import collections
import os
import random
import torch
from torch.utils.data import IterableDataset, DataLoader
import pandas as pd
import glob
from typing import List, Dict, Any, Optional, Iterator
import pyarrow.parquet as pq
from transformers import AutoTokenizer
from torchvision import transforms
import json
from PIL import Image
class RefinedWebDataset(IterableDataset):
def __init__(self,
data_path,
rank: int = 0,
world_size: int = 1,
shuffle=True,
repeat=True,
buffer_size=1000,
max_length=8000,
num_workers=1):
super().__init__()
self.files = sorted(glob.glob(data_path))
self.rank = rank
self.world_size = world_size
self.shuffle = shuffle
self.repeat = repeat
self.buffer_size = buffer_size
self.max_length = max_length
self.num_workers = num_workers
self.files = self.files[self.rank::self.world_size]
def read_parquet_file(self, file_path):
table = pq.read_table(file_path, columns=["content"])
df = table.to_pandas()
for _, row in df.iterrows():
yield {"content": row["content"]}
def __iter__(self):
while True:
file_list = self.files
if self.shuffle:
random.shuffle(file_list)
for file in file_list:
data_generator = self.read_parquet_file(file)
buffer = []
for data in data_generator:
text = data["content"].replace("\n", "")
if len(text) > self.max_length:
start_index = random.randint(0, len(text) - self.max_length - 1)
selected_text = text[start_index:start_index + self.max_length]
else:
selected_text = text
buffer.append({"input_ids": selected_text})
if len(buffer) >= self.buffer_size:
if self.shuffle:
random.shuffle(buffer)
for item in buffer:
yield item
buffer = []
if buffer:
if self.shuffle:
random.shuffle(buffer)
for item in buffer:
yield item
if not self.repeat:
break
def collate_fn(self, batch):
batched = collections.defaultdict(list)
for data in batch:
for k, v in data.items():
batched[k].append(v)
for k, v in batched.items():
if k not in ('key', 'input_ids', 'similarity'):
batched[k] = torch.stack(v, dim=0)
return batched
class ChatDataset(IterableDataset):
def __init__(self,
data_path,
rank: int = 0,
world_size: int = 1,
shuffle=True,
repeat=True,
buffer_size=1000,
max_length=8000,
num_workers=1,
tokenizer=None):
super().__init__()
self.files = sorted(glob.glob(data_path))
self.rank = rank
self.world_size = world_size
self.shuffle = shuffle
self.repeat = repeat
self.buffer_size = buffer_size
self.max_length = max_length
self.num_workers = num_workers
self.tokenizer = tokenizer
self.files = self.files[self.rank::self.world_size]
def read_parquet_file(self, file_path):
table = pq.read_table(file_path, columns=["content"])
df = table.to_pandas()
for _, row in df.iterrows():
yield {"content": row["content"]}
def __iter__(self):
while True:
file_list = self.files
if self.shuffle:
random.shuffle(file_list)
for file in file_list:
data_generator = self.read_parquet_file(file)
buffer = []
for data in data_generator:
text = data["content"]
if self.tokenizer is None:
if len(text) > self.max_length:
start_index = random.randint(0, len(text) - self.max_length - 1)
selected_text = text[start_index:start_index + self.max_length]
else:
selected_text = text
else:
if len(self.tokenizer(text)['input_ids']) < self.max_length:
selected_text = text
else:
continue
buffer.append({"input_ids": selected_text})
if len(buffer) >= self.buffer_size:
if self.shuffle:
random.shuffle(buffer)
for item in buffer:
yield item
buffer = []
if buffer:
if self.shuffle:
random.shuffle(buffer)
for item in buffer:
yield item
if not self.repeat:
break
def collate_fn(self, batch):
batched = collections.defaultdict(list)
for data in batch:
for k, v in data.items():
batched[k].append(v)
for k, v in batched.items():
if k not in ('key', 'input_ids', 'similarity'):
batched[k] = torch.stack(v, dim=0)
return batched
class R2iDataset(IterableDataset):
def __init__(self,
data_path,
rank: int = 0,
world_size: int = 1,
shuffle=True,
repeat=True,
buffer_size=1000,
max_length=8000,
num_workers=1,
resolution=256,
tokenizer=None):
super().__init__()
self.data_path = data_path
self.rank = rank
self.world_size = world_size
self.shuffle = shuffle
self.repeat = repeat
self.buffer_size = buffer_size
self.max_length = max_length
self.num_workers = num_workers
self.tokenizer = tokenizer
self.resolution = resolution
def __iter__(self):
while True:
subdirs = sorted([d for d in glob.glob(os.path.join(self.data_path, "*")) if os.path.isdir(d)])
if self.shuffle:
random.shuffle(subdirs)
subdirs = subdirs[self.rank::self.world_size]
subdirs = ['/data_storage/lbw/datasets/laion-aesthetics-12m-images-2/00000']
for subdir in subdirs:
all_files = glob.glob(os.path.join(subdir, "*.*"))
base_names = set()
for file_path in all_files:
base_name = os.path.splitext(os.path.basename(file_path))[0]
base_names.add(base_name)
base_names = list(base_names)
if self.shuffle:
random.shuffle(base_names)
buffer = []
for base_name in base_names:
jpg_path = os.path.join(subdir, f"{base_name}.jpg")
caption_path = os.path.join(subdir, f"{base_name}.caption")
shortcaption_path = os.path.join(subdir, f"{base_name}.shortcaption")
if not os.path.exists(jpg_path):
continue
try:
image = Image.open(jpg_path).convert("RGB")
caption = ""
if os.path.exists(caption_path):
with open(caption_path, "r", encoding="utf-8") as f:
caption = f.read().strip()
short_caption = ""
if os.path.exists(shortcaption_path):
with open(shortcaption_path, "r", encoding="utf-8") as f:
short_caption = f.read().strip()
transformed_image = image_transform_clip({"images": image}, resolution=self.resolution)["images"]
if self.tokenizer is not None:
if len(self.tokenizer(caption)['input_ids']) > self.max_length - 2:
continue
prompt = (
'<|start_header_id|>user<|end_header_id|>\n'
"You should first think out a more detailed version of the description and then provide the user with the image. The detailed description is enclosed within tags, i.e. detailed description here image here\n"
f"{short_caption}"
'<|start_header_id|>assistant<|end_header_id|>\n'
f"{caption}"
)
sample = {
"images": transformed_image,
"input_ids": prompt,
}
buffer.append(sample)
if len(buffer) >= self.buffer_size:
if self.shuffle:
random.shuffle(buffer)
for item in buffer:
yield item
buffer = []
except Exception as e:
print(f"Error processing {jpg_path}: {e}")
continue
if buffer:
if self.shuffle:
random.shuffle(buffer)
for item in buffer:
yield item
if not self.repeat:
break
def collate_fn(self, batch):
batched = collections.defaultdict(list)
for data in batch:
for k, v in data.items():
batched[k].append(v)
for k, v in batched.items():
if k not in ('key', 'input_ids', 'similarity'):
batched[k] = torch.stack(v, dim=0)
return batched
class VQADataset(IterableDataset):
def __init__(self,
json_path: str,
image_root: str,
tokenizer = None,
rank: int = 0,
world_size: int = 1,
shuffle: bool = True,
repeat: bool = True,
buffer_size: int = 100,
resolution: int = 256,
max_length: int = 8000,
num_workers: int = 1,
image_transform_method: str = "squash"):
super().__init__()
self.json_path = json_path
self.image_root = image_root
self.tokenizer = tokenizer
self.rank = rank
self.world_size = world_size
self.shuffle = shuffle
self.repeat = repeat
self.buffer_size = buffer_size
self.resolution = resolution
self.max_length = max_length
self.num_workers = num_workers
self.image_transform_method = image_transform_method
try:
with open(self.json_path, 'r', encoding='utf-8') as f:
raw_data = json.load(f)
except FileNotFoundError:
print(f"Error: Data file not found at {self.json_path}")
self.list_data_dict = []
except json.JSONDecodeError:
print(f"Error: Could not decode JSON from {self.json_path}")
self.list_data_dict = []
else:
self.list_data_dict = [item for item in raw_data if 'image' in item and 'conversations' in item]
self.list_data_dict = self.list_data_dict[self.rank::self.world_size]
def __iter__(self):
sot_token = '<|startoftext|>'
assistant_prompt_suffix = '<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n'
while True:
current_data_list = list(self.list_data_dict)
if self.shuffle:
random.shuffle(current_data_list)
buffer = []
for item in current_data_list:
image_relative_path = item.get('image')
conversations = item.get('conversations', [])
if not image_relative_path or not conversations or len(conversations) < 2:
continue
num_total_messages = len(conversations)
if num_total_messages % 2 != 0:
conversations = conversations[:-1]
num_total_messages -= 1
if num_total_messages < 2: continue
num_turns = num_total_messages // 2
if num_turns == 0:
continue
selected_num_turns = random.randint(1, num_turns)
selected_conversations = conversations[:selected_num_turns * 2]
image_path = os.path.join(self.image_root, image_relative_path)
try:
image = Image.open(image_path).convert("RGB")
if self.image_transform_method == "squash":
transformed_image = image_transform_squash({"images": image}, resolution=self.resolution)["images"]
elif self.image_transform_method == "pad":
transformed_image = image_transform_pad({"images": image}, resolution=self.resolution)["images"]
else:
transformed_image = image_transform_clip({"images": image}, resolution=self.resolution)["images"]
first_human_message = selected_conversations[0]['value']
processed_message = first_human_message.replace('\n', '').replace('\n', '')
current_selection_messages = list(selected_conversations)
current_selection_messages[0] = dict(current_selection_messages[0])
current_selection_messages[0]['value'] = processed_message
messages = []
for turn in current_selection_messages:
role = "user" if turn["from"] == "human" else "assistant"
messages.append({"role": role, "content": turn["value"]})
formatted_text = self.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
if formatted_text.startswith(sot_token):
formatted_text = formatted_text[len(sot_token):]
if formatted_text.endswith(assistant_prompt_suffix):
formatted_text = formatted_text[:-len(assistant_prompt_suffix)]
token_ids = self.tokenizer(formatted_text)['input_ids']
if len(token_ids) > self.max_length:
continue
sample = {
"images": transformed_image,
"input_ids": formatted_text,
}
buffer.append(sample)
if len(buffer) >= self.buffer_size:
if self.shuffle:
random.shuffle(buffer)
for buf_item in buffer:
yield buf_item
buffer = []
except FileNotFoundError:
print(f"Warning: Image file not found at {image_path}, skipping item.")
continue
except Exception as e:
print(f"Warning: Error processing item with image {image_path}: {e}, skipping.")
continue
if buffer:
if self.shuffle:
random.shuffle(buffer)
for buf_item in buffer:
yield buf_item
if not self.repeat:
break
def collate_fn(self, batch):
batched = collections.defaultdict(list)
for data in batch:
for k, v in data.items():
batched[k].append(v)
for k, v in batched.items():
if k not in ('key', 'input_ids', 'similarity'):
batched[k] = torch.stack(v, dim=0)
return batched
def image_transform_clip(sample, resolution=256):
image = sample["images"]
image = transforms.Resize(resolution, interpolation=transforms.InterpolationMode.BICUBIC)(image)
image = transforms.CenterCrop((resolution, resolution))(image)
image = transforms.ToTensor()(image)
image = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True)(image)
sample["images"] = image
return sample
def image_transform_squash(sample, resolution=256):
image = sample["images"]
image = transforms.Resize((resolution, resolution), interpolation=transforms.InterpolationMode.BICUBIC)(image)
image = transforms.ToTensor()(image)
image = transforms.Normalize(mean=[0.5, 0.5, 0.5],std=[0.5, 0.5, 0.5])(image)
sample["images"] = image
return sample
def image_transform_pad(sample, resolution=256, fill_color=(255, 255, 255)):
image = sample["images"]
w, h = image.size
if w == h:
padded_image = image
elif w < h:
padding_needed = h - w
padding_left = padding_needed // 2
padding_right = padding_needed - padding_left
pad_transform = transforms.Pad((padding_left, 0, padding_right, 0), fill=fill_color, padding_mode='constant')
padded_image = pad_transform(image)
else:
padding_needed = w - h
padding_top = padding_needed // 2
padding_bottom = padding_needed - padding_top
pad_transform = transforms.Pad((0, padding_top, 0, padding_bottom), fill=fill_color, padding_mode='constant')
padded_image = pad_transform(image)
image_resized = transforms.Resize((resolution, resolution), interpolation=transforms.InterpolationMode.BICUBIC)(padded_image)
image_tensor = transforms.ToTensor()(image_resized)
image_normalized = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])(image_tensor)
sample["images"] = image_normalized
return sample
if __name__ == '__main__':
data_path = "/data_storage/shared/datasets/falcon-refinedweb/data/data/*.parquet"
dataset = RefinedWebDataset(
data_path=data_path,
max_length=8000,
buffer_size=0,
)
from torch.utils.data import DataLoader
train_dataloader = DataLoader(
dataset,
batch_size=1,
sampler=None,
collate_fn=dataset.collate_fn,
num_workers=0
)
print("Starting data loading test...")
for i, batch in enumerate(train_dataloader):
if i == 0:
print(batch)
print(f"Batch size: {len(batch['input_ids'])}")
print(f"First sample length: {len(batch['input_ids'][0])}")
if i >= 5:
break
print("Data loading test complete")