202 lines
7.3 KiB
Python
202 lines
7.3 KiB
Python
|
import torch
|
|||
|
|
import torchvision.transforms as T
|
|||
|
|
from PIL import Image
|
|||
|
|
from torchvision.transforms.functional import InterpolationMode
|
|||
|
|
import requests
|
|||
|
|
from io import BytesIO
|
|||
|
|
from urllib.parse import urlparse
|
|||
|
|
|
|||
|
|
IMAGENET_MEAN = (0.485, 0.456, 0.406)
|
|||
|
|
IMAGENET_STD = (0.229, 0.224, 0.225)
|
|||
|
|
|
|||
|
|
def build_transform(input_size):
|
|||
|
|
transform = T.Compose([
|
|||
|
|
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
|
|||
|
|
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
|
|||
|
|
T.ToTensor(),
|
|||
|
|
T.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD)
|
|||
|
|
])
|
|||
|
|
return transform
|
|||
|
|
|
|||
|
|
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
|
|||
|
|
best_ratio_diff = float('inf')
|
|||
|
|
best_ratio = (1, 1)
|
|||
|
|
area = width * height
|
|||
|
|
for ratio in target_ratios:
|
|||
|
|
target_aspect_ratio = ratio[0] / ratio[1]
|
|||
|
|
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
|
|||
|
|
if ratio_diff < best_ratio_diff:
|
|||
|
|
best_ratio_diff = ratio_diff
|
|||
|
|
best_ratio = ratio
|
|||
|
|
elif ratio_diff == best_ratio_diff:
|
|||
|
|
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
|
|||
|
|
best_ratio = ratio
|
|||
|
|
return best_ratio
|
|||
|
|
|
|||
|
|
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
|
|||
|
|
orig_width, orig_height = image.size
|
|||
|
|
aspect_ratio = orig_width / orig_height
|
|||
|
|
|
|||
|
|
target_ratios = set(
|
|||
|
|
(i, j) for n in range(min_num, max_num + 1)
|
|||
|
|
for i in range(1, n + 1)
|
|||
|
|
for j in range(1, n + 1)
|
|||
|
|
if i * j <= max_num and i * j >= min_num
|
|||
|
|
)
|
|||
|
|
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
|
|||
|
|
|
|||
|
|
target_aspect_ratio = find_closest_aspect_ratio(
|
|||
|
|
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
|
|||
|
|
|
|||
|
|
target_width = image_size * target_aspect_ratio[0]
|
|||
|
|
target_height = image_size * target_aspect_ratio[1]
|
|||
|
|
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
|
|||
|
|
|
|||
|
|
resized_img = image.resize((target_width, target_height))
|
|||
|
|
processed_images = []
|
|||
|
|
for i in range(blocks):
|
|||
|
|
box = (
|
|||
|
|
(i % (target_width // image_size)) * image_size,
|
|||
|
|
(i // (target_width // image_size)) * image_size,
|
|||
|
|
((i % (target_width // image_size)) + 1) * image_size,
|
|||
|
|
((i // (target_width // image_size)) + 1) * image_size
|
|||
|
|
)
|
|||
|
|
split_img = resized_img.crop(box)
|
|||
|
|
processed_images.append(split_img)
|
|||
|
|
|
|||
|
|
assert len(processed_images) == blocks
|
|||
|
|
if use_thumbnail and len(processed_images) != 1:
|
|||
|
|
thumbnail_img = image.resize((image_size, image_size))
|
|||
|
|
processed_images.append(thumbnail_img)
|
|||
|
|
return processed_images
|
|||
|
|
|
|||
|
|
def load_image(image_file, input_size=448, max_num=12):
|
|||
|
|
# 处理不同类型的输入
|
|||
|
|
if isinstance(image_file, Image.Image):
|
|||
|
|
image = image_file.convert('RGB')
|
|||
|
|
elif isinstance(image_file, str) and bool(urlparse(image_file).netloc):
|
|||
|
|
try:
|
|||
|
|
response = requests.get(image_file, timeout=10)
|
|||
|
|
response.raise_for_status()
|
|||
|
|
image = Image.open(BytesIO(response.content)).convert('RGB')
|
|||
|
|
except Exception as e:
|
|||
|
|
raise ValueError(f"无法从URL加载图片: {str(e)}")
|
|||
|
|
elif isinstance(image_file, str):
|
|||
|
|
image = Image.open(image_file).convert('RGB')
|
|||
|
|
elif isinstance(image_file, bytes):
|
|||
|
|
image = Image.open(BytesIO(image_file)).convert('RGB')
|
|||
|
|
else:
|
|||
|
|
raise ValueError(f"不支持的图片格式: {type(image_file)}")
|
|||
|
|
|
|||
|
|
transform = build_transform(input_size=input_size)
|
|||
|
|
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
|||
|
|
pixel_values = [transform(image) for image in images]
|
|||
|
|
pixel_values = torch.stack(pixel_values)
|
|||
|
|
return pixel_values
|
|||
|
|
|
|||
|
|
def prepare_training_data(question: str, answer: str, tokenizer, image_path=None, input_size=448, max_num=12, num_image_token: int = 256):
|
|||
|
|
"""
|
|||
|
|
准备完整的训练数据（包括输入和标签）
|
|||
|
|
|
|||
|
|
Args:
|
|||
|
|
question (str): 用户输入的问题
|
|||
|
|
answer (str): 助手的回答文本
|
|||
|
|
tokenizer: 分词器
|
|||
|
|
image_path: 图片路径、URL或PIL Image对象
|
|||
|
|
input_size (int): 图片输入尺寸
|
|||
|
|
max_num (int): 最大图片块数
|
|||
|
|
num_image_token (int): 每张图片的token数量
|
|||
|
|
|
|||
|
|
Returns:
|
|||
|
|
dict: 包含模型训练所需的所有输入
|
|||
|
|
"""
|
|||
|
|
# 1. 处理图像输入
|
|||
|
|
pixel_values = None
|
|||
|
|
if image_path is not None:
|
|||
|
|
pixel_values = load_image(image_path, input_size=input_size, max_num=max_num)[-1:] ### 只取最后一张图片
|
|||
|
|
if torch.cuda.is_available():
|
|||
|
|
pixel_values = pixel_values.to(torch.bfloat16)
|
|||
|
|
|
|||
|
|
# 2. 确保问题包含图片标记
|
|||
|
|
if pixel_values is not None and '<image>' not in question:
|
|||
|
|
question = '<image>\n' + question
|
|||
|
|
|
|||
|
|
# 3. 根据pixel_values确定num_patches
|
|||
|
|
num_patches = pixel_values.shape[0] if pixel_values is not None else 0
|
|||
|
|
|
|||
|
|
# 4. 构造完整的对话内容
|
|||
|
|
system_msg = "你是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。"
|
|||
|
|
|
|||
|
|
# 5. 替换图片标记
|
|||
|
|
if num_patches > 0:
|
|||
|
|
image_tokens = "<img>" + "<IMG_CONTEXT>" * (num_image_token * num_patches) + "</img>"
|
|||
|
|
question = question.replace('<image>', image_tokens, 1)
|
|||
|
|
|
|||
|
|
# 6. 构造完整的query（包含答案）
|
|||
|
|
full_prompt = (
|
|||
|
|
f"<|im_start|>system\n{system_msg}<|im_end|>\n"
|
|||
|
|
f"<|im_start|>user\n{question}<|im_end|>\n"
|
|||
|
|
f"<|im_start|>assistant\n{answer}<|im_end|>\n"
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 7. 转换为模型输入格式
|
|||
|
|
model_inputs = tokenizer(
|
|||
|
|
full_prompt,
|
|||
|
|
return_tensors="pt",
|
|||
|
|
add_special_tokens=False
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 8. 构造labels
|
|||
|
|
input_ids = model_inputs["input_ids"]
|
|||
|
|
labels = input_ids.clone()
|
|||
|
|
|
|||
|
|
assistant_start_token = "<|im_start|>assistant\n"
|
|||
|
|
assistant_token_ids = tokenizer(assistant_start_token, add_special_tokens=False)["input_ids"]
|
|||
|
|
|
|||
|
|
assistant_start_pos = None
|
|||
|
|
for i in range(len(input_ids[0]) - len(assistant_token_ids)):
|
|||
|
|
if input_ids[0][i:i+len(assistant_token_ids)].tolist() == assistant_token_ids:
|
|||
|
|
assistant_start_pos = i
|
|||
|
|
break
|
|||
|
|
|
|||
|
|
if assistant_start_pos is not None:
|
|||
|
|
labels[0, :assistant_start_pos] = -100
|
|||
|
|
|
|||
|
|
return {
|
|||
|
|
"input_ids": input_ids,
|
|||
|
|
"attention_mask": model_inputs["attention_mask"],
|
|||
|
|
"labels": labels,
|
|||
|
|
"pixel_values": pixel_values
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
# 使用示例：
|
|||
|
|
"""
|
|||
|
|
import torch
|
|||
|
|
from transformers import AutoTokenizer
|
|||
|
|
|
|||
|
|
# 初始化tokenizer
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained("path_to_tokenizer")
|
|||
|
|
|
|||
|
|
# 准备示例数据
|
|||
|
|
question = "请描述这张图片中的内容"
|
|||
|
|
answer = "这是一张美丽的风景照，画面中有青山绿水。"
|
|||
|
|
image_path = "./examples/image1.jpg" # 或URL，或PIL Image对象
|
|||
|
|
|
|||
|
|
# 准备训练数据
|
|||
|
|
training_data = prepare_training_data(
|
|||
|
|
question=question,
|
|||
|
|
answer=answer,
|
|||
|
|
tokenizer=tokenizer,
|
|||
|
|
image_path=image_path,
|
|||
|
|
input_size=448,
|
|||
|
|
max_num=12
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# training_data 包含:
|
|||
|
|
# {
|
|||
|
|
# "input_ids": tensor([[...]]), # 完整对话的token ids
|
|||
|
|
# "attention_mask": tensor([[...]]), # 注意力掩码
|
|||
|
|
# "labels": tensor([[...]]), # 带有-100标记的标签
|
|||
|
|
# "pixel_values": tensor([[...]]) # 图像数据
|
|||
|
|
# }
|
|||
|
|
"""
|