generate_prompt.py

import os

import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import json
from tqdm import tqdm
import os
import argparse

def read_json(file_path): 
    with open(file_path, 'r', encoding='utf-8') as file:
        data = json.load(file)
    return data

def write_json(file_path, data):
    with open(file_path, 'w', encoding='utf-8') as file:
        json.dump(data, file, ensure_ascii=False, indent=4)

# default: Load the model on the available device(s)
print(torch.cuda.device_count())
#model_path = "/home/zbz5349/WorkSpace/aigeeks/Qwen2.5-VL/ckpt_7B"
# model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
#     model_path, torch_dtype="auto", device_map="auto"
# )


# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
parser = argparse.ArgumentParser()
parser.add_argument("--model_path", type=str, default="/home/zbz5349/WorkSpace/aigeeks/Qwen2.5-VL/ckpt_7B")
parser.add_argument("--begin", type=int, default=0)
parser.add_argument("--end", type=int, default=4635)
parser.add_argument("--batch_size", type=int, default=16)
parser.add_argument("--data_path", type=str, default="/home/zbz5349/WorkSpace/aigeeks/Qwen2.5-VL/magicbrush_dataset/dataset.json")
parser.add_argument("--prompt_path", type=str, default="/home/zbz5349/WorkSpace/aigeeks/Qwen2.5-VL/magicbrush_dataset/gen.json")

args = parser.parse_args()
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    args.model_path,
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

# default processor
processor = AutoProcessor.from_pretrained(args.model_path)
print(model.device)




data = read_json(args.data_path)
save_data = []
correct_num = 0
begin = args.begin
end = args.end
batch_size = args.batch_size
json_path = args.prompt_path


for batch_idx in tqdm(range(begin, end, batch_size)):
    batch = data[batch_idx:min(batch_idx+batch_size, end)] 
    print(len(batch))
    image_list = []
    input_text_list = []
    data_list = []
    save_list = []
    sd_ans = []
    # while True:
    for idx, i in enumerate(batch):
        save_ =  {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": "",
                },
            {"type": "text",
             "text": "Please help me write a prompt for image editing on this picture. The requirements are as follows: complex editing instructions should include two to five simple editing instructions involving spatial relationships (simple editing instructions such as ADD: add an object to the left of a certain object, DELETE: delete a certain object, MODIFY: change a certain object into another object). We hope that the editing instructions can have simple reasoning and can also include some abstract concept-based editing (such as making the atmosphere more romantic, or making the diet healthier, or making the boy more handsome and the girl more beautiful, etc.). Please give me clear editing instructions and also consider whether such editing instructions are reasonable."},
            ],
            "result":""
        }
        #idx_real = batch_idx * batch_size + idx
        messages = batch[idx]
        save_['content'][0]['image'] = messages['content'][0]['image']
        save_['content'][1]['text'] = messages['content'][1]['text']

        data_list.append(messages)
        save_list.append(save_)
    #print(len(data_list))
    text = processor.apply_chat_template(data_list, tokenize=False, add_generation_prompt=True)
    #print(len(text))
    image_inputs, video_inputs = process_vision_info(data_list)
    inputs = processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=True,
        return_tensors="pt",
    )
    inputs = inputs.to(model.device)

    # Inference: Generation of the output
    generated_ids = model.generate(**inputs, max_new_tokens=128)
    #print(generated_ids.shape)
    generated_ids_trimmed = [
        out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
    ]
    output_text = processor.batch_decode(
        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
    )
    #print(output_text)
    save_['result'] = output_text
    save_data.append(save_)

    if batch_idx % 4 ==0:
        write_json(json_path,save_data)
        print(len(save_data))

write_json(json_path,save_data)