File size: 9,631 Bytes
9ed01de | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 | import copy
import json
import os
from dataclasses import dataclass
from typing import Dict, Sequence
import torch
import transformers
from PIL import Image
from torch.utils.data import Dataset
from src.constants import IGNORE_INDEX
from .utils import (expand2square, load_video, preprocess,
preprocess_multimodal, rank0_print)
class SingleDataset(Dataset):
"""Dataset for supervised fine-tuning."""
def __init__(
self,
data_paths: str,
data_weights: str,
tokenizer: transformers.PreTrainedTokenizer,
data_args,
):
super(SingleDataset, self).__init__()
list_data_dict = []
for data_path, data_weight in zip(data_paths, data_weights):
data_dict = json.load(open(data_path, "r"))
list_data_dict += data_dict * data_weight
rank0_print("Formatting inputs...Skip in lazy mode")
self.tokenizer = tokenizer
self.list_data_dict = list_data_dict
self.data_args = data_args
def __len__(self):
return len(self.list_data_dict)
@property
def lengths(self):
length_list = []
for sample in self.list_data_dict:
img_tokens = 128 if "image" in sample else 0
length_list.append(
sum(len(conv["value"].split()) for conv in sample["conversations"])
+ img_tokens
)
return length_list
@property
def modality_lengths(self):
length_list = []
for sample in self.list_data_dict:
cur_len = sum(
len(conv["value"].split()) for conv in sample["conversations"]
)
cur_len = cur_len if "image" in sample else -cur_len
length_list.append(cur_len)
return length_list
def next_rand(self):
import random
return random.randint(0, len(self) - 1)
def __getitem__(self, i) -> Dict[str, torch.Tensor]:
while True:
try:
sources = self.list_data_dict[i]
if isinstance(i, int):
sources = [sources]
sources_org = copy.deepcopy(sources)
assert (
len(sources) == 1
), "Don't know why it is wrapped to a list" # FIXME
if "image" in sources_org[0]:
image_file = sources_org[0]["image"]
image_folder = self.data_args.image_folder
processor = self.data_args.image_processor
from pathlib import Path
# if not Path(os.path.join(image_folder, image_file)).exists():
# i = self.next_rand()
# continue
if isinstance(image_file, list):
# Multiple Images as Input
try:
image = [
Image.open(os.path.join(image_folder, imfile)).convert(
"RGB"
)
for imfile in image_file
]
except Exception as ex:
print(ex)
i = self.next_rand()
continue
if self.data_args.image_aspect_ratio == "pad":
image = [
expand2square(
img,
tuple(int(x * 255) for x in processor.image_mean),
)
for img in image
]
image = processor.preprocess(image, return_tensors="pt")[
"pixel_values"
]
else:
image = processor.preprocess(image, return_tensors="pt")[
"pixel_values"
]
elif os.path.join(image_folder, image_file).endswith("mp4"):
# Video as Input
image = load_video(os.path.join(image_folder, image_file))
if self.data_args.image_aspect_ratio == "pad":
image = [
expand2square(
img,
tuple(int(x * 255) for x in processor.image_mean),
)
for img in image
]
image = processor.preprocess(image, return_tensors="pt")[
"pixel_values"
]
else:
image = processor.preprocess(image, return_tensors="pt")[
"pixel_values"
]
else:
try:
image = Image.open(
os.path.join(image_folder, image_file)
).convert("RGB")
except Exception as ex:
print(ex)
i = self.next_rand()
continue
if self.data_args.image_aspect_ratio == "pad":
image = expand2square(
image, tuple(int(x * 255) for x in processor.image_mean)
)
image = processor.preprocess(image, return_tensors="pt")[
"pixel_values"
]
else:
image = processor.preprocess(image, return_tensors="pt")[
"pixel_values"
]
sources = preprocess_multimodal(
copy.deepcopy([e["conversations"] for e in sources]),
self.data_args,
)
else:
sources = copy.deepcopy([e["conversations"] for e in sources])
data_dict = preprocess(
sources,
self.tokenizer,
has_image=("image" in sources_org[0]),
)
if isinstance(i, int):
data_dict = dict(
input_ids=data_dict["input_ids"][0],
labels=data_dict["labels"][0],
)
# default task_type: "score", level_probs: [-10000] * 5
data_dict["task_type"] = sources_org[0].get("task_type", "score")
data_dict["level_probs"] = sources_org[0].get("level_probs", [-10000] * 5)
# image exist in the data
if "image" in sources_org[0]:
data_dict["image"] = image
elif self.data_args.is_multimodal:
# image does not exist in the data, but the model is multimodal
crop_size = self.data_args.image_processor.crop_size
data_dict["image"] = torch.zeros(
3, crop_size["height"], crop_size["width"]
)
return data_dict
except Exception as ex:
print(ex)
i = self.next_rand()
continue
@dataclass
class DataCollatorForSupervisedDataset(object):
"""Collate examples for supervised fine-tuning."""
tokenizer: transformers.PreTrainedTokenizer
def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
input_ids, labels = tuple(
[instance[key] for instance in instances] for key in ("input_ids", "labels")
)
input_ids = torch.nn.utils.rnn.pad_sequence(
input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id
)
labels = torch.nn.utils.rnn.pad_sequence(
labels, batch_first=True, padding_value=IGNORE_INDEX
)
input_ids = input_ids[:, : self.tokenizer.model_max_length]
labels = labels[:, : self.tokenizer.model_max_length]
batch = dict(
input_type="single",
input_ids=input_ids,
labels=labels,
attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
)
batch["task_types"] = [instance["task_type"] for instance in instances]
batch["level_probs"] = torch.tensor([instance["level_probs"] for instance in instances])
if "image" in instances[0]:
images = [instance["image"] for instance in instances]
if all(x is not None and x.shape == images[0].shape for x in images):
batch["images"] = torch.stack(images)
else:
batch["images"] = images
return batch
def make_single_data_module(
tokenizer: transformers.PreTrainedTokenizer, data_args
) -> Dict:
"""Make dataset and collator for supervised fine-tuning."""
train_dataset = SingleDataset(
tokenizer=tokenizer,
data_paths=data_args.data_paths,
data_weights=data_args.data_weights,
data_args=data_args,
)
data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
return dict(
train_dataset=train_dataset, eval_dataset=None, data_collator=data_collator
)
|