stable-diffusion-finetune/ldm/data/laion.py

314 lines
11 KiB
Python
Raw Normal View History

2022-05-26 23:51:29 +02:00
import webdataset as wds
from PIL import Image
import io
import os
2022-05-27 11:46:04 +02:00
import torchvision
from PIL import Image
import glob
import random
import numpy as np
import pytorch_lightning as pl
2022-05-26 23:51:29 +02:00
from tqdm import tqdm
2022-05-27 11:46:04 +02:00
from omegaconf import OmegaConf
from einops import rearrange
import torch
2022-05-30 22:34:17 +02:00
from webdataset.handlers import warn_and_continue
2022-05-27 11:46:04 +02:00
from ldm.util import instantiate_from_config
2022-05-31 15:56:29 +02:00
class DataWithWings(torch.utils.data.IterableDataset):
2022-05-31 15:20:27 +02:00
def __init__(self, min_size, transform=None, target_transform=None):
self.min_size = min_size
self.transform = transform if transform is not None else nn.Identity()
self.target_transform = target_transform if target_transform is not None else nn.Identity()
self.kv = OnDiskKV(file='/home/ubuntu/laion5B-watermark-safety-ordered', key_format='q', value_format='ee')
self.kv_aesthetic = OnDiskKV(file='/home/ubuntu/laion5B-aesthetic-tags-kv', key_format='q', value_format='e')
self.pwatermark_threshold = 0.8
self.punsafe_threshold = 0.5
self.aesthetic_threshold = 5.
self.total_samples = 0
self.samples = 0
location = 'pipe:aws s3 cp --quiet s3://s-datasets/laion5b/laion2B-data/{000000..231349}.tar -'
self.inner_dataset = wds.DataPipeline(
wds.ResampledShards(location),
wds.tarfile_to_samples(handler=wds.warn_and_continue),
wds.shuffle(1000, handler=wds.warn_and_continue),
wds.decode('pilrgb', handler=wds.warn_and_continue),
wds.map(self._add_tags, handler=wds.ignore_and_continue),
wds.select(self._filter_predicate),
wds.map_dict(jpg=self.transform, txt=self.target_transform, punsafe=self._punsafe_to_class, handler=wds.warn_and_continue),
wds.to_tuple('jpg', 'txt', 'punsafe', handler=wds.warn_and_continue),
)
@staticmethod
def _compute_hash(url, text):
if url is None:
url = ''
if text is None:
text = ''
total = (url + text).encode('utf-8')
return mmh3.hash64(total)[0]
def _add_tags(self, x):
hsh = self._compute_hash(x['json']['url'], x['txt'])
pwatermark, punsafe = self.kv[hsh]
aesthetic = self.kv_aesthetic[hsh][0]
return {**x, 'pwatermark': pwatermark, 'punsafe': punsafe, 'aesthetic': aesthetic}
def _punsafe_to_class(self, punsafe):
return torch.tensor(punsafe >= self.punsafe_threshold).long()
def _filter_predicate(self, x):
try:
return x['pwatermark'] < self.pwatermark_threshold and x['aesthetic'] >= self.aesthetic_threshold and x['json']['original_width'] >= self.min_size and x['json']['original_height'] >= self.min_size
except:
return False
def __iter__(self):
return iter(self.inner_dataset)
2022-05-27 11:46:04 +02:00
def dict_collation_fn(samples, combine_tensors=True, combine_scalars=True):
"""Take a list of samples (as dictionary) and create a batch, preserving the keys.
If `tensors` is True, `ndarray` objects are combined into
tensor batches.
:param dict samples: list of samples
:param bool tensors: whether to turn lists of ndarrays into a single ndarray
:returns: single sample consisting of a batch
:rtype: dict
"""
keys = set.intersection(*[set(sample.keys()) for sample in samples])
batched = {key: [] for key in keys}
2022-05-27 11:46:04 +02:00
for s in samples:
[batched[key].append(s[key]) for key in batched]
result = {}
for key in batched:
if isinstance(batched[key][0], (int, float)):
if combine_scalars:
result[key] = np.array(list(batched[key]))
elif isinstance(batched[key][0], torch.Tensor):
if combine_tensors:
result[key] = torch.stack(list(batched[key]))
elif isinstance(batched[key][0], np.ndarray):
if combine_tensors:
result[key] = np.array(list(batched[key]))
else:
result[key] = list(batched[key])
return result
class WebDataModuleFromConfig(pl.LightningDataModule):
def __init__(self, tar_base, batch_size, train=None, validation=None,
2022-06-10 11:08:42 +02:00
test=None, num_workers=4, multinode=True, min_size=None,
2022-05-27 11:46:04 +02:00
**kwargs):
super().__init__(self)
print(f'Setting tar base to {tar_base}')
self.tar_base = tar_base
self.batch_size = batch_size
self.num_workers = num_workers
self.train = train
self.validation = validation
self.test = test
2022-05-30 22:34:17 +02:00
self.multinode = multinode
2022-06-10 11:08:42 +02:00
self.min_size = min_size # filter out very small images
2022-05-27 11:46:04 +02:00
def make_loader(self, dataset_config, train=True):
if 'image_transforms' in dataset_config:
image_transforms = [instantiate_from_config(tt) for tt in dataset_config.image_transforms]
else:
image_transforms = []
image_transforms.extend([torchvision.transforms.ToTensor(),
torchvision.transforms.Lambda(lambda x: rearrange(x * 2. - 1., 'c h w -> h w c'))])
image_transforms = torchvision.transforms.Compose(image_transforms)
if 'transforms' in dataset_config:
transforms_config = OmegaConf.to_container(dataset_config.transforms)
else:
transforms_config = dict()
2022-05-30 22:34:17 +02:00
transform_dict = {dkey: load_partial_from_config(transforms_config[dkey])
if transforms_config[dkey] != 'identity' else identity
for dkey in transforms_config}
2022-05-27 11:46:04 +02:00
img_key = dataset_config.get('image_key', 'jpeg')
transform_dict.update({img_key: image_transforms})
2022-06-13 01:26:31 +02:00
if 'postprocess' in dataset_config:
postprocess = instantiate_from_config(dataset_config['postprocess'])
else:
postprocess = None
2022-05-27 11:46:04 +02:00
shuffle = dataset_config.get('shuffle', 0)
2022-05-30 22:34:17 +02:00
shardshuffle = shuffle > 0
2022-05-27 11:46:04 +02:00
nodesplitter = wds.shardlists.split_by_node if self.multinode else wds.shardlists.single_node_only
2022-05-30 22:34:17 +02:00
tars = os.path.join(self.tar_base, dataset_config.shards)
dset = wds.WebDataset(
tars,
nodesplitter=nodesplitter,
2022-05-31 22:54:59 +02:00
shardshuffle=shardshuffle,
2022-06-13 12:59:50 +02:00
handler=wds.warn_and_continue).repeat().shuffle(shuffle)
2022-05-30 22:34:17 +02:00
print(f'Loading webdataset with {len(dset.pipeline[0].urls)} shards.')
2022-05-30 16:04:35 +02:00
2022-05-27 11:46:04 +02:00
dset = (dset
.select(self.filter_keys)
2022-05-31 22:54:59 +02:00
.decode('pil', handler=wds.warn_and_continue)
2022-06-10 11:44:55 +02:00
.select(self.filter_size)
2022-05-31 15:21:55 +02:00
.map_dict(**transform_dict, handler=wds.warn_and_continue)
2022-06-13 01:26:31 +02:00
)
if postprocess is not None:
dset = dset.map(postprocess)
dset = (dset
2022-05-27 11:46:04 +02:00
.batched(self.batch_size, partial=False,
2022-06-13 01:26:31 +02:00
collation_fn=dict_collation_fn)
2022-05-27 11:46:04 +02:00
)
loader = wds.WebLoader(dset, batch_size=None, shuffle=False,
num_workers=self.num_workers)
2022-05-30 22:34:17 +02:00
return loader
2022-05-27 11:46:04 +02:00
2022-06-10 11:08:42 +02:00
def filter_size(self, x):
if self.min_size is None:
return True
try:
return x['json']['original_width'] >= self.min_size and x['json']['original_height'] >= self.min_size
except Exception:
return False
def filter_keys(self, x):
try:
2022-06-10 11:43:49 +02:00
return ("jpg" in x) and ("txt" in x)
except Exception:
return False
2022-05-27 11:46:04 +02:00
def train_dataloader(self):
2022-05-30 22:34:17 +02:00
return self.make_loader(self.train)
2022-05-27 11:46:04 +02:00
def val_dataloader(self):
2022-05-30 22:34:17 +02:00
return self.make_loader(self.validation, train=False)
2022-05-27 11:46:04 +02:00
def test_dataloader(self):
2022-05-30 22:34:17 +02:00
return self.make_loader(self.test, train=False)
2022-05-27 11:46:04 +02:00
2022-05-26 23:51:29 +02:00
2022-06-13 01:26:31 +02:00
from ldm.modules.image_degradation import degradation_fn_bsr_light
class AddLR(object):
def __init__(self, factor):
self.factor = factor
def pt2np(self, x):
x = ((x+1.0)*127.5).clamp(0, 255).to(dtype=torch.uint8).detach().cpu().numpy()
return x
def np2pt(self, x):
x = torch.from_numpy(x)/127.5-1.0
return x
def __call__(self, sample):
# sample['jpg'] is tensor hwc in [-1, 1] at this point
x = self.pt2np(sample['jpg'])
x = degradation_fn_bsr_light(x, sf=self.factor)['image']
x = self.np2pt(x)
sample['lr'] = x
return sample
2022-05-30 22:34:17 +02:00
def example00():
2022-05-26 23:51:29 +02:00
url = "pipe:aws s3 cp s3://s-datasets/laion5b/laion2B-data/000000.tar -"
dataset = wds.WebDataset(url)
example = next(iter(dataset))
for k in example:
print(k, type(example[k]))
print(example["__key__"])
for k in ["json", "txt"]:
print(example[k].decode())
image = Image.open(io.BytesIO(example["jpg"]))
outdir = "tmp"
os.makedirs(outdir, exist_ok=True)
2022-05-27 11:46:04 +02:00
image.save(os.path.join(outdir, example["__key__"] + ".png"))
2022-05-26 23:51:29 +02:00
def load_example(example):
return {
"key": example["__key__"],
"image": Image.open(io.BytesIO(example["jpg"])),
"text": example["txt"].decode(),
}
for i, example in tqdm(enumerate(dataset)):
ex = load_example(example)
print(ex["image"].size, ex["text"])
if i >= 100:
break
2022-05-30 22:34:17 +02:00
def example01():
# the first laion shards contain ~10k examples each
url = "pipe:aws s3 cp s3://s-datasets/laion5b/laion2B-data/{000000..000002}.tar -"
batch_size = 3
shuffle_buffer = 10000
dset = wds.WebDataset(
url,
nodesplitter=wds.shardlists.split_by_node,
shardshuffle=True,
)
dset = (dset
.shuffle(shuffle_buffer, initial=shuffle_buffer)
.decode('pil', handler=warn_and_continue)
.batched(batch_size, partial=False,
collation_fn=dict_collation_fn)
)
num_workers = 2
loader = wds.WebLoader(dset, batch_size=None, shuffle=False, num_workers=num_workers)
batch_sizes = list()
keys_per_epoch = list()
for epoch in range(5):
keys = list()
for batch in tqdm(loader):
batch_sizes.append(len(batch["__key__"]))
keys.append(batch["__key__"])
for bs in batch_sizes:
assert bs==batch_size
print(f"{len(batch_sizes)} batches of size {batch_size}.")
batch_sizes = list()
keys_per_epoch.append(keys)
for i_batch in [0, 1, -1]:
print(f"Batch {i_batch} of epoch {epoch}:")
print(keys[i_batch])
print("next epoch.")
if __name__ == "__main__":
2022-06-08 23:00:53 +02:00
#example01()
from omegaconf import OmegaConf
from torch.utils.data.distributed import DistributedSampler
from torch.utils.data import IterableDataset
from torch.utils.data import DataLoader, RandomSampler, Sampler, SequentialSampler
from pytorch_lightning.trainer.supporters import CombinedLoader, CycleIterator
2022-06-13 01:26:31 +02:00
#config = OmegaConf.load("configs/stable-diffusion/txt2img-1p4B-multinode-clip-encoder-high-res-512.yaml")
config = OmegaConf.load("configs/stable-diffusion/txt2img-upscale-clip-encoder-f16-1024.yaml")
2022-06-08 23:00:53 +02:00
datamod = WebDataModuleFromConfig(**config["data"]["params"])
dataloader = datamod.train_dataloader()
for batch in dataloader:
print(batch.keys())
print(batch["jpg"].shape)
break