jaeikkim
Reinit Space without binary assets
7bfbdc3
# coding=utf-8
# Copyright 2025 MMaDA Team
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import bisect
import csv
import logging
import itertools
import json
import math
import os
import hashlib
import contextlib
from pathlib import Path
from accelerate import Accelerator
from itertools import chain
# Video real-time?
import os.path as osp
import time
import requests
import random
import re
import datasets
import pandas as pd
from functools import partial
from typing import List, Optional, Union, Dict, Any, Sequence
from glob import glob
from tqdm import tqdm
import numpy as np
import cv2
from PIL import Image
import torch
from dataclasses import dataclass
from datasets import Dataset as HFDataset
from datasets import load_dataset, get_dataset_config_names
from io import BytesIO
Image.warnings.simplefilter('error', Image.DecompressionBombWarning)
import webdataset as wds
import yaml
from braceexpand import braceexpand
from torch.utils.data import default_collate, Dataset
from torchvision import transforms
from transformers import PreTrainedTokenizer
from datasets import (
load_dataset,
load_from_disk,
DatasetDict,
DownloadConfig,
get_dataset_config_names,
concatenate_datasets,
)
import warnings
from training.utils import image_transform as utils_image_transform, image_transform_squash as utils_image_transform_squash
from webdataset.tariterators import (
base_plus_ext,
tar_file_expander,
url_opener,
valid_sample,
)
logger = logging.getLogger(__name__)
S2T_INSTRUCTION = ["Transcribe the given audio.",
"Write down what you hear in the audio.",
"Provide a transcript for the given speech.",
"What does the speaker in the audio say?",
"Convert the speech in the audio to text.",
"Listen to the audio and write out the text."]
T2S_INSTRUCTION = ["Generate speech for the given text.",
"Read the given sentence aloud.",
"Say the given words.",
"Convert the given text into spoken audio.",
"Speak the given text.",
"Synthesize the text into speech."]
V2T_INSTRUCTION = ["Describe the video in detail.",
"Please provide a detailed description of the video.",
"What is happening in the video?",
"Describe the content of the video in detail.",]
V2S_INSTRUCTION = [
"Generate speech that describes the given video.",
"Narrate the events happening in the video.",
"Produce spoken audio describing the video content.",
"Convert the video into a detailed spoken narration.",
"Speak a description of what is shown in the video.",
"Synthesize speech that explains the content of the video.",
]
person_token = ["a person", "someone", "somebody"]
def replace_person_token(t):
"Used for CC12M - handles all case variations of <person> tag"
t = re.sub(r"<person>([,\s]*(and)*[,\s]*<person>)+", " people ", t, flags=re.IGNORECASE)
person_pattern = re.compile(r"<person>", re.IGNORECASE)
while person_pattern.search(t):
match = person_pattern.search(t)
t = t[:match.start()] + f" {random.choice(person_token)} " + t[match.end():]
return t
def filter_keys(key_set):
def _f(dictionary):
return {k: v for k, v in dictionary.items() if k in key_set}
return _f
def group_by_keys_nothrow(data, keys=base_plus_ext, lcase=True, suffixes=None, handler=None, src=None):
"""Return function over iterator that groups key, value pairs into samples.
:param keys: function that splits the key into key and extension (base_plus_ext)
:param lcase: convert suffixes to lower case (Default value = True)
"""
current_sample = None
for filesample in data:
assert isinstance(filesample, dict)
if "fname" not in filesample.keys():
print(f"fname not in filesample.keys(): {filesample}")
print(f"src: {src}")
continue
fname, value = filesample["fname"], filesample["data"]
prefix, suffix = keys(fname)
if prefix is None:
continue
if lcase:
suffix = suffix.lower()
if current_sample is None or prefix != current_sample["__key__"] or suffix in current_sample:
if valid_sample(current_sample):
yield current_sample
current_sample = dict(__key__=prefix, __url__=filesample["__url__"])
if suffixes is None or suffix in suffixes:
current_sample[suffix] = value
if valid_sample(current_sample):
yield current_sample
def tarfile_to_samples_nothrow(src, handler=wds.warn_and_continue):
# NOTE this is a re-impl of the webdataset impl with group_by_keys that doesn't throw
streams = url_opener(src, handler=handler)
files = tar_file_expander(streams, handler=handler) # [{fname,data,__url__}, ...] __url__ 字段标识当前读取的文件来自哪个 tar 包
samples = group_by_keys_nothrow(files, handler=handler, src=src)
return samples
def image_transform(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 conditional_image_transform(sample, resolution=256):
url = sample.get("__url__", "")
special_datasets = ['ai2d', 'clevr', 'docvqa', 'geo']
use_squash = False
for keyword in special_datasets:
if keyword in url:
use_squash = True
break
if use_squash:
return image_transform_squash(sample, resolution)
else:
return image_transform(sample, resolution)
def remove_prefix(caption):
caption = caption.replace('The image features ', '').replace('The image presents ', '').replace(
"The image you've sent is, ", '').replace("In the center of the image, ", '').replace(
"The image showcases ", '').replace("The image is ", '').replace(
"The image captures ", '').replace("In the given image ", '').replace(
"The image portrays ", '').replace("In the image, ", '').replace("In this image, we see ", '').replace(
"The image depicts ", '').replace("This is ", '').replace("In this image, ", '').replace(
"This image captures ", '')
return caption
def filter_long_samples(sample):
return sample.get('input_ids') is not None
class Text2ImageDataset:
def __init__(
self,
train_shards_path_or_url: Union[str, List[str]],
tokenizer: PreTrainedTokenizer,
max_seq_length: int,
num_train_examples: int,
per_gpu_batch_size: int,
global_batch_size: int,
num_workers: int,
resolution: int = 256,
shuffle_buffer_size: int = 1000,
pin_memory: bool = False,
persistent_workers: bool = False,
external_caption_path: Optional[str] = '',
external_journeydb_caption_path: Optional[str] = '',
external_laion12m_caption_path: Optional[str] = '',
external_cc12m_caption_path: Optional[str] = '',
external_text_to_image_2M_512_caption_path: Optional[str] = '',
external_ai2d_caption_path: Optional[str] = '',
external_clevr_caption_path: Optional[str] = '',
external_docvqa_caption_path: Optional[str] = '',
external_geo_caption_path: Optional[str] = '',
is_captioning: bool = False,
add_caption_prompt: bool = False,
long_caption: bool = True,
shuffle: bool = True,
):
if f"{train_shards_path_or_url}.yaml" in os.listdir('./configs'):
with open(f"./configs/{train_shards_path_or_url}.yaml") as f:
train_shards_path_or_url = yaml.safe_load(f)
self.long_caption = long_caption
self.external_caption_path = external_caption_path
self.external_journeydb_caption_path = external_journeydb_caption_path
self.external_laion12m_caption_path = external_laion12m_caption_path
self.external_cc12m_caption_path = external_cc12m_caption_path
self.external_text_to_image_2M_512_caption_path = external_text_to_image_2M_512_caption_path
self.is_captioning = is_captioning
self.add_caption_prompt = add_caption_prompt
if self.add_caption_prompt:
with open("./training/questions.json") as f:
self.caption_prompt = json.load(f)
# self.caption_prompt = ['USER: \n' + prompt + ' ASSISTANT:' for prompt in self.caption_prompt]
self.caption_prompt = ['<|start_header_id|>user<|end_header_id|>\n' + prompt + '<eot_id><|start_header_id|>assistant<|end_header_id|>\n' for prompt in self.caption_prompt]
else:
self.caption_prompt = None
if external_journeydb_caption_path != '':
with open(external_journeydb_caption_path) as file:
self.journeydb_caption = json.load(file)
else:
self.journeydb_caption = None
if external_ai2d_caption_path!= '':
self.ai2d_caption = pd.read_csv(external_ai2d_caption_path)
if external_clevr_caption_path!= '':
self.clevr_caption = pd.read_csv(external_clevr_caption_path)
if external_docvqa_caption_path!= '':
self.docvqa_caption = pd.read_csv(external_docvqa_caption_path)
if external_geo_caption_path!= '':
self.geo_caption = pd.read_csv(external_geo_caption_path)
def tokenize(text):
if tokenizer is not None:
text = replace_person_token(text)
encoding = tokenizer(
text,
truncation=True,
max_length=2 * max_seq_length,
padding=False,
return_tensors="pt"
)
full_input_ids = encoding.input_ids[0]
if len(full_input_ids) > max_seq_length:
return None
else:
return text
else:
return text
if not isinstance(train_shards_path_or_url, str):
train_shards_path_or_url = [list(braceexpand(urls)) for urls in train_shards_path_or_url]
# flatten list using itertools
train_shards_path_or_url = list(itertools.chain.from_iterable(train_shards_path_or_url))
if external_caption_path != '':
processing_pipeline = [
wds.decode("pil", handler=wds.ignore_and_continue),
wds.map(self.load_external_caption, handler=wds.ignore_and_continue),
wds.rename(
images="jpg;png;jpeg;webp",
input_ids="text;txt;caption",
handler=wds.warn_and_continue,
),
wds.map(partial(conditional_image_transform, resolution=resolution), handler=wds.warn_and_continue),
wds.map(filter_keys(set(["images", "input_ids"]))),
wds.map_dict(
input_ids=tokenize,
handler=wds.warn_and_continue,
),
wds.select(filter_long_samples),
]
else:
processing_pipeline = [
wds.decode("pil", handler=wds.ignore_and_continue),
wds.rename(
images="jpg;png;jpeg;webp",
input_ids="text;txt;caption",
handler=wds.warn_and_continue,
),
wds.map(partial(conditional_image_transform, resolution=resolution), handler=wds.warn_and_continue),
wds.map(filter_keys(set(["images", "input_ids"]))),
wds.map_dict(
input_ids=tokenize,
handler=wds.warn_and_continue,
),
wds.select(filter_long_samples),
]
pipeline = [
wds.ResampledShards(train_shards_path_or_url),
tarfile_to_samples_nothrow,
wds.shuffle(shuffle_buffer_size),
*processing_pipeline,
wds.batched(per_gpu_batch_size, partial=False, collation_fn=default_collate),
]
num_batches = math.ceil(num_train_examples / global_batch_size)
num_worker_batches = math.ceil(num_train_examples / (global_batch_size * num_workers)) # per dataloader worker
num_batches = num_worker_batches * num_workers
num_samples = num_batches * global_batch_size
self._train_dataset = wds.DataPipeline(*pipeline).with_epoch(num_worker_batches)
self._train_dataloader = wds.WebLoader(
self._train_dataset,
batch_size=None,
shuffle=False,
num_workers=num_workers,
pin_memory=pin_memory,
persistent_workers=persistent_workers,
)
# add meta-data to dataloader instance for convenience
self._train_dataloader.num_batches = num_batches
self._train_dataloader.num_samples = num_samples
def load_external_caption(self, sample):
if 'SA1B' in sample['__key__'] or 'sa' in sample['__key__']:
captionf = f"{self.external_caption_path}/{sample['__key__'].split('/')[-1]}.txt"
if os.path.exists(captionf):
with open(captionf, "r") as reader:
captions = reader.readlines()[0].replace('\n', '')
else:
captions = ""
# for captioning
if self.is_captioning:
if self.add_caption_prompt is not None:
prompt = random.sample(self.caption_prompt, 1)[0]
sample['txt'] = prompt + captions
else:
sample['txt'] = captions
# for generation
else:
# randomly choose short and long captions
if random.random() < 0.5:
sample['txt'] = captions.split('.')[0]
else:
sample['txt'] = captions
sample['txt'] = remove_prefix(sample['txt'])
return sample
elif 'laion' in sample['__url__']:
url_part = sample['__url__'].split('/')[-1].split('.')[0]
key = sample['__key__'].split('/')[-1]
captionf = os.path.join(self.external_laion12m_caption_path, url_part, f"{key}.caption")
if os.path.exists(captionf):
with open(captionf, "r") as reader:
captions = reader.read().strip()
else:
captions = ""
# for captioning
if self.is_captioning:
if self.add_caption_prompt is not None:
prompt = random.sample(self.caption_prompt, 1)[0]
sample['txt'] = prompt + captions
else:
sample['txt'] = captions
# for generation
else:
# randomly choose short and long captions
if random.random() < 0.5:
sample['txt'] = captions.split('.')[0]
else:
sample['txt'] = captions
sample['txt'] = remove_prefix(sample['txt'])
return sample
elif 'cc12m' in sample['__url__']:
url_part = sample['__url__'].split('/')[-1].split('.')[0]
key = sample['__key__'].split('/')[-1]
captionf = os.path.join(self.external_cc12m_caption_path, url_part, f"{key}.caption")
if os.path.exists(captionf):
with open(captionf, "r") as reader:
captions = reader.read().strip()
else:
captions = ""
# for captioning
if self.is_captioning:
if self.add_caption_prompt is not None:
prompt = random.sample(self.caption_prompt, 1)[0]
sample['txt'] = prompt + captions
else:
sample['txt'] = captions
# for generation
else:
# randomly choose short and long captions
if random.random() < 0.5:
sample['txt'] = captions.split('.')[0]
else:
sample['txt'] = captions
sample['txt'] = remove_prefix(sample['txt'])
return sample
elif "text-to-image-2M" in sample['__url__']:
if "json" in sample and "prompt" in sample["json"]:
captions = sample["json"]["prompt"]
else:
print(f"sample has no json or prompt: {sample}")
captions = ""
sample['txt'] = captions
return sample
elif 'ai2d' in sample['__url__']:
key = sample['__key__'].split('/')[-1]
df_row = self.ai2d_caption[self.ai2d_caption['image'].astype(str) == key + '.png']
if len(df_row) == 0:
print(f"No captions available for key {sample['__key__']}")
return sample
elif len(df_row) > 1:
# print(f"Multiple captions available for key {sample['__key__']}")
df_row = df_row.sample(1)
question = df_row['question'].values[0]
solution = df_row['solution'].values[0]
caption = (
'<|start_header_id|>user<|end_header_id|>\n'
"You should first think about the reasoning process in the mind and then provide the user with the answer. The reasoning process is enclosed within <think> </think> tags, i.e. <think> reasoning process here </think> answer here\n"
f"{question}\n"
'<eot_id><|start_header_id|>assistant<|end_header_id|>\n'
f"{solution}"
)
sample['txt'] = caption
return sample
elif 'clevr' in sample['__url__']:
key = sample['__key__'].split('/')[-1]
df_row = self.clevr_caption[self.clevr_caption['image'].astype(str) == key + ".jpg"]
if len(df_row) == 0:
print(f"No captions available for key {sample['__key__']}")
return sample
elif len(df_row) > 1:
# print(f"Multiple captions available for key {sample['__key__']}")
df_row = df_row.sample(1)
question = df_row['question'].values[0]
solution = df_row['solution'].values[0]
caption = (
'<|start_header_id|>user<|end_header_id|>\n'
"You should first think about the reasoning process in the mind and then provide the user with the answer. The reasoning process is enclosed within <think> </think> tags, i.e. <think> reasoning process here </think> answer here\n"
f"{question}\n"
'<eot_id><|start_header_id|>assistant<|end_header_id|>\n'
f"{solution}"
)
sample['txt'] = caption
return sample
elif 'docvqa' in sample['__url__']:
key = sample['__key__'].split('/')[-1]
df_row = self.docvqa_caption[self.docvqa_caption['image'].astype(str) == key + ".png"]
if len(df_row) == 0:
print(f"No captions available for key {sample['__key__']}")
return sample
elif len(df_row) > 1:
# print(f"Multiple captions available for key {sample['__key__']}")
df_row = df_row.sample(1)
question = df_row['question'].values[0]
solution = df_row['solution'].values[0]
caption = (
'<|start_header_id|>user<|end_header_id|>\n'
"You should first think about the reasoning process in the mind and then provide the user with the answer. The reasoning process is enclosed within <think> </think> tags, i.e. <think> reasoning process here </think> answer here\n"
f"{question}\n"
'<eot_id><|start_header_id|>assistant<|end_header_id|>\n'
f"{solution}"
)
sample['txt'] = caption
return sample
elif 'geo' in sample['__url__']:
key = sample['__key__'].split('/')[-1]
df_row = self.geo_caption[self.geo_caption['image'].astype(str) == key + ".jpg"]
if len(df_row) == 0:
print(f"No captions available for key {sample['__key__']}")
return sample
elif len(df_row) > 1:
# print(f"Multiple captions available for key {sample['__key__']}")
df_row = df_row.sample(1)
question = df_row['question'].values[0]
solution = df_row['solution'].values[0]
caption = (
'<|start_header_id|>user<|end_header_id|>\n'
"You should first think about the reasoning process in the mind and then provide the user with the answer. The reasoning process is enclosed within <think> </think> tags, i.e. <think> reasoning process here </think> answer here\n"
f"{question}\n"
'<eot_id><|start_header_id|>assistant<|end_header_id|>\n'
f"{solution}"
)
sample['txt'] = caption
return sample
elif self.journeydb_caption is not None and sample['__key__'] in self.journeydb_caption:
captions_list = self.journeydb_caption[sample['__key__']]
if len(captions_list) == 0:
print(f"No captions available for key {sample['__key__']}")
return sample
sample['txt'] = random.sample(captions_list, 1)[0]
return sample
else:
print(f"none exist sample: {sample}")
return sample
@property
def train_dataset(self):
return self._train_dataset
@property
def train_dataloader(self):
return self._train_dataloader
# +++++ S2T/T2S Dataset Definition +++++
class SpeechTextDataset(Dataset):
def __init__(self, dataset : str, subset : str, split : Optional[str] = None):
self.dataset_name = dataset
if self.dataset_name == "gigaspeech": # subset is either "xs" or "xl"
self.hgf_dataset : datasets.Dataset = load_dataset("speechcolab/gigaspeech", subset, split=split)
elif self.dataset_name == "librispeech":
root_path = "/home/work/AIDAS/data/audio/LibriSpeech"
self.dataset_path = root_path + "/" + subset # subset is like "train-clean-100", etc
if split is not None:
warnings.warn(f"Split parameter '{split}' is provided but will not be used for LibriSpeech dataset.")
# librispeech path processing
self.subdirs_path = sorted(list(glob(self.dataset_path + "/*/*")))
self.subdirs_len = [len(glob(subdir + "/*.flac")) for subdir in self.subdirs_path]
self.subdirs_len_accum = list(itertools.accumulate(self.subdirs_len))
# handle wrong subset name
if len(self.subdirs_path) == 0:
raise ValueError(f"Invalid subset name '{subset}' for LibriSpeech dataset. Available subsets are: train-clean-100, train-clean-360")
elif self.dataset_name == "commonvoice":
self.commonvoice_path = "/home/work/AIDAS/data/audio/commonvoice/cv-corpus-22.0-2025-06-20/en"
if split is not None:
warnings.warn(f"Split parameter '{split}' is provided but will not be used for commonvoice dataset.")
self.tsv = pd.read_csv(self.commonvoice_path + f"/{subset}.tsv", sep="\t", usecols=["path", "sentence"])
else:
raise ValueError(f"Unsupported dataset: {dataset}. Supported datasets are: gigaspeech, librispeech, commonvoice.")
def __len__(self):
if self.dataset_name == "gigaspeech":
return len(self.hgf_dataset)
elif self.dataset_name == "librispeech":
return self.subdirs_len_accum[-1]
else: # commonvoice
return len(self.tsv)
def __getitem__(self, idx):
audio_path : str; text : str
if self.dataset_name == "gigaspeech":
sample = self.hgf_dataset[idx]
audio_path = sample["audio"]["path"]
text = sample["text"]
elif self.dataset_name == "librispeech":
# idx overflow
if idx >= self.subdirs_len_accum[-1]:
raise IndexError(f"Index {idx} is out of bounds for the dataset with length {len(self)}.")
# audio_path (flac)
subdir_idx = bisect.bisect_right(self.subdirs_len_accum, idx)
flac_idx = idx - self.subdirs_len_accum[subdir_idx - 1] if subdir_idx > 0 else idx
audio_path = sorted(list(glob(self.subdirs_path[subdir_idx]+"/*.flac")))[flac_idx]
# text
txt_path = glob(self.subdirs_path[subdir_idx]+"/*.txt")
assert len(txt_path) == 1, f"Expected one txt file in {self.subdirs_path[subdir_idx]}, found {len(txt_path)}"
with open(txt_path[0], "r") as f:
txt = f.readlines()
text = " ".join(txt[flac_idx].split(" ")[1:]) # rip off the header, e.g., "103-1240-0007 [TEXT]"
else: # commonvoice
audio_path = self.commonvoice_path + "/clips/" + self.tsv.iloc[idx]["path"]
text = self.tsv.iloc[idx]["sentence"]
return {"audio_path": audio_path, "text": text}
class MixedSpeechTextDataset(Dataset):
def __init__(self, dataset_configs: list):
"""
Initializes and combines multiple speech datasets.
Args:
dataset_configs (list): A list of configuration dictionaries,
where each dict defines a dataset to load.
"""
self.dataset_metadata = []
self.dataset_lengths = []
self._sha1 = hashlib.sha1
# Iterate through the list of dataset configurations from the YAML file
for config in dataset_configs:
name = config['name']
subset = config.get('subset')
split = config.get('split')
use_tokens = bool(config.get("use_precomputed_tokens", False))
token_root = config.get("precomputed_tokens_root")
token_root_path = Path(token_root).expanduser() if token_root else None
print(f"Initializing dataset: {name} (Subset: {subset}, Split: {split})")
# --- Gigaspeech ---
if name == "gigaspeech":
hgf_dataset = datasets.load_dataset("speechcolab/gigaspeech", subset, split=split)
self.dataset_metadata.append({
"name": name,
"data": hgf_dataset,
"use_precomputed_tokens": use_tokens and token_root_path is not None,
"precomputed_tokens_root": token_root_path,
})
self.dataset_lengths.append(len(hgf_dataset))
# --- LibriSpeech ---
elif name == "librispeech":
root_path = "/home/work/AIDAS/data/audio/LibriSpeech"
dataset_path = f"{root_path}/{subset}"
if split is not None:
warnings.warn(f"Split parameter '{split}' is provided but will not be used for LibriSpeech.")
subdirs_path = sorted(glob(f"{dataset_path}/*/*"))
if not subdirs_path:
raise ValueError(f"Invalid subset for LibriSpeech or path not found: {dataset_path}")
subdirs_len = [len(glob(f"{subdir}/*.flac")) for subdir in subdirs_path]
subdirs_len_accum = list(itertools.accumulate(subdirs_len))
metadata = {
"name": name,
"subdirs_path": subdirs_path,
"subdirs_len_accum": subdirs_len_accum,
"use_precomputed_tokens": use_tokens and token_root_path is not None,
"precomputed_tokens_root": token_root_path,
}
self.dataset_metadata.append(metadata)
self.dataset_lengths.append(subdirs_len_accum[-1])
# --- Common Voice ---
elif name == "commonvoice":
commonvoice_path = "/home/work/AIDAS/data/audio/commonvoice/cv-corpus-22.0-2025-06-20/en"
if split is not None:
warnings.warn(f"Split parameter '{split}' is provided but will not be used for Common Voice.")
tsv_path = f"{commonvoice_path}/{subset}.tsv"
tsv = pd.read_csv(tsv_path, sep="\t", usecols=["path", "sentence"])
metadata = {
"name": name,
"data_root": f"{commonvoice_path}/clips/",
"tsv": tsv,
"use_precomputed_tokens": use_tokens and token_root_path is not None,
"precomputed_tokens_root": token_root_path,
}
self.dataset_metadata.append(metadata)
self.dataset_lengths.append(len(tsv))
else:
raise ValueError(f"Unsupported dataset: {name}.")
# Calculate cumulative lengths to map a global index to a specific dataset
self.cumulative_lengths = list(itertools.accumulate(self.dataset_lengths))
# print(f"✅ All datasets loaded for the SPEECH!. Total length: {self.__len__()} samples.")
def __len__(self):
"""Returns the total number of samples across all datasets."""
return self.cumulative_lengths[-1] if self.cumulative_lengths else 0
def __getitem__(self, idx):
"""
Fetches a sample from the combined dataset.
It first determines which dataset the global index `idx` belongs to,
calculates the local index within that dataset, and then retrieves the item.
"""
if idx >= self.__len__():
raise IndexError(f"Index {idx} is out of bounds for the combined dataset with length {self.__len__()}.")
# Find which dataset the index belongs to
dataset_idx = bisect.bisect_right(self.cumulative_lengths, idx)
# Calculate the local index within that dataset
local_idx = idx - self.cumulative_lengths[dataset_idx - 1] if dataset_idx > 0 else idx
metadata = self.dataset_metadata[dataset_idx]
dataset_name = metadata["name"]
dataset_length = self.dataset_lengths[dataset_idx]
audio_path: str
text: str
audio_tokens: Optional[torch.Tensor]
max_retry = 5
retry = 0
while retry < max_retry:
try:
audio_tokens = None
if dataset_name == "gigaspeech":
sample = metadata["data"][local_idx]
audio_path = sample["audio"]["path"]
text = sample["text"]
# Preprocess special words to punctuation
text = (
text.replace(" <COMMA>", ",")
.replace(" <PERIOD>", ".")
.replace(" <QUESTIONMARK>", "?")
.replace(" <EXCLAMATIONMARK>", "!")
)
elif dataset_name == "librispeech":
# Find the specific subdirectory and file using the local index
subdir_idx = bisect.bisect_right(metadata["subdirs_len_accum"], local_idx)
flac_idx = local_idx - metadata["subdirs_len_accum"][subdir_idx - 1] if subdir_idx > 0 else local_idx
subdir_path = metadata["subdirs_path"][subdir_idx]
audio_path = sorted(glob(f"{subdir_path}/*.flac"))[flac_idx]
# Read the corresponding transcript
txt_path = glob(f"{subdir_path}/*.txt")[0]
with open(txt_path, "r") as f:
line = f.readlines()[flac_idx]
text = " ".join(line.strip().split(" ")[1:])
else: # commonvoice
row = metadata["tsv"].iloc[local_idx]
audio_path = metadata["data_root"] + row["path"]
text = row["sentence"]
# Preprocess lower case to upper case
text = text.upper()
audio_tokens = self._maybe_load_precomputed_tokens(audio_path, metadata)
return {
"audio_path": audio_path,
"text": text,
"audio_tokens": audio_tokens,
}
except Exception as exc:
print(f"[MixedSpeechTextDataset] Failed to load sample from '{dataset_name}' at local index {local_idx}: {exc!r}")
retry += 1
if retry >= max_retry:
break
local_idx = random.randint(0, dataset_length - 1)
continue
raise RuntimeError(f"Unable to fetch a valid sample from dataset '{dataset_name}' after {max_retry} retries.")
def _maybe_load_precomputed_tokens(self, audio_path: str, metadata: dict) -> Optional[torch.Tensor]:
if not metadata.get("use_precomputed_tokens"):
return None
root: Optional[Path] = metadata.get("precomputed_tokens_root")
if root is None:
return None
if not root.exists():
logger.warning("Precomputed token root missing: %s", root)
return None
key = os.path.abspath(audio_path)
digest = self._sha1(key.encode("utf-8")).hexdigest()
token_path = root / digest[:2] / digest[2:4] / f"{digest}.pt"
if not token_path.exists():
logger.warning("Precomputed audio tokens not found: %s", token_path)
return None
try:
tokens = torch.load(token_path, map_location="cpu")
if isinstance(tokens, torch.Tensor):
return tokens.clone()
if isinstance(tokens, (list, tuple)):
return torch.tensor(tokens, dtype=torch.long)
logger.warning("Unexpected token format in %s (type=%s)", token_path, type(tokens))
except Exception as exc:
logger.warning("Failed to load precomputed tokens %s: %s", token_path, exc)
return None
class Speech2SpeechDataset(Dataset):
"""
Mixed dataset of emova-sft and InstructS2S-200K.
Return value of __getitem__ indicates a pair of (user, assistant) message (single-turn).
Critically, the return type for emova_sft and instructs2s are different:
emova_sft: tuple[list[int], list[int], Any]
instructs2s: tuple[str, str, Optional[Any]]
So in the main training code, we need to handle both types of return values.
Notes:
- Use `s2s_collate_fn` within DataLoader.
- For emova_sft, Tensor cannot be returned because padding is done later in the main training code.
For reference:
Total samples: 496514
emova-sft (speech-text): 73.6k
emova-sft (speech-image): 71.5k
InstructS2S-200K: 422856
"""
def __init__(self, dataset_configs: list):
self.dataset_configs = dataset_configs # currently this arg is not used
## emova-sft (text + image splits)
emova_sft_text = load_dataset("Emova-ollm/emova-sft-4m", "emova-speech-text-en", split='train')
emova_sft_image = load_dataset("Emova-ollm/emova-sft-4m", "emova-speech-image-en", split='train')
def _maybe_cast_image_columns(ds):
for column in ("image", "images"):
if column in ds.column_names:
try:
ds = ds.cast_column(column, datasets.Image(decode=False))
except Exception:
# Column may already be raw bytes/str; ignore and keep as-is
pass
return ds
emova_sft_text = _maybe_cast_image_columns(emova_sft_text)
emova_sft_image = _maybe_cast_image_columns(emova_sft_image)
def emova_sft_preprocess_batch(batch):
# Extract conversations from the batch
conversations_list = batch['conversations']
usr_ids_list = []
asst_ids_list = []
images_list = []
def normalize_emova_ids(ids: str) -> list[int]:
unit_numbers = ids.replace('<|speech_', '').replace('|>', ' ').strip()
unit_ids = [int(unit) for unit in unit_numbers.split(" ")]
return unit_ids
# Process each conversation in the batch
for conversations in conversations_list:
usr_raw = conversations[0]['value']
asst_raw = conversations[1]['value']
usr_ids: str = usr_raw.split("\n\nuser question speech:")[-1].strip()
asst_ids: str = json.loads(asst_raw)['assistant response speech'].strip()
usr_ids_list.append(normalize_emova_ids(usr_ids))
asst_ids_list.append(normalize_emova_ids(asst_ids))
raw_images = (
batch.get('image')
or batch.get('images')
or batch.get('image_base64')
or [None] * len(conversations_list)
)
if not isinstance(raw_images, (list, tuple)):
raw_images = [raw_images] * len(conversations_list)
else:
raw_images = list(raw_images)
if len(raw_images) != len(conversations_list):
# Align lengths by padding/truncating with None without decoding payloads
adjusted = raw_images[:len(conversations_list)]
if len(adjusted) < len(conversations_list):
adjusted.extend([None] * (len(conversations_list) - len(adjusted)))
raw_images = adjusted
images_list.extend(raw_images)
# Return a dictionary with lists of processed data
return {
"usr_ids": usr_ids_list,
"asst_ids": asst_ids_list,
"image": images_list,
}
self.emova_sft_text = emova_sft_text.map(
emova_sft_preprocess_batch,
batched=True,
batch_size=1024,
remove_columns=['conversations'],
desc="Processing emova-sft (text)",
num_proc=16
)
self.emova_sft_image = emova_sft_image.map(
emova_sft_preprocess_batch,
batched=True,
batch_size=1024,
remove_columns=['conversations'],
desc="Processing emova-sft (image)",
num_proc=16
)
self._emova_text_len = len(self.emova_sft_text)
self._emova_image_len = len(self.emova_sft_image)
self._emova_total_len = self._emova_text_len + self._emova_image_len
## InstructS2S-200K (with caching)
instructs2s_rootdir = "/home/work/AIDAS/data/InstructS2S-200K/en/wav"
self.instructs2s_wav_pair_paths = []
pairs_txt = os.path.join(instructs2s_rootdir, "pairs.txt")
if os.path.isfile(pairs_txt):
with open(pairs_txt, "r") as f:
for line in tqdm(f, desc="Loading InstructS2S-200K paths from cached file"):
line = line.strip()
if not line:
continue
parts = line.split()
if len(parts) >= 2:
self.instructs2s_wav_pair_paths.append((parts[0], parts[1]))
else:
instructs2s_wav_dirs = [p for p in glob(os.path.join(instructs2s_rootdir, "*")) if os.path.isdir(p)]
# Walk each directory and collect (user, assistant) wav pairs
for dir_path in tqdm(instructs2s_wav_dirs, desc="Processing instructs2s-200k"):
dir_name = os.path.basename(dir_path)
k = 1
while True:
user_wav = os.path.join(dir_path, f"{dir_name}-{k}-user.wav")
assistant_wav = os.path.join(dir_path, f"{dir_name}-{k}-assistant.wav")
if os.path.isfile(user_wav) and os.path.isfile(assistant_wav):
self.instructs2s_wav_pair_paths.append((user_wav, assistant_wav))
k += 1
continue
break
with open(pairs_txt, "w") as f:
for u, a in self.instructs2s_wav_pair_paths:
f.write(f"{u} {a}\n")
## Mixed dataset (ordered)
self.mixed_dataset = [self.emova_sft_text, self.emova_sft_image, self.instructs2s_wav_pair_paths]
def __len__(self):
return sum([len(dataset) for dataset in self.mixed_dataset])
def __getitem__(self, idx) -> Union[tuple[list[int], list[int], Any], tuple[str, str, Optional[Any]]]:
if idx < self._emova_text_len: # emova_sft text split
sample = self.emova_sft_text[idx]
elif idx < self._emova_total_len: # emova_sft image split
sample = self.emova_sft_image[idx - self._emova_text_len]
else: # instructs2s
local_idx = idx - self._emova_total_len
usr_wav, asst_wav = self.instructs2s_wav_pair_paths[local_idx]
return usr_wav, asst_wav, None # tuple[str, str, Optional[image]]; wav file paths
usr_ids = sample['usr_ids']
asst_ids = sample['asst_ids']
image = sample.get('image')
return usr_ids, asst_ids, image # tuple[list[int], list[int], image]
def s2s_collate_fn(batch):
"""
Collate function for Speech2SpeechDataset.
"""
emova_data = []
instructs2s_data = []
for item in batch:
if isinstance(item[0], list): # emova_sft: tuple[list[int], list[int]]
emova_data.append(item)
else: # instructs2s: tuple[str, str]
instructs2s_data.append(item)
return {
'emova_sft': emova_data,
'instructs2s': instructs2s_data,
}
class VideoCaptionDataset(Dataset):
def __init__(
self,
transform,
tokenizer,
max_seq_length: int,
resolution: int = 256,
panda70m_path = "/home/work/AIDAS/data/video/panda70m/panda70m_training_2m",
openvid1m_path = "/home/work/AIDAS/data/video/openvid1m/video",
webvid10m_path = "/home/work/AIDAS/data/video/webvid10m",
llavavid_path = "/home/work/AIDAS/data/video/LLaVA-Video-178K",
dataset_name = "openvid1m",
llavavid_local_files_only: bool = False,
llavavid_skip_configs: Optional[Sequence[str]] = None,
llavavid_skip_video_patterns: Optional[Sequence[str]] = None,
sample_method='uniform',
num_frames: int = 8,
vq_model=None,
):
available_datasets = ['panda70m', 'openvid1m', 'webvid10m', 'llavavid']
if dataset_name not in available_datasets:
raise ValueError(f"Invalid dataset name: {dataset_name}. Available datasets: {available_datasets}")
self.max_seq_length = max_seq_length
self.transform = transform
self.vq_model = vq_model
self.tokenizer = tokenizer
self.resolution = resolution
self.sample_method = sample_method
self.dataset_name = dataset_name
self.num_frames = num_frames
self.llavavid_local_files_only = llavavid_local_files_only
self.llavavid_skip_configs = set(llavavid_skip_configs or [])
self.llavavid_skip_video_patterns = tuple(llavavid_skip_video_patterns or [])
self.caption_prompt = V2T_INSTRUCTION
self.caption_prompt = ['<|start_header_id|>user<|end_header_id|>\n' + prompt + '<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n' for prompt in self.caption_prompt]
self.webvid10m_path = webvid10m_path
if dataset_name == 'panda70m':
self.vid_data = self._collect_panda70m(panda70m_path)
self.dataset_root = panda70m_path
elif dataset_name == 'webvid10m':
self.vid_data = self._collect_webvid10m(webvid10m_path)
self.dataset_root = webvid10m_path
elif dataset_name == 'openvid1m':
self.vid_data = self._collect_openvid1m(openvid1m_path)
self.dataset_root = openvid1m_path
elif dataset_name == 'llavavid':
self.vid_data = self._collect_llavavid(llavavid_path)
self.dataset_root = Path(llavavid_path)
self.llavavid_video_root = Path(llavavid_path)
else:
raise ValueError(f"Invalid dataset name: {dataset_name}. Available datasets: panda70m, webvid10m")
def _get_caption_prompt(self):
"""
Get a random caption prompt from the list of caption prompts.
"""
return np.random.choice(self.caption_prompt)
def _tokenize(self, text):
if self.tokenizer is not None:
input_ids = self.tokenizer(
text,
truncation=True,
max_length=2 * self.max_seq_length,
padding=False,
return_tensors="pt"
)[0]
if len(input_ids) > self.max_seq_length:
return None
else:
return input_ids
else:
raise ValueError("Tokenizer is not provided.")
def _collect_webvid10m(self, root_path):
print("Loading videos from WebVid10m dataset...")
csv_path = osp.join(root_path, "webvid-10M-train.csv")
webvid_pd = pd.read_csv(csv_path)
self.dataset_length = len(webvid_pd)
print(f"{len(webvid_pd)} videos has been loaded.")
return webvid_pd
def _collect_panda70m(self, root_path):
video_caption_pairs = []
subdirs = sorted(os.listdir(root_path))
print("Loading videos from panda70m dataset...")
for subdir in subdirs:
full_subdir = os.path.join(root_path, subdir)
if not os.path.isdir(full_subdir):
continue
video_paths = glob(os.path.join(full_subdir, "*.mp4"))
for video_path in video_paths:
caption_path = video_path.replace(".mp4", ".txt")
if os.path.exists(caption_path):
with open(caption_path, 'r') as f:
caption = f.read().strip()
prompt = self._get_caption_prompt()
video_caption_pairs.append({
"video": video_path,
"caption": prompt + caption
})
print(f"{len(video_caption_pairs)} videos has been loaded.")
return video_caption_pairs
def _collect_openvid1m(self, root_path):
csv_path = osp.join(root_path, "OpenVid-1M.csv")
openvid_pd = pd.read_csv(csv_path)
self.dataset_length = len(openvid_pd)
print(f"{len(openvid_pd)} videos has been loaded.")
return openvid_pd
def _collect_llavavid(
self,
root_path="lmms-lab/LLaVA-Video-178K",
cache_dir="/home/work/AIDAS/huggingface/datasets"
):
"""
Collect all available (and locally cached) subsets of the LLaVA-Video-178K dataset.
Handles both on-disk exports (each config stored as subfolders of splits) and remote configs.
Returns a single flattened HuggingFace Dataset that concatenates every successfully loaded config.
"""
DATASET_NAME = root_path
local_root = Path(DATASET_NAME)
configs: list[str]
using_local_dirs = local_root.exists()
configs = []
if using_local_dirs:
for p in sorted(local_root.iterdir()):
if not p.is_dir():
continue
if p.name.startswith("."):
continue
split_exists = any((p / split_name).exists() for split_name in ("open_ended", "caption", "multi_choice"))
if not split_exists:
continue
configs.append(p.name)
if not configs:
using_local_dirs = False
if not configs:
try:
configs = get_dataset_config_names(DATASET_NAME)
using_local_dirs = False
except Exception as e:
raise RuntimeError(f"Failed to fetch configs for {DATASET_NAME}: {e}")
skip_configs = getattr(self, "llavavid_skip_configs", set())
if skip_configs:
existing = [cfg for cfg in configs if cfg in skip_configs]
if existing:
print(f"LLaVA-Vid: skipping configs {existing}")
configs = [cfg for cfg in configs if cfg not in skip_configs]
if not configs:
raise RuntimeError("All LLaVA-Video configs were skipped; nothing left to load.")
def _add_config_column(dataset: HFDataset, cfg_name: str, row_count: int):
"""Attach the originating config name so downstream can locate videos."""
if dataset is None or not cfg_name:
return dataset
if "llavavid_config" in dataset.column_names:
return dataset
return dataset.add_column("llavavid_config", [cfg_name] * row_count)
def _flatten_dataset(ds_obj, label: str, cfg_name: str):
"""Convert DatasetDicts into a single Dataset and report the row count."""
if ds_obj is None:
return None, 0
if isinstance(ds_obj, DatasetDict):
splits = [split for split in ds_obj.values()]
if not splits:
print(f"Skipping {label}: dataset dict has no splits.")
return None, 0
total_rows = sum(len(split) for split in splits)
if len(splits) == 1:
return splits[0], total_rows
try:
merged = concatenate_datasets(splits)
except Exception as merge_err:
print(f"Skipping {label}: failed to concatenate splits: {merge_err}")
return None, 0
dataset = merged
else:
dataset = ds_obj
try:
total_rows = len(dataset)
except Exception as len_err:
print(f"Skipping {label}: unable to compute dataset length ({len_err}).")
return None, 0
dataset = _add_config_column(dataset, cfg_name, total_rows)
return dataset, total_rows
def _load_local_config(cfg_name: str):
"""Attempt to read a single config from disk, handling split sub-directories if needed."""
cfg_root = local_root / cfg_name
if not cfg_root.exists():
return None, 0
# First try loading the directory directly (Dataset or DatasetDict exports).
try:
ds_direct = load_from_disk(str(cfg_root))
except Exception as direct_err:
print(f"Failed to load config {cfg_name} via load_from_disk: {direct_err}.")
else:
ds_flat, ds_count = _flatten_dataset(ds_direct, cfg_name, cfg_name)
if ds_flat is not None and ds_count > 0:
return ds_flat, ds_count
# Fallback: iterate over split sub-directories (caption/open_ended/multi_choice, etc.).
split_dirs = [p for p in sorted(cfg_root.iterdir()) if p.is_dir()]
if not split_dirs:
return None, 0
split_datasets = []
for split_dir in split_dirs:
try:
split_ds = load_from_disk(str(split_dir))
except Exception as split_err:
print(f"Skipping {cfg_name}/{split_dir.name}: {split_err}")
continue
split_datasets.append(split_ds)
if not split_datasets:
return None, 0
split_total = sum(len(split_ds) for split_ds in split_datasets)
if len(split_datasets) == 1:
dataset = split_datasets[0]
else:
try:
dataset = concatenate_datasets(split_datasets)
except Exception as merge_err:
print(f"Skipping {cfg_name}: failed to concatenate split datasets: {merge_err}")
return None, 0
dataset = _add_config_column(dataset, cfg_name, split_total)
return dataset, split_total
datasets_loaded = []
total_count = 0
for cfg in configs:
ds = None
cfg_count = 0
if using_local_dirs:
ds, cfg_count = _load_local_config(cfg)
if ds is None or cfg_count == 0:
download_cfg = None
if self.llavavid_local_files_only:
download_cfg = DownloadConfig(local_files_only=True)
try:
remote_ds = load_dataset(
DATASET_NAME,
name=cfg,
cache_dir=cache_dir,
verification_mode="no_checks",
download_config=download_cfg,
)
except Exception as remote_err:
print(f"Skipping {cfg}: {remote_err}")
continue
ds, cfg_count = _flatten_dataset(remote_ds, cfg, cfg)
if ds is None or cfg_count == 0:
print(f"Skipping {cfg}: dataset empty after flattening.")
continue
datasets_loaded.append(ds)
total_count += cfg_count
if not datasets_loaded:
raise RuntimeError("No valid configs could be loaded!")
if len(datasets_loaded) == 1:
global_dataset = datasets_loaded[0]
else:
try:
global_dataset = concatenate_datasets(datasets_loaded)
except Exception as merge_err:
print(f"Failed to concatenate configs in one step: {merge_err}. Trying pairwise concatenation.")
try:
combined = datasets_loaded[0]
for ds_next in datasets_loaded[1:]:
combined = concatenate_datasets([combined, ds_next])
global_dataset = combined
except Exception as pair_err:
raise RuntimeError(f"Unable to merge LLaVA-Video configs: {pair_err}") from pair_err
# Filter out samples whose video path matches known-bad patterns (e.g., missing shareVideoGPTV frames)
skip_patterns = getattr(self, "llavavid_skip_video_patterns", tuple())
if skip_patterns:
def _matches_skip(entry: dict[str, Any]) -> bool:
video_entry = entry.get("video")
if not isinstance(video_entry, str):
return False
return any(pattern in video_entry for pattern in skip_patterns)
def _filter_dataset(ds_obj):
if isinstance(ds_obj, list):
filtered_list = []
removed_total = 0
for item in ds_obj:
filtered_item, removed_item = _filter_dataset(item)
removed_total += removed_item
if filtered_item is None:
continue
filtered_list.append(filtered_item)
return filtered_list, removed_total
elif isinstance(ds_obj, HFDataset):
before = len(ds_obj)
filtered = ds_obj.filter(lambda ex: not _matches_skip(ex))
removed = before - len(filtered)
return filtered, removed
elif isinstance(ds_obj, dict):
return (None, 1) if _matches_skip(ds_obj) else (ds_obj, 0)
else:
return ds_obj, 0
global_dataset, removed_samples = _filter_dataset(global_dataset)
if removed_samples > 0:
total_count -= removed_samples
print(f"LLaVA-Vid: skipped {removed_samples} samples matching patterns {skip_patterns}.")
print(f"LLaVA-Vid: {len(datasets_loaded)} configs loaded.")
print(f"LLaVA-Vid: {total_count:,} total samples loaded.")
self.dataset_length = total_count
return global_dataset
def __len__(self):
return len(self.vid_data)
def __getitem__(self, idx):
max_try_count = 50
for try_count in range(max_try_count):
try:
data = self._sample_data(idx)
except Exception as exc:
logger.warning(
"VideoCaptionDataset failed to fetch index %s on attempt %s/%s: %s",
idx,
try_count + 1,
max_try_count,
exc,
)
idx = random.randint(0, self.dataset_length - 1)
continue
if data is not None:
return {
"video": data["video"],
"caption": data["caption"],
}
idx = random.randint(0, self.dataset_length - 1)
logger.warning(
"VideoCaptionDataset exhausted %s attempts without a valid sample; returning None.",
max_try_count,
)
return None
def _sample_data_webvid10m(self):
store_path = osp.join(self.webvid10m_path, "video_store")
row = self.video_caption_pairs['webvid10m'].sample(1).iloc[0]
video_id = str(row["videoid"])
url = row["contentUrl"]
caption = row["name"]
video_path = osp.join(store_path, f"{video_id}.mp4")
if not osp.exists(video_path): # not downloaded yet
download_video_url(url, video_path)
# print(video_id)
# print(_whoami_str())
return video_path, caption
def _sample_data(self, idx):
if self.dataset_name == 'webvid10m':
# currently randomly sample from the dataset
video_path, caption = self._sample_data_webvid10m()
elif self.dataset_name == 'panda70m':
raise NotImplementedError("Panda70m is not implemented yet.")
# video_path, caption = self._sample_data_panda70m()
elif self.dataset_name == 'openvid1m':
data_row = self.vid_data.iloc[idx]
video_path = osp.join(self.dataset_root, "video", data_row["video"])
caption = data_row["caption"]
elif self.dataset_name == 'llavavid':
data_row = self.vid_data[idx]
video_entry = data_row['video']
cfg_name = data_row.get('llavavid_config') if isinstance(data_row, dict) else None
caption = data_row['conversations'] # this is a list of turns in llavavid
resolved_video_path = None
if isinstance(video_entry, str):
candidate_paths = []
video_path_obj = Path(video_entry)
if video_path_obj.is_absolute() and video_path_obj.exists():
resolved_video_path = video_path_obj
else:
if hasattr(self, "llavavid_video_root"):
base_root = Path(self.llavavid_video_root)
if cfg_name:
candidate_paths.append(base_root / cfg_name / video_entry)
candidate_paths.append(base_root / video_entry)
# Also allow treating the stored value as relative to current dir.
candidate_paths.append(Path(video_entry))
for candidate in candidate_paths:
if candidate.exists():
resolved_video_path = candidate
break
if resolved_video_path is None:
logger.warning(
"LLaVA-Video sample missing video file: %s (config=%s)",
video_entry,
cfg_name,
)
return None
if resolved_video_path.suffix.lower() == ".mkv":
logger.warning(
"LLaVA-Video skipping MKV file: %s (config=%s)",
resolved_video_path,
cfg_name,
)
return None
video_path = str(resolved_video_path)
else:
raise ValueError(f"Invalid dataset name: {self.dataset_name}. Available datasets: panda70m, webvid10m, openvid1m")
try:
frames = load_video_mp4(
video_path=video_path,
sample_method=self.sample_method,
num_frames=self.num_frames,
resolution=self.resolution,
transform=self.transform,
strict=False,
)
except Exception as exc:
logger.warning(
"LLaVA-Video sample failed to load (%s): %s",
video_path,
exc,
)
return None
if frames is None:
logger.warning(
"LLaVA-Video sample timed out while reading frames (%s); skipping sample.",
video_path,
)
return None
return {
"video": frames, # torch tensor (T, C, H, W)
"caption": caption # input_ids (seq_len); str
}
def download_video_url(url: str, save_path, timeout=10, max_retries=3) -> bool:
for attempt in range(1, max_retries + 1):
try:
with requests.get(url, stream=True, timeout=timeout) as r:
r.raise_for_status()
with open(save_path, 'wb') as f:
for chunk in r.iter_content(chunk_size=8192):
if chunk:
f.write(chunk)
return True # Success
except Exception as e:
print(f"[Attempt {attempt}/{max_retries}] Download failed: {e}")
if attempt < max_retries:
sleep_time = 2 ** (attempt - 1) # exponential backoff: 1,2,4,8,...
time.sleep(sleep_time)
else:
return False # all attempts failed
return False
def load_video_mp4(
video_path,
sample_method: str = 'uniform',
num_frames: int = 8,
resolution: int = 256,
transform=None,
*,
per_frame_timeout: float = 1.5,
read_retry_interval: float = 0.05,
strict: bool = True,
):
"""
Load video frames and return them as a list of PIL images.
Args:
video_path: Path to the video file.
sample_method: Sampling method, 'uniform' or 'random'.
num_frames: Number of frames to sample from the video.
per_frame_timeout: Max seconds to block while seeking/reading a frame.
read_retry_interval: Delay between read retries while waiting for a frame.
strict: When False, return None on timeout/seek failure instead of raising.
Returns:
List[Image.Image] | None (if strict=False and a timeout/seek failure occurs)
"""
with open(os.devnull, "w") as devnull, contextlib.redirect_stderr(devnull):
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
raise IOError(f"Could not open video file {video_path}")
if per_frame_timeout <= 0:
per_frame_timeout = 0.1
if read_retry_interval <= 0:
read_retry_interval = 0.01
def _read_frame_with_timeout(frame_index: Optional[int] = None):
deadline = time.monotonic() + per_frame_timeout
attempts = 0
while True:
if frame_index is not None:
cap.set(cv2.CAP_PROP_POS_FRAMES, int(frame_index))
ret, frame = cap.read()
if ret and frame is not None:
return frame
attempts += 1
if time.monotonic() >= deadline:
return None
time.sleep(min(read_retry_interval, max(deadline - time.monotonic(), 0.0)))
try:
frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
except Exception:
frame_count = -1
if frame_count is None or frame_count <= 0:
# Fallback: attempt to read sequentially but stop early on failure
frames = []
try:
while len(frames) < num_frames:
frame = _read_frame_with_timeout()
if frame is None:
break
frames.append(Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)))
finally:
cap.release()
if len(frames) < num_frames:
msg = f"Video {video_path} has insufficient frames ({len(frames)})."
if strict:
raise ValueError(msg)
logger.warning("%s Skipping sample.", msg)
return None
selected = frames
else:
if frame_count < num_frames:
cap.release()
msg = f"Video {video_path} has insufficient frames ({frame_count})."
if strict:
raise ValueError(msg)
logger.warning("%s Skipping sample.", msg)
return None
if sample_method == 'uniform':
indices = np.linspace(0, frame_count - 1, num_frames).astype(int)
elif sample_method == 'random':
indices = np.sort(np.random.choice(frame_count, num_frames, replace=False))
else:
cap.release()
raise ValueError(f"Sampling method {sample_method} not supported.")
selected = []
try:
for idx in indices:
frame = _read_frame_with_timeout(idx)
if frame is None:
msg = (
f"Timed out ({per_frame_timeout:.2f}s) seeking frame {idx} in {video_path}"
)
if strict:
raise TimeoutError(msg)
logger.warning("%s. Skipping sample.", msg)
return None
selected.append(Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)))
finally:
cap.release()
sampled_frames = []
for frame in selected:
if transform:
frame = transform(frame, resolution=resolution)
sampled_frames.append(frame)
return sampled_frames
class VideoSpeechDataset(Dataset):
"""Loads paired video clips and speech audio paths or pre-tokenized speech."""
def __init__(
self,
*,
transform=None,
resolution: int = 256,
num_frames: int = 8,
video_root: str = "/home/work/AIDAS/data/video/openvid1m/video/video",
audio_root: str = "/home/work/AIDAS/data/video-speech",
speech_dir_name: str = "openvid-speech-trunc",
index_path: str = "/home/work/AIDAS/data/video-speech/openvid-speech.csv",
sample_method: str = "uniform",
precomputed_tokens_root: Optional[str] = None,
) -> None:
self.transform = transform
self.resolution = resolution
self.num_frames = num_frames
self.sample_method = sample_method or "uniform"
if self.sample_method not in {"uniform", "random"}:
logger.warning("Unknown sample_method '%s', defaulting to 'uniform'", self.sample_method)
self.sample_method = "uniform"
self.video_root = Path(video_root).expanduser().resolve()
audio_base = Path(audio_root).expanduser()
if speech_dir_name:
audio_base = audio_base / speech_dir_name
self.audio_root = audio_base.resolve()
self.index_path = Path(index_path).expanduser().resolve()
if not self.index_path.exists():
raise FileNotFoundError(f"VideoSpeechDataset index not found: {self.index_path}")
self.precomputed_tokens_root = (
Path(precomputed_tokens_root).expanduser().resolve()
if precomputed_tokens_root
else None
)
if self.precomputed_tokens_root is not None and not self.precomputed_tokens_root.exists():
logger.warning(
"Precomputed speech token root %s missing; falling back to raw audio paths.",
self.precomputed_tokens_root,
)
self.precomputed_tokens_root = None
self._samples: list[tuple[Path, Path]] = []
self._token_cache: Dict[str, torch.Tensor] = {}
self._token_cache_limit = 4096
self._load_index()
if not self._samples:
raise RuntimeError(f"VideoSpeechDataset found no valid samples in {self.index_path}")
def _load_index(self) -> None:
missing = 0
with self.index_path.open("r", newline="") as csvfile:
reader = csv.reader(csvfile)
for row in reader:
if not row:
continue
base = row[0].strip()
if not base:
continue
if base.lower().endswith(".wav"):
base = base[:-4]
video_path = self.video_root / f"{base}.mp4"
audio_path = self.audio_root / f"{base}.wav"
if not video_path.is_file() or not audio_path.is_file():
missing += 1
continue
self._samples.append((video_path, audio_path))
if missing:
logger.info(
"VideoSpeechDataset skipped %d entries missing media (index=%s)",
missing,
self.index_path,
)
def __len__(self) -> int:
return len(self._samples)
def _transform_frame(self, image: Image.Image, resolution: int) -> torch.Tensor:
if self.transform is None:
return utils_image_transform(image, resolution)
try:
return self.transform(image, resolution=resolution)
except TypeError:
return self.transform(image)
def _resolve_token_path(self, audio_path: Path) -> Optional[Path]:
if self.precomputed_tokens_root is None:
return None
digest = hashlib.sha1(os.path.abspath(str(audio_path)).encode("utf-8")).hexdigest()
return self.precomputed_tokens_root / digest[:2] / digest[2:4] / f"{digest}.pt"
def _get_precomputed_tokens(self, audio_path: Path) -> Optional[torch.Tensor]:
cache_key = os.path.abspath(str(audio_path))
cached = self._token_cache.get(cache_key)
if cached is not None:
return cached.clone()
token_path = self._resolve_token_path(audio_path)
if token_path is None or not token_path.exists():
return None
try:
tokens = torch.load(token_path, map_location="cpu")
except Exception as exc:
logger.warning("Failed to load precomputed speech tokens %s: %s", token_path, exc)
return None
if not isinstance(tokens, torch.Tensor):
return None
tokens = tokens.to(dtype=torch.long, copy=False)
if len(self._token_cache) < self._token_cache_limit:
self._token_cache[cache_key] = tokens
return tokens.clone()
def _prepare_speech_entry(self, audio_path: Path):
tokens = self._get_precomputed_tokens(audio_path)
if tokens is not None:
return tokens
return str(audio_path)
def __getitem__(self, idx: int) -> Dict[str, Any]:
video_path, audio_path = self._samples[idx]
frames = load_video_mp4(
str(video_path),
sample_method=self.sample_method,
num_frames=self.num_frames,
resolution=self.resolution,
transform=self._transform_frame,
)
speech_entry = self._prepare_speech_entry(audio_path)
return {
"video": frames,
"speech": speech_entry,
}
class TextImageInterleavedDataset:
"""
HF-backed dataset that yields rows of:
{
"image_paths": [str, ...], # absolute paths (no decoding)
"user_text": str,
"assistant_text": str,
}
"""
def __init__(
self,
*,
configs: Union[str, Sequence[str], None] = None, # default: all configs
split: str = "train",
data_root: str = "/home/work/AIDAS/data/TIGER-Lab/Mantis-Instruct",
max_images: Optional[int] = None,
filter_empty: bool = True,
resolution: int = 256,
# sampling controls
per_config_fraction: float = 1/7, # ← sample 1/7 PER CONFIG
sample_seed: int = 42,
# kept for compatibility, not used in this 1/7-per-config version
max_samples: Optional[int] = 1_000_000,
local_data_root: Optional[str] = None,
local_data_files: Optional[Dict[str, Any]] = None,
local_files_only: bool = False,
):
self.data_root = data_root
self.split = self._normalize_split(split)
self.max_images = max_images
self.filter_empty = filter_empty
self.resolution = resolution
self.local_data_root = local_data_root
self.local_data_files = local_data_files or {}
self._download_config = DownloadConfig(local_files_only=True) if local_files_only else None
# cache transforms
self._tfm_crop = transforms.Compose([
transforms.Resize(resolution, interpolation=transforms.InterpolationMode.BICUBIC),
transforms.CenterCrop((resolution, resolution)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5,0.5,0.5], std=[0.5,0.5,0.5]),
])
self._tfm_squash = transforms.Compose([
transforms.Resize((resolution, resolution), interpolation=transforms.InterpolationMode.BICUBIC),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5,0.5,0.5], std=[0.5,0.5,0.5]),
])
# ---- resolve configs ----
if configs is None or configs == "all":
cfgs = self._resolve_configs_from_local()
if not cfgs:
cfgs = sorted(get_dataset_config_names("TIGER-Lab/Mantis-Instruct"))
elif isinstance(configs, str):
cfgs = [configs]
else:
cfgs = list(configs)
self.configs = cfgs
rng = np.random.default_rng(sample_seed)
per_cfg_ds: List[HFDataset] = []
acc = Accelerator()
for cfg in cfgs:
base_ds = self._load_base_dataset(cfg, acc)
if base_ds is None:
continue
# --- SAMPLE 1/7 OF BASE ROWS *PER CONFIG* (before any map/expansion) ---
n = len(base_ds)
if n == 0:
continue
k = max(1, int(np.floor(n * per_config_fraction)))
# reproducible uniform sample without replacement
sel_idx = rng.choice(n, size=k, replace=False)
base_ds = base_ds.select(list(sel_idx))
# locate image dir for this (cfg, split)
img_dir = self._resolve_img_dir(cfg, self.split)
if img_dir is None:
raise FileNotFoundError(f"No image dir for config='{cfg}', split='{self.split}'")
# (1) attach constants
def add_const_cols(batch):
m = len(next(iter(batch.values()))) if batch else 0
return {"config": [cfg]*m, "img_dir": [img_dir]*m}
ds = base_ds.map(add_const_cols, batched=True)
# (2) normalize image column → absolute string paths
image_key = self._guess_image_key(ds.column_names)
def make_abs_paths(batch):
bases = batch["img_dir"] # list[str] per row
rels = batch[image_key] # per-row: list[dict]|dict|list[str]|str|None
def dict_to_rel(d: Dict[str, Any]) -> Optional[str]:
# typical HF Image: {"path": "...", "bytes": ...}
for k in ("path", "file_name", "filepath", "image_path", "name"):
v = d.get(k)
if isinstance(v, str) and v:
return v
# nested
img = d.get("image")
if isinstance(img, dict):
v = img.get("path")
if isinstance(v, str) and v:
return v
return None
out_paths = []
for base, r in zip(bases, rels):
# normalize r → list[str]
if r is None:
row = []
elif isinstance(r, str):
row = [r]
elif isinstance(r, dict):
s = dict_to_rel(r)
row = [s] if s else []
elif isinstance(r, list):
tmp = []
for x in r:
if isinstance(x, str):
tmp.append(x)
elif isinstance(x, dict):
s = dict_to_rel(x)
if s:
tmp.append(s)
row = tmp
else:
row = []
# join to absolute (keep absolute if already)
abs_paths = [p if os.path.isabs(p) else os.path.join(base, p) for p in row if isinstance(p, str)]
# cap if requested
if self.max_images is not None and len(abs_paths) > self.max_images:
abs_paths = abs_paths[: self.max_images]
out_paths.append(abs_paths)
return {"image_paths": out_paths}
ds = ds.map(make_abs_paths, batched=True)
# (3) expand conversation: one row per (user → assistant) turn
conv_key = "conversation"
def expand_turns(batch):
image_paths_list = batch["image_paths"]
conversations = batch.get(conv_key, [[]] * len(image_paths_list))
out_img_paths, out_user, out_assistant = [], [], []
for img_paths, conv in zip(image_paths_list, conversations):
conv = conv or []
# walk adjacent pairs
i = 0
while i < len(conv) - 1:
a, b = conv[i], conv[i + 1]
if (isinstance(a, dict) and isinstance(b, dict)
and a.get("role") == "user" and b.get("role") == "assistant"):
user_text = (a.get("content") or "").strip()
assistant_text = (b.get("content") or "").strip()
if (not self.filter_empty) or assistant_text:
out_img_paths.append(img_paths)
out_user.append(user_text)
out_assistant.append(assistant_text)
i += 2
else:
i += 1
return {
"image_paths": out_img_paths,
"user_text": out_user,
"assistant_text": out_assistant,
}
ds = ds.map(expand_turns, batched=True, remove_columns=ds.column_names)
if self.filter_empty:
ds = ds.filter(lambda e: bool(e["assistant_text"]))
per_cfg_ds.append(ds)
if not per_cfg_ds:
raise ValueError("Empty dataset after per-config sampling and preprocessing.")
self.dataset = concatenate_datasets(per_cfg_ds) if len(per_cfg_ds) > 1 else per_cfg_ds[0]
self.dataset = self.dataset.with_format("python")
print(f"[HF Dataset] per-config 1/7 sampled; configs={self.configs}, split='{self.split}', rows={len(self.dataset)}")
# ---- public API ----
def __len__(self):
return len(self.dataset)
def __getitem__(self, idx):
start_idx = idx
attempts = 0
max_attempts = 10
while attempts < max_attempts:
ex = self.dataset[idx]
text = (
"<|start_header_id|>user<|end_header_id|>\n"
f"{ex['user_text']}\n"
"<eot_id><|start_header_id|>assistant<|end_header_id|>\n"
f"{ex['assistant_text']}"
)
paths = ex["image_paths"]
imgs: list[torch.Tensor] = []
for path in paths:
img = self._load_and_transform_one(path)
if img is not None:
imgs.append(img)
if imgs:
return {
"images": imgs,
"text": text,
}
attempts += 1
idx = (idx + 1) % len(self.dataset)
if idx == start_idx:
break
raise RuntimeError("TextImageInterleavedDataset: no valid images found after retries.")
# ---- helpers ----
@staticmethod
def _normalize_split(split: str) -> str:
s = split.lower()
return {"val": "validation", "dev": "validation"}.get(s, s)
def _resolve_img_dir(self, cfg: str, split: str) -> Optional[str]:
# Typical local layout:
# {data_root}/{cfg}/{split}_images
# {data_root}/{cfg}/images
cand1 = os.path.join(self.data_root, cfg, f"{split}_images")
cand2 = os.path.join(self.data_root, cfg, "images")
for c in (cand1, cand2):
if os.path.isdir(c):
return c
return None
def _load_and_transform_one(self, path: str):
try:
with Image.open(path) as im:
im = im.convert("RGB")
except FileNotFoundError:
return None
except Exception:
return None
return self._tfm_crop(im)
@staticmethod
def _guess_image_key(cols: List[str]) -> str:
for k in ("images", "image_paths", "imgs", "paths", "image"):
if k in cols:
return k
raise KeyError(f"Cannot find image column among {cols}")
def _resolve_configs_from_local(self) -> List[str]:
cfgs: List[str] = []
if self.local_data_root:
root = Path(self.local_data_root)
if root.is_dir():
for entry in sorted(root.iterdir()):
if not entry.is_dir():
continue
if self._has_split_data(entry):
cfgs.append(entry.name)
if not cfgs and self.local_data_files:
cfgs = [k for k in sorted(self.local_data_files.keys()) if k != "default"]
return cfgs
def _has_split_data(self, cfg_path: Path) -> bool:
split_dir = cfg_path / self.split
if split_dir.is_dir():
return True
alt_dirs = [
cfg_path / f"{self.split}.dataset",
cfg_path / f"{self.split}.arrow",
]
for candidate in alt_dirs:
if candidate.is_dir():
return True
patterns = [
cfg_path / self.split / "*.arrow",
cfg_path / self.split / "*.parquet",
cfg_path / f"{self.split}/*.arrow",
cfg_path / f"{self.split}/*.parquet",
cfg_path / f"{self.split}*.arrow",
cfg_path / f"{self.split}*.parquet",
]
for pattern in patterns:
if glob(str(pattern)):
return True
return False
def _load_base_dataset(self, cfg: str, acc: Accelerator) -> Optional[HFDataset]:
base_ds: Optional[HFDataset] = None
if self.local_data_root is not None:
# print(self.local_data_root)
base_ds = self._load_from_local_root(cfg)
if base_ds is None and self.local_data_files:
base_ds = self._load_from_local_data_files(cfg)
if base_ds is not None:
return base_ds
kwargs = {}
if self._download_config is not None:
kwargs["download_config"] = self._download_config
if acc.num_processes > 1:
acc.wait_for_everyone()
try:
base_ds = load_dataset(
"TIGER-Lab/Mantis-Instruct",
cfg,
split=self.split,
**kwargs,
)
except Exception as exc:
if self._download_config is not None:
raise RuntimeError(
f"Failed to load local dataset for config='{cfg}'. "
"Ensure that the dataset is cached or provide 'local_data_root'."
) from exc
raise
finally:
if acc.num_processes > 1:
acc.wait_for_everyone()
return base_ds
def _load_from_local_root(self, cfg: str) -> Optional[HFDataset]:
cfg_root = os.path.join(self.local_data_root, cfg)
if not os.path.exists(cfg_root):
return None
candidates = [
cfg_root,
os.path.join(cfg_root, self.split),
os.path.join(cfg_root, f"{self.split}.dataset"),
]
for path in candidates:
if not os.path.isdir(path):
continue
try:
loaded = load_from_disk(path)
if isinstance(loaded, DatasetDict):
if self.split in loaded:
return loaded[self.split]
continue
return loaded
except Exception:
continue
patterns = [
os.path.join(cfg_root, f"{self.split}.parquet"),
os.path.join(cfg_root, f"{self.split}/*.parquet"),
os.path.join(cfg_root, f"{self.split}_*.parquet"),
os.path.join(cfg_root, f"{self.split}.json"),
os.path.join(cfg_root, f"{self.split}.jsonl"),
os.path.join(cfg_root, f"{self.split}/*.jsonl"),
os.path.join(cfg_root, f"{self.split}.arrow"),
os.path.join(cfg_root, f"{self.split}/*.arrow"),
]
for pattern in patterns:
files = sorted(glob(pattern))
if files:
return self._load_from_files(files)
return None
def _load_from_local_data_files(self, cfg: str) -> Optional[HFDataset]:
spec = self.local_data_files.get(cfg) or self.local_data_files.get("default")
if spec is None:
return None
if isinstance(spec, str):
entries = [spec]
loader = None
elif isinstance(spec, dict):
loader = spec.get("type") or spec.get("loader") or spec.get("format")
files = spec.get(self.split) or spec.get("files")
if files is None:
return None
entries = files if isinstance(files, list) else [files]
else:
entries = list(spec)
loader = None
resolved_files: list[str] = []
for entry in entries:
if not entry:
continue
matched = sorted(glob(entry))
if matched:
resolved_files.extend(matched)
elif os.path.exists(entry):
resolved_files.append(entry)
if not resolved_files:
return None
return self._load_from_files(resolved_files, loader_hint=loader)
def _load_from_files(self, files: list[str], loader_hint: Optional[str] = None) -> Optional[HFDataset]:
if not files:
return None
ext = Path(files[0]).suffix.lower()
loader = loader_hint
if loader is None:
if ext in (".parquet",):
loader = "parquet"
elif ext in (".json", ".jsonl"):
loader = "json"
elif ext in (".arrow", ".feather"):
loader = "arrow"
if loader == "parquet":
return load_dataset("parquet", data_files={self.split: files}, split=self.split)
if loader in {"json", "jsonl"}:
return load_dataset("json", data_files={self.split: files}, split=self.split)
if loader == "arrow":
datasets = [HFDataset.from_file(path) for path in files]
return concatenate_datasets(datasets) if len(datasets) > 1 else datasets[0]
return None
class HFInstructionTextDataset(Dataset):
"""Mixed instruction-following text dataset sourced from multiple HF corpora."""
HF_SOURCES = (
{
"name": "openai/gsm8k",
"config": "main",
"split": "train",
"user_key": "question",
"assistant_key": "answer",
},
{
"name": "qwedsacf/grade-school-math-instructions",
"config": None,
"split": "train",
"user_key": "INSTRUCTION",
"assistant_key": "RESPONSE",
},
{
"name": "alespalla/chatbot_instruction_prompts",
"config": None,
"split": "train",
"user_key": "prompt",
"assistant_key": "response",
},
{
"name": "TIGER-Lab/MathInstruct",
"config": None,
"split": "train",
"user_key": "instruction",
"assistant_key": "output",
},
)
def __init__(
self,
*,
split: str = "train",
max_samples_per_source: Optional[int] = None,
max_total_samples: Optional[int] = None,
seed: int = 42,
) -> None:
self.split = split
self.seed = seed
self.samples: List[str] = []
rng = random.Random(seed)
for source in self.HF_SOURCES:
desired_split = source.get("split", split)
try:
dataset_name = source["name"]
dataset_config = source.get("config")
if dataset_config is not None:
hf_ds = load_dataset(dataset_name, dataset_config, split=desired_split)
else:
hf_ds = load_dataset(dataset_name, split=desired_split)
except Exception as exc:
print(f"[HFInstructionTextDataset] Failed to load {source['name']}: {exc}")
continue
if max_samples_per_source is not None and len(hf_ds) > max_samples_per_source:
hf_ds = hf_ds.shuffle(seed=seed).select(range(max_samples_per_source))
user_key = source["user_key"]
assistant_key = source["assistant_key"]
for example in hf_ds:
user_raw = str(example.get(user_key, "")).strip()
assistant_raw = str(example.get(assistant_key, "")).strip()
if not user_raw or not assistant_raw:
continue
formatted = self._format_dialogue(user_raw, assistant_raw)
if formatted:
self.samples.append(formatted)
if not self.samples:
raise ValueError("HFInstructionTextDataset loaded zero valid samples.")
rng.shuffle(self.samples)
if max_total_samples is not None:
self.samples = self.samples[: max_total_samples]
@staticmethod
def _format_dialogue(user_text: str, assistant_text: str) -> str:
return (
"<|start_header_id|>user<|end_header_id|>\n"
f"{user_text}\n"
"<|eot_id><|start_header_id|>assistant<|end_header_id|>\n"
f"{assistant_text}"
)
def __len__(self) -> int:
return len(self.samples)
def __getitem__(self, index: int) -> Dict[str, str]:
return {"input_ids": self.samples[index]}
@staticmethod
def collate_fn(batch: List[Dict[str, str]]) -> Dict[str, List[str]]:
return {"input_ids": [example["input_ids"] for example in batch]}
class TextToImage2MDataset(Dataset):
"""Loads jackyhate/text-to-image-2M for text-to-image fine-tuning."""
def __init__(
self,
split: str = "train",
resolution: int = 256,
dataset_name: str = "jackyhate/text-to-image-2M",
cache_dir: str | None = None,
local_files_only: bool = False,
) -> None:
self.resolution = resolution
self.dataset_name = dataset_name
self.cache_dir = cache_dir
self.local_files_only = local_files_only
download_cfg = None
if local_files_only:
download_cfg = DownloadConfig(local_files_only=True)
self._dataset = load_dataset(
dataset_name,
split=split,
cache_dir=cache_dir,
download_config=download_cfg,
)
def __len__(self) -> int:
return len(self._dataset)
def __getitem__(self, idx: int) -> Dict[str, Any]:
sample = self._dataset[idx]
prompt = None
json_meta = sample.get("json")
if isinstance(json_meta, dict):
prompt = json_meta.get("prompt")
if prompt is None:
prompt = sample.get("prompt", "")
image_field = sample.get("jpg") or sample.get("image")
if image_field is None:
raise KeyError("Expected image field 'jpg' in text-to-image-2M sample")
if isinstance(image_field, Image.Image):
image = image_field.convert("RGB")
elif isinstance(image_field, bytes):
image = Image.open(BytesIO(image_field)).convert("RGB")
else:
image = Image.fromarray(np.array(image_field)).convert("RGB")
image_tensor = utils_image_transform(image, self.resolution)
return {
"input_prompt": prompt,
"output_prompt": None,
"edit_prompt": None,
"inverse_prompt": None,
"input_image": image_tensor,
"output_image": image_tensor,
}
class HQEditX2IDataset(Dataset):
def __init__(
self,
split: str = "train",
resolution: int = 256,
dataset_name: str = "UCSC-VLAA/HQ-Edit",
cache_dir: str = "/home/work/AIDAS/huggingface/datasets",
):
self.resolution = resolution
self.cache_dir = cache_dir # retained for API compatibility
self._dataset = load_dataset(dataset_name, split=split)
def __len__(self) -> int:
return len(self._dataset)
def __getitem__(self, idx: int) -> Dict[str, Any]:
sample = self._dataset[idx]
input_tensor = utils_image_transform(sample['input_image'].convert("RGB"), self.resolution)
output_tensor = utils_image_transform(sample['output_image'].convert("RGB"), self.resolution)
return {
"input_prompt": sample["input"],
"output_prompt": sample["output"],
"edit_prompt": sample["edit"],
"inverse_prompt": sample["inverse_edit"],
"input_image": input_tensor,
"output_image": output_tensor,
}
class CombinedX2IDataset(Dataset):
"""Round-robin combination of multiple x2i-style datasets."""
def __init__(self, datasets: Sequence[Dataset]):
if not datasets:
raise ValueError("CombinedX2IDataset requires at least one dataset.")
self.datasets = list(datasets)
self.lengths = [len(ds) for ds in self.datasets]
if any(length == 0 for length in self.lengths):
raise ValueError("Underlying x2i dataset has zero length.")
self.cumulative = list(itertools.accumulate(self.lengths))
self.total_length = self.cumulative[-1]
def __len__(self) -> int:
return self.total_length
def __getitem__(self, idx: int) -> Dict[str, Any]:
if idx < 0 or idx >= self.total_length:
raise IndexError(f"Index {idx} out of bounds for CombinedX2IDataset of length {self.total_length}")
dataset_idx = bisect.bisect_right(self.cumulative, idx)
prev = self.cumulative[dataset_idx - 1] if dataset_idx > 0 else 0
local_idx = idx - prev
return self.datasets[dataset_idx][local_idx]
class OpenImageI2IDataset(Dataset):
"""
Image-to-image dataset built from local Open Images edit JSONL files.
Supports three JSONL schemas:
* SFT-style single turn edits (text + output_image + local_input_image)
* Preference data; only positive edits (output_image) are used by default
* Multi-turn edits which are flattened into single-turn pairs
"""
def __init__(
self,
resolution: int = 256,
image_root: str | None = None,
sft_jsonl: Union[str, Sequence[str], None] = None,
pref_jsonl: Union[str, Sequence[str], None] = None,
multi_turn_jsonl: Union[str, Sequence[str], None] = None,
prefer_summarized_text: bool = True,
pref_positive_only: bool = True,
skip_missing: bool = True,
max_samples_per_source: int | None = None,
max_total_samples: int | None = None,
seed: int | None = None,
) -> None:
self.resolution = resolution
self.image_root = image_root
self.prefer_summarized_text = prefer_summarized_text
self.pref_positive_only = pref_positive_only
self.skip_missing = skip_missing
self._rng = random.Random(seed if seed is not None else 0)
self._per_source_limit = self._coerce_positive_int(max_samples_per_source)
self._total_limit = self._coerce_positive_int(max_total_samples)
self._samples: list[dict[str, str]] = []
self._stats: dict[str, int] = {
"sft": 0,
"pref": 0,
"multi_turn": 0,
"missing_paths": 0,
"invalid_records": 0,
}
sft_paths = self._coerce_paths(sft_jsonl)
pref_paths = self._coerce_paths(pref_jsonl)
multi_turn_paths = self._coerce_paths(multi_turn_jsonl)
for path in sft_paths:
self._samples.extend(self._load_single_turn_file(path, source_key="sft"))
for path in pref_paths:
if not self.pref_positive_only:
logger.warning("OpenImageI2IDataset currently only supports positive preference pairs.")
self._samples.extend(self._load_single_turn_file(path, source_key="pref"))
for path in multi_turn_paths:
self._samples.extend(self._load_multi_turn_file(path))
if self._total_limit is not None and len(self._samples) > self._total_limit:
self._rng.shuffle(self._samples)
self._samples = self._samples[: self._total_limit]
if not self._samples:
raise ValueError("OpenImageI2IDataset could not load any valid examples.")
logger.info(
"Loaded %d OpenImage i2i samples (sft=%d, pref=%d, multi_turn=%d, missing_paths=%d, invalid=%d).",
len(self._samples),
self._stats["sft"],
self._stats["pref"],
self._stats["multi_turn"],
self._stats["missing_paths"],
self._stats["invalid_records"],
)
def __len__(self) -> int:
return len(self._samples)
def __getitem__(self, idx: int) -> Dict[str, Any]:
record = self._samples[idx]
input_image = Image.open(record["input_path"]).convert("RGB")
target_image = Image.open(record["target_path"]).convert("RGB")
input_tensor = utils_image_transform(input_image, self.resolution)
target_tensor = utils_image_transform(target_image, self.resolution)
prompt = record["prompt"]
return {
"input_prompt": prompt,
"output_prompt": None,
"edit_prompt": prompt,
"inverse_prompt": None,
"input_image": input_tensor,
"output_image": target_tensor,
}
def _load_single_turn_file(self, path: str, *, source_key: str) -> list[dict[str, str]]:
file_path = os.path.abspath(os.path.expanduser(path))
if not os.path.exists(file_path):
logger.warning("OpenImageI2IDataset: JSONL file not found: %s", file_path)
return []
base_dir = os.path.dirname(file_path)
samples: list[dict[str, str]] = []
with open(file_path, "r", encoding="utf-8") as handle:
for line in handle:
line = line.strip()
if not line:
continue
try:
record = json.loads(line)
except json.JSONDecodeError:
self._stats["invalid_records"] += 1
continue
prompt = self._select_prompt(record)
input_path = self._resolve_path(record.get("local_input_image"), base_dir=base_dir)
output_path = self._resolve_path(record.get("output_image"), base_dir=base_dir)
sample = self._build_sample(prompt, input_path, output_path)
if sample:
samples.append(sample)
if self._per_source_limit is not None and len(samples) > self._per_source_limit:
self._rng.shuffle(samples)
samples = samples[: self._per_source_limit]
self._stats[source_key] += len(samples)
return samples
def _load_multi_turn_file(self, path: str) -> list[dict[str, str]]:
file_path = os.path.abspath(os.path.expanduser(path))
if not os.path.exists(file_path):
logger.warning("OpenImageI2IDataset: JSONL file not found: %s", file_path)
return []
base_dir = os.path.dirname(file_path)
samples: list[dict[str, str]] = []
with open(file_path, "r", encoding="utf-8") as handle:
for line in handle:
line = line.strip()
if not line:
continue
try:
record = json.loads(line)
except json.JSONDecodeError:
self._stats["invalid_records"] += 1
continue
multi_samples = self._expand_multi_turn(record, base_dir=base_dir)
if multi_samples:
samples.extend(multi_samples)
if self._per_source_limit is not None and len(samples) > self._per_source_limit:
self._rng.shuffle(samples)
samples = samples[: self._per_source_limit]
self._stats["multi_turn"] += len(samples)
return samples
def _expand_multi_turn(self, record: dict, *, base_dir: str) -> list[dict[str, str]]:
prompts = record.get("metadata_edit_turn_prompts") or []
files = record.get("files") or []
if not prompts or not files:
self._stats["invalid_records"] += 1
return []
outputs: dict[int, str] = {}
final_image: str | None = None
for entry in files:
file_id = entry.get("id")
url = entry.get("url")
if not file_id or not url:
continue
if file_id.startswith("edit_turn"):
try:
idx = int(file_id.replace("edit_turn", "").strip())
except ValueError:
continue
outputs[idx] = url
elif file_id == "final_image":
final_image = url
current_input = self._resolve_path(record.get("local_input_image"), base_dir=base_dir)
if not current_input:
return []
samples: list[dict[str, str]] = []
for turn_idx, prompt in enumerate(prompts, start=1):
target_rel = outputs.get(turn_idx)
if target_rel is None:
if turn_idx == len(prompts):
target_rel = final_image
else:
break
target_path = self._resolve_path(target_rel, base_dir=base_dir)
if not target_path:
break
sample = self._build_sample(prompt, current_input, target_path)
if not sample:
break
samples.append(sample)
current_input = target_path
return samples
def _select_prompt(self, record: dict) -> str | None:
if self.prefer_summarized_text and record.get("summarized_text"):
return record.get("summarized_text")
if record.get("text"):
return record.get("text")
return record.get("metadata_edit_turn_prompt")
def _build_sample(self, prompt: str | None, input_path: str | None, target_path: str | None) -> dict[str, str] | None:
if not prompt:
self._stats["invalid_records"] += 1
return None
if not input_path or not target_path:
return None
return {
"prompt": str(prompt).strip(),
"input_path": input_path,
"target_path": target_path,
}
def _resolve_path(self, path: str | None, *, base_dir: str | None = None) -> str | None:
if not path or path.startswith("http://") or path.startswith("https://"):
return None
candidates: list[str] = []
normalized = path.replace("\\", "/")
if os.path.isabs(normalized):
candidates.append(os.path.normpath(normalized))
else:
if self.image_root:
candidates.append(os.path.normpath(os.path.join(self.image_root, normalized)))
if base_dir:
candidates.append(os.path.normpath(os.path.join(base_dir, normalized)))
for candidate in candidates:
if not self.skip_missing or os.path.exists(candidate):
return candidate
self._stats["missing_paths"] += 1
return None
def _coerce_paths(self, value: Union[str, Sequence[str], None]) -> list[str]:
if value is None:
return []
if isinstance(value, str):
values = [value]
else:
values = [item for item in value if item]
return [os.path.abspath(os.path.expanduser(path)) for path in values]
@staticmethod
def _coerce_positive_int(value: Any) -> int | None:
if value is None:
return None
try:
int_value = int(value)
except (TypeError, ValueError):
return None
return int_value if int_value > 0 else None
# import os, socket
# from typing import Optional
# def _dist_identity():
# """Return a dict with rank info from env/torch if available."""
# info = {}
# # Env fallbacks for different launchers (torchrun/SLURM/MPI)
# def _get(*keys) -> Optional[int]:
# for k in keys:
# v = os.environ.get(k)
# if v is not None:
# try:
# return int(v)
# except ValueError:
# return None
# return None
# info["rank"] = _get("RANK", "SLURM_PROCID", "OMPI_COMM_WORLD_RANK")
# info["local_rank"] = _get("LOCAL_RANK", "SLURM_LOCALID", "MPI_LOCALRANKID")
# info["node_rank"] = _get("NODE_RANK", "SLURM_NODEID")
# info["world_size"] = _get("WORLD_SIZE", "SLURM_NTASKS", "OMPI_COMM_WORLD_SIZE")
# info["hostname"] = socket.gethostname()
# info["pid"] = os.getpid()
# # Optional: torch.distributed status
# try:
# import torch.distributed as dist
# info["dist_initialized"] = dist.is_available() and dist.is_initialized()
# if info["dist_initialized"]:
# info["rank"] = dist.get_rank()
# info["world_size"] = dist.get_world_size()
# info["backend"] = dist.get_backend()
# except Exception:
# info["dist_initialized"] = False
# # Optional: DataLoader worker ID
# try:
# from torch.utils.data import get_worker_info
# wi = get_worker_info()
# info["worker_id"] = wi.id if wi is not None else None
# except Exception:
# info["worker_id"] = None
# return info
# def _whoami_str():
# i = _dist_identity()
# return (
# f"[PROC] rank={i['rank']} local_rank={i['local_rank']} node_rank={i['node_rank']} "
# f"world={i['world_size']} worker={i['worker_id']} "
# f"host={i['hostname']} pid={i['pid']} "
# f"{'(backend='+i['backend']+')' if i.get('backend') else ''}"
# )
if __name__ == '__main__':
pass