recursively searches for `.env` in all folders starting from work dir
dotenv.load_dotenv(override=True)
def div_up(x: int, y: int) -> int:
return (x + y - 1) // y
# https://stackoverflow.com/questions/1006289/how-to-find-out-the-number-of-cpus-using-python/55423170#55423170
def num_cpu_cores():
try:
import psutil
return psutil.cpu_count(logical=False)
except ImportError:
return len(os.sched_getaffinity(0))
class TestLMDataModule:
def test_wikitext2(self):
batch_size = 7
dataset_name = 'wikitext'
dataset_config_name = 'wikitext-2-raw-v1'
data_dir = Path(os.getenv('DATA_DIR', current_dir.parent.parent / 'data'))
cache_dir = data_dir / 'wikitext-2' / 'cache'
max_length = 1024
datamodule = LMDataModule(dataset_name, tokenizer_name='gpt2',
dataset_config_name=dataset_config_name,
max_length=max_length, cache_dir=cache_dir,
add_eos=False, batch_size=batch_size, num_workers=4)
datamodule.prepare_data()
datamodule.setup(stage='fit')
train_loader = datamodule.train_dataloader()
val_loader = datamodule.val_dataloader()
datamodule.setup(stage='test')
test_loader = datamodule.test_dataloader()
train_len = 2391884
val_len = 247289
test_len = 283287
assert len(train_loader) == div_up((train_len - 1) // max_length, batch_size)
assert len(val_loader) == div_up((val_len - 1) // max_length, batch_size)
assert len(test_loader) == div_up((test_len - 1) // max_length, batch_size)
for loader in [train_loader, val_loader, test_loader]:
x, y = next(iter(loader))
assert x.dim() == 2
assert x.shape == (batch_size, max_length)
assert x.dtype == torch.long
assert torch.allclose(x[:, 1:], y[:, :-1])
def test_wikitext103(self):
batch_size = 7
dataset_name = 'wikitext'
dataset_config_name = 'wikitext-103-raw-v1'
data_dir = Path(os.getenv('DATA_DIR', current_dir.parent.parent / 'data'))
cache_dir = data_dir / 'wikitext-103' / 'cache'
max_length = 1024
datamodule = LMDataModule(dataset_name, tokenizer_name='gpt2',
dataset_config_name=dataset_config_name,
max_length=max_length, cache_dir=cache_dir,
add_eos=False, batch_size=batch_size, num_workers=4)
datamodule.prepare_data()
datamodule.setup(stage='fit')
train_loader = datamodule.train_dataloader()
val_loader = datamodule.val_dataloader()
datamodule.setup(stage='test')
test_loader = datamodule.test_dataloader()
train_len = 117920140
val_len = 247289
test_len = 283287
assert len(train_loader) == div_up((train_len - 1) // max_length, batch_size)
assert len(val_loader) == div_up((val_len - 1) // max_length, batch_size)
assert len(test_loader) == div_up((test_len - 1) // max_length, batch_size)
for loader in [train_loader, val_loader, test_loader]:
x, y = next(iter(loader))
assert x.dim() == 2
assert x.shape == (batch_size, max_length)
assert x.dtype == torch.long
assert torch.allclose(x[:, 1:], y[:, :-1])
def test_openwebtext(self):
batch_size = 8
dataset_name = 'openwebtext'
dataset_config_name = None
data_dir = Path(os.getenv('DATA_DIR', current_dir.parent.parent / 'data'))
cache_dir = data_dir / 'openwebtext' / 'cache'
max_length = 1024
datamodule = LMDataModule(dataset_name, tokenizer_name='gpt2',
dataset_config_name=dataset_config_name,
max_length=max_length, cache_dir=cache_dir,
add_eos=True, batch_size=batch_size,
num_workers=num_cpu_cores() // 2)
datamodule.prepare_data()
datamodule.setup(stage='fit')
train_loader = datamodule.train_dataloader()
val_loader = datamodule.val_dataloader()
datamodule.setup(stage='test')
test_loader = datamodule.test_dataloader()
train_len = 9035582198
val_len = 4434897
test_len = 4434897
assert len(train_loader) == div_up((train_len - 1) // max_length, batch_size)
assert len(val_loader) == div_up((val_len - 1) // max_length, batch_size)
assert len(test_loader) == div_up((test_len - 1) // max_length, batch_size)
for loader in [train_loader, val_loader, test_loader]:
x, y = next(iter(loader))
assert x.dim() == 2
assert x.shape == (batch_size, max_length)
assert x.dtype == torch.long
assert torch.allclose(x[:, 1:], y[:, :-1])
def test_lambada(self):
batch_size = 8
dataset_name = 'lambada'
dataset_config_name = None
data_dir = Path(os.getenv('DATA_DIR', current_dir.parent.parent / 'data'))
cache_dir = data_dir / 'lambada' / 'cache'
max_length = 1024
datamodule = LMDataModule(dataset_name, tokenizer_name='gpt2',
dataset_config_name=dataset_config_name,
max_length=max_length, cache_dir=cache_dir,
add_eos=True, batch_size=batch_size,
num_workers=64)
datamodule.prepare_data()
datamodule.setup(stage='fit')
train_loader = datamodule.train_dataloader()
val_loader = datamodule.val_dataloader()
datamodule.setup(stage='test')
test_loader = datamodule.test_dataloader()
train_len = 9035582198
val_len = 4434897
test_len = 4434897
assert len(train_loader) == div_up((train_len - 1) // max_length, batch_size)
assert len(val_loader) == div_up((val_len - 1) // max_length, batch_size)
assert len(test_loader) == div_up((test_len - 1) // max_length, batch_size)
for loader in [train_loader, val_loader, test_loader]:
x, y = next(iter(loader))
assert x.dim() == 2
assert x.shape == (batch_size, max_length)
assert x.dtype == torch.long
assert torch.allclose(x[:, 1:], y[:, :-1])
def test_the_pile(self):
batch_size = 8
dataset_name = 'the_pile'
dataset_config_name = None
data_dir = Path(os.getenv('DATA_DIR', current_dir.parent.parent / 'data'))
cache_dir = data_dir / 'the_pile' / 'cache'
max_length = 2048
# Dataset is too large to fit into memory, need to use disk for concatenation
datamodule = LMDataModule(dataset_name, tokenizer_name='gpt2',
dataset_config_name=dataset_config_name,
max_length=max_length, cache_dir=cache_dir,
add_eos=True, batch_size=batch_size,
num_workers=num_cpu_cores() // 2, use_shmem=False)
datamodule.prepare_data()
datamodule.setup(stage='fit')
train_loader = datamodule.train_dataloader()
val_loader = datamodule.val_dataloader()
datamodule.setup(stage='test')
test_loader = datamodule.test_dataloader()
train_len = 374337375694
val_len = 383326395
test_len = 373297018
assert len(train_loader) == div_up((train_len - 1) // max_length, batch_size)
assert len(val_loader) == div_up((val_len - 1) // max_length, batch_size)
assert len(test_loader) == div_up((test_len - 1) // max_length, batch_size)
for loader in [train_loader, val_loader, test_loader]:
x, y = next(iter(loader))
assert x.dim() == 2
assert x.shape == (batch_size, max_length)
assert x.dtype == torch.long
assert torch.allclose(x[:, 1:], y[:, :-1])
def test_pg19(self):
batch_size = 8
dataset_name = 'pg19'
dataset_config_name = None
data_dir = Path(os.getenv('DATA_DIR', current_dir.parent.parent / 'data'))
cache_dir = data_dir / 'pg19' / 'cache'
max_length = 2048
# Dataset is too large to fit into memory, need to use disk for concatenation
datamodule = LMDataModule(dataset_name, tokenizer_name='gpt2',
dataset_config_name=dataset_config_name,
max_length=max_length, cache_dir=cache_dir,
add_eos=True, batch_size=batch_size,
num_workers=num_cpu_cores() // 2)
datamodule.prepare_data()
datamodule.setup(stage='fit')
train_loader = datamodule.train_dataloader()
val_loader = datamodule.val_dataloader()
datamodule.setup(stage='test')
test_loader = datamodule.test_dataloader()
train_len = 3066544128
val_len = 4653056
test_len = 10584064
assert len(train_loader) == div_up((train_len - 1) // max_length, batch_size)
assert len(val_loader) == div_up((val_len - 1) // max_length, batch_size)
assert len(test_loader) == div_up((test_len - 1) // max_length, batch_size)
for loader in [train_loader, val_loader, test_loader]:
x, y = next(iter(loader))
assert x.dim() == 2
assert x.shape == (batch_size, max_length)
assert x.dtype == torch.long
assert torch.allclose(x[:, 1:], y[:, :-1])
|