部署效果跟huggingface上demo效果有差距 #81
Description
我尝试自己部署了showui-2b模型，但效果似乎不如huggingface上的demo演示效果，能否抽空解答我的疑问？感谢。
以下是测试结果（采用同样的提示词：”点击我的“）：
DEMO RESULT（结果正确）：
MY RESULT（结果错误）：
以下是我的推理代码：
import torch
from qwen_vl_utils import process_vision_info
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from PIL import Image, ImageDraw, ImageFont
from utils import draw_action_on_image_xy
import ast
import time
import json
class ShowUI:
def __init__(self):
model_path = "./weights/showui-2b"
start_time = time.time()
self.model = Qwen2VLForConditionalGeneration.from_pretrained(
model_path,
torch_dtype=torch.bfloat16,
device_map={"": 0}
)
self.min_pixels = 256*28*28
self.max_pixels = 1344*28*28
self.processor = AutoProcessor.from_pretrained(model_path, min_pixels=self.min_pixels, max_pixels=self.max_pixels)
load_time = time.time() - start_time
print('-------------------------')
print(f"model load time: {load_time:.2f} s")
self.nav_system = """You are an assistant trained to navigate the {_APP} screen.
Given a task instruction and a screen observation, output the executable action.
Here is the action space:
{_ACTION_SPACE}
"""
self.nav_format = """
Format the action as a dictionary with the following keys:
{"action": "ACTION_TYPE", "value": "element", "position": [x,y]}
If value or position is not applicable, set it as "null".
Position represents the relative coordinates on the screenshot and should be scaled to a range of 0-1.
"""
self.action_map = {
'web': """
1. `CLICK`: Click on an element, value is not applicable and the position [x,y] is required.
2. `INPUT`: Type a string into an element, value is a string to type and the position [x,y] is required.
3. `SELECT`: Select a value for an element, value is not applicable and the position [x,y] is required.
4. `HOVER`: Hover on an element, value is not applicable and the position [x,y] is required.
5. `ANSWER`: Answer the question, value is the answer and the position is not applicable.
6. `ENTER`: Enter operation, value and position are not applicable.
7. `SCROLL`: Scroll the screen, value is the direction to scroll and the position is not applicable.
8. `SELECT_TEXT`: Select some text content, value is not applicable and position [[x1,y1], [x2,y2]] is the start and end position of the select operation.
9. `COPY`: Copy the text, value is the text to copy and the position is not applicable.
10. 'IDENTIFY': Identify the text, value is the text to identify and the position [[x1,y1], [x2,y2]] is the start and end position of the element.
11. 'ASSERTION': Assert an element exists, value is not applicable and the position [x,y] is the coordinates of the element.
""",
'phone': """
1. `INPUT`: Type a string into an element, value is not applicable and the position [x,y] is required.
2. `SWIPE`: Swipe the screen, value is not applicable and the position [[x1,y1], [x2,y2]] is the start and end position of the swipe operation.
3. `TAP`: Tap on an element, value is not applicable and the position [x,y] is required.
4. `ANSWER`: Answer the question, value is the status (e.g., 'task complete') and the position is not applicable.
5. `ENTER`: Enter operation, value and position are not applicable.
6. 'ASSERTION': Assert an element exists, value is not applicable and the position [x,y] is the coordinates of the element.
"""
}
def infer(self, img_url, query, split):
system_prompt = self.nav_system.format(_APP=split, _ACTION_SPACE=self.action_map[split]) + self.nav_format
messages = [
{
"role": "user",
"content": [
{"type": "text", "text": system_prompt},
{"type": "text", "text": f'Task: {query}'},
{"type": "image", "image": img_url, "min_pixels": self.min_pixels, "max_pixels": self.max_pixels},
],
}
]
text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = self.processor(text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt")
inputs = inputs.to("cuda")
generated_ids = self.model.generate(**inputs, max_new_tokens=512)
generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
outputs = self.processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
return outputs
def cleanup(self):
torch.cuda.empty_cache()``