micbench_xcomposer.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
ckpt_path = "internlm/internlm-xcomposer2-vl-7b"
tokenizer = AutoTokenizer.from_pretrained(ckpt_path, trust_remote_code=True)
# `torch_dtype=torch.float16` 可以令模型以 float16 精度加载，否则 transformers 会将模型加载为 float32，导致显存不足
model = AutoModelForCausalLM.from_pretrained(ckpt_path, torch_dtype=torch.float16, trust_remote_code=True, device_map={"":"cuda:1"})

model.max_length = 5120

image_folder = "/home/ps/Downloads/datasets/micbench/micbench_imgs/"
questions_file = "/home/ps/Downloads/datasets/micbench/micbench_{}.json"
answers_file = "/home/ps/Downloads/InternLM-XComposer/q-bench2-xcomposer2-micbench-{}.jsonl"

split = "test"

import json
from ast import literal_eval
from PIL import Image
import numpy as np
from tqdm import tqdm


with open(questions_file.format(split)) as f:
    if questions_file.endswith("json"):
        llvqa_data = json.load(f)
    else:
        llvqa_data = [literal_eval(line) for line in f.readlines()]
    print(llvqa_data[0])
    
    model = model.eval() 

correct_ = np.zeros((3,4))
all_ = np.zeros((3,4))

correct= 0

pbar = tqdm(total=len(llvqa_data))

for i, llddata in enumerate(llvqa_data):
    img_path_list = llddata["img_path"]
    if len(img_path_list) > 3:
        pbar.update(1)
        continue
    images = []
    for img_path in img_path_list:
        image = Image.open(image_folder+img_path).convert("RGB")
        image = model.vis_processor(image)
        images.append(image)
    image = torch.stack(images)
    options_prompt = ''
    for choice, ans in zip(["A.", "B.", "C.", "D."], llddata["candidates"]):
        options_prompt += f"{choice} {ans}\n"
        if ans == llddata["correct_ans"]:
            llddata["correct_choice"] = choice[0]
    query =  '<ImageHere> '*len(img_path_list) + 'Please answer this question by choosing the correct choice.' 
    query +=  "Context: N/A\nQuestion: " + llddata["question"] + '\nOptions: ' + options_prompt
    print(img_path_list, query)
    with torch.no_grad():
        with torch.cuda.amp.autocast():
            response, history = model.chat(tokenizer, query=query, image=image, history=[], do_sample=False, max_new_tokens=6)
    outputs = response.replace("The answer is ", "")
    if llddata["correct_choice"] in outputs: 
        correct += 1
        #correct_[llddata["type"]][llddata["concern"]]+= 1
    #all_[llddata["type"]][llddata["concern"]]+= 1
    
    with open(answers_file.format(split), "a") as wf:
        json.dump(llddata, wf)
    
    pbar.update(1)
    pbar.set_description("[Running Accuracy]: {:.4f},[Response]: {}, [Correct Ans]: {}, , [Prog]: {}".format(correct/(i+1), outputs, llddata.get("correct_choice", -1), i+1))