Giters

bensonbs / GLIP

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

GLIP 實戰

GLIP 簡介

GLIP: Grounded Language-Image Pre-training 一種用於學習對象級、語言感知和語義豐富的視覺表示的基礎語言圖像預訓練 (GLIP) 模型。

統一了目標檢測和短語**基礎以進行預訓練。

  • 模型使用

    範例1

    想要框選旁邊正在加藥的桶子輸入 bucket 水桶 ，但使用模型只認得小的水桶

    利用用語義說明 blue rags on big bucket 藍色抹布在大水桶 上成功框選到加藥桶

    範例2

    John wearing white suit and Ben wearing black clothes 模型沒有訓練過防護衣，但是可以使用white suit 白色套裝也能成功框選，且模型能夠根據語意推斷哪個是John 哪個是 Ben。

  • Docker 建置環境

  • 範例程式

    • 打開window終端機輸入
    • docker run -it -v C:\Users\ZME\AI:/workspace -p 18888:8888 -p 15000:5000 --name GLIP_gpu --gpus all pengchuanzhang/pytorch:ubuntu20.04_torch1.9-cuda11.3-nccl2.9.9

  • 參數詳解

  • -it 是 --interactive +的縮寫 --tty 。當您 docker run 使用此命令時，它會帶您直接進入容器

  • -v 共用資料夾 主機共用路徑 : docker內共路徑

  • -p 是虛擬機內外port 轉發 主機 port : docker port

  • --name Containers 名稱

  • --gpus all docker 使用所有 gpu

  • 依照Cuda版本選擇映像檔

    • Cuda 11.3 映像檔
      • pengchuanzhang/pytorch:ubuntu20.04_torch1.9-cuda11.3-nccl2.9.9
    • Cuda10.2 映像檔
      • pengchuanzhang/maskrcnn:ubuntu18-py3.7-cuda10.2-pytorch1.9

  • Docker 環境啟動

1. 選擇 Containers

2. 打開終端機 OPEN IN TERMINAL

3. 輸入jupyter lab

4. 複製 token

- *http://hostname:**8888**/token=`b2c967cb247f79da97803bd7b66d52ca527cf78b1c52974e``*

5. 瀏覽器輸入`

- http://localhost:18888/ 進入 jupyter lab

6. 貼上 token 設定 jupyter lab 密碼

建立Docker內程式環境

1. 打開Jupyter Lab 終端機

2. 輸入 git clone https://github.com/microsoft/GLIP 下載GLIP github

3. 進入 GLIP 資料夾 cd GLIP

4. 創建 Demo.ipynb

5. 安裝所需套件

  ! pip install einops shapely timm yacs tensorboardX ftfy prettytable pymongo
  ! pip install transformers 
  ! pip install ipywidgets
  ! python setup.py build develop --user

6. import 所需套件

  import matplotlib.pyplot as plt
  import matplotlib.pylab as pylab
  
  import requests
  from io import BytesIO
  from PIL import Image
  import numpy as np
  pylab.rcParams['figure.figsize'] = 20, 12
  from maskrcnn_benchmark.config import cfg
  from maskrcnn_benchmark.engine.predictor_glip import GLIPDemo
  
  def load(url):
      """
      Given an url of an image, downloads the image and
      returns a PIL image
      """
      response = requests.get(url)
      pil_image = Image.open(BytesIO(response.content)).convert("RGB")
      # convert to BGR format
      image = np.array(pil_image)[:, :, [2, 1, 0]]
      return image
  
  def imshow(img, caption):
    plt.imshow(img[:, :, [2, 1, 0]])
    plt.axis("off")
    plt.figtext(0.5, 0.09, caption, wrap=True, horizontalalignment='center', fontsize=20)

7. 下載模型權重

Use this command for evaluate the GLPT-T model weight_file 檔案較大提供GD 載點 將glip_tiny_model_o365_goldg_cc_sbu.pth 放在 MODEL 目錄底下

  # ! wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/models/glip_tiny_model_o365_goldg_cc_sbu.pth -O MODEL/glip_tiny_model_o365_goldg_cc_sbu.pth
  config_file = "configs/pretrain/glip_Swin_T_O365_GoldG.yaml"
  weight_file = "MODEL/glip_tiny_model_o365_goldg_cc_sbu.pth"

update the config options with the config file manual override some options

  cfg.local_rank = 0
  cfg.num_gpus = 1
  cfg.merge_from_file(config_file)
  cfg.merge_from_list(["MODEL.WEIGHT", weight_file])
  cfg.merge_from_list(["MODEL.DEVICE", "cuda"])

(第一次執行要下載資料比較久，notebook進度條有bug看不到會比較空虛)

  glip_demo = GLIPDemo(
      cfg,
      min_image_size=800,
      confidence_threshold=0.7,
      show_mask_heatmaps=False
  )

可以改使用.py執行可正常使用進度條 創建download.py

  from maskrcnn_benchmark.config import cfg
  from maskrcnn_benchmark.engine.predictor_glip import GLIPDemo
  
  
  if __name__ == '__main__':
  
      # Use this command for evaluate the GLPT-T model
      # ! wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/models/glip_tiny_model_o365_goldg_cc_sbu.pth -O MODEL/glip_tiny_model_o365_goldg_cc_sbu.pth
      config_file = "configs/pretrain/glip_Swin_T_O365_GoldG.yaml"
      weight_file = "MODEL/glip_tiny_model_o365_goldg_cc_sbu.pth"
  
      # Use this command to evaluate the GLPT-L model
      # ! wget https://penzhanwu2bbs.blob.core.windows.net/data/GLIPv1_Open/models/glip_large_model.pth -O MODEL/glip_large_model.pth
      # config_file = "configs/pretrain/glip_Swin_L.yaml"
      # weight_file = "MODEL/glip_large_model.pth"
  
      # update the config options with the config file
      # manual override some options
      cfg.local_rank = 0
      cfg.num_gpus = 1
      cfg.merge_from_file(config_file)
      cfg.merge_from_list(["MODEL.WEIGHT", weight_file])
      cfg.merge_from_list(["MODEL.DEVICE", "cuda"])
      glip_demo = GLIPDemo(
          cfg,
          min_image_size=800,
          confidence_threshold=0.7,
          show_mask_heatmaps=False
      )

Bug修正

GLIPDemo class 在執行時會未定義 self.color 取代為 255 不影響使用

    glip_demo.color=255

讀取圖像方式

- Internet 讀取圖像
	- `image = load('http://farm4.staticflickr.com/3693/9472793441_b7822c00de_z.jpg')`
- PIL Imgae
	- `image = np.array(Image.open('./4.png'))[:,:,::-1]`
	- 1. 讀取圖片
	- 2. PIL格式轉為np.array
	- 3. `[:,:,::-1]` RGB -> BGR (PIL圖片顯示為RGB cv2為BGR)
- open-cv
	- `cv2.imread('./9.png')`

成果展示 1

- ### 模型檢測
	- 檢測 **a man,the cone,  hard hat **
	- 檢測 人、三角錐、安全帽

  image = np.array(Image.open('./4.png'))[:,:,::-1]
  caption = 'a man . the cone . hard hat' # the caption can also be the simple concatonation of some random categories.
  result, _ = glip_demo.run_on_web_image(image, caption, 0.5)
  imshow(result, caption)

成果展示 2

- 利用語意說明  **a man wearing hard hat and cone next to him**
- 有個人戴著安全帽，三角錐在他身旁

  image = np.array(Image.open('./4.png'))[:,:,::-1]
  caption = 'a man wearing hard hat and cone next to him' # the caption can also be the simple concatonation of some random categories.
  result, _ = glip_demo.run_on_web_image(image, caption, 0.5)
  imshow(result, caption)

輸出Pytorch Yolov5 標註檔

將cv2 左上 -> 右下 框選方式改為 yolov5 格式 中心點 長寬

```
  def xyxy2xywhn(x, w=640, h=640, clip=False, eps=0.0):
      # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] normalized where xy1=top-left, xy2=bottom-right
      if clip:
	  clip_coords(x, (h - eps, w - eps))  # warning: inplace clip
      y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
      y[:, 0] = ((x[:, 0] + x[:, 2]) / 2) / w  # x center
      y[:, 1] = ((x[:, 1] + x[:, 3]) / 2) / h  # y center
      y[:, 2] = (x[:, 2] - x[:, 0]) / w  # width
      y[:, 3] = (x[:, 3] - x[:, 1]) / h  # height
      return y
  ```

模型predictions 篩選閥值，取得標籤，寫出.txt輸出位置與影像輸入相同

```
  def make_label(img_path,caption):
      image = cv2.imread(img_path)
      txt_path = os.path.splitext(img_path)[0]
      predictions = glip_demo.compute_prediction(image, caption)
      glip_demo.confidence_threshold = 0.5
      top_predictions = glip_demo._post_process_fixed_thresh(predictions)
      boxs = top_predictions.bbox
      index = top_predictions.get_field("labels")
      labels = [glip_demo.entities[i - 1]for i in index]
      h,w,_ = image.shape
      xywhs = xyxy2xywhn(x=boxs,w=w,h=h)
      
      with open(f'{txt_path}.txt', 'a') as f:
	  for cls,xywh in zip(index,xywhs):
	      line = (cls-1, *xywh)
	      f.write(('%g ' * len(line)).rstrip() % line + '\n')
	      
      with open(os.path.join(''.join(os.path.split(txt_path)[:-1]),'labels.txt'), 'w') as f:
	  for c in glip_demo.entities:
	      f.write(c+ '\n')
  ```

測試 Function

```
  img_path = './pic/4.png'
  caption = 'a man wearing hard hat and cone next to him' 
  make_label(image,caption)
  ```

輸出結果 ./pic/4.txt

```
  0 0.770998 0.547756 0.072838 0.212098
  1 0.779665 0.513667 0.0603401 0.136979
  2 0.936433 0.467944 0.034296 0.0650564
  4 0.368129 0.365073 0.0397727 0.0945886
  0 0.777506 0.514885 0.0635377 0.141048
  
  ```

API 架設

程式測試完成利用 (本機)<-->(Docker) 的API，傳輸 (圖片+語意)<-->(類別+座標)

  • 本機將圖片轉為base64字串+語意透過API傳輸
  • 透過docker內的FastAPI回傳Json格式類別+座標

本機設置

在建置container預留了 -p 15000:5000 給API連接

  • 設置API PORT 連至 15000 與 docker 內 5000 連接
  • base64_str 為 base64格式的圖片，型態為字串
  • caption 需要GLIP辨識的語意
import base64
from PIL import Image
from io import BytesIO
import requests,json

def api(img_base64):
    r = requests.post('http://127.0.0.1:15000/upload', 
    data = {
	'base64_str':f'{img_base64}',
	'caption': 'a man hard hat'
    })

    return json.loads(r.text)
  • 將PIL圖片轉為base64連同語意一起發送到API
    caption = 'cone . a man'
    image = Image.open(r'temp\outputs\3703(G2A_L10_1_D)GED1_2022_11_07_7_57_59_0000.png')
    buffered = BytesIO()
    image.save(buffered, format="JPEG")
    img_base64 = base64.b64encode(buffered.getvalue())
    api(str(img_base64)[2:-1])

Docker設置

pip install python-multipart
pip install fastapi

新增main.py

import re
import os
import torch
import base64
import uvicorn
import numpy as np

from io import BytesIO
from PIL import Image
from typing import Union
from fastapi import FastAPI, File, Form
from pydantic import BaseModel

from maskrcnn_benchmark.config import cfg
from maskrcnn_benchmark.engine.predictor_glip import GLIPDemo

將收到的 base64 圖片轉為 PIL 格式

def base64_to_image(base64_str, image_path=None):
    base64_data = re.sub('^data:image/.+;base64,', '', base64_str)
    byte_data = base64.b64decode(base64_data)
    image_data = BytesIO(byte_data)
    img = Image.open(image_data)
    if image_path:
	img.save(image_path)
    return img

坐標系轉換

def xyxy2xywhn(x, w=640, h=640, clip=False, eps=0.0):
# Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] normalized where xy1=top-left, xy2=bottom-right
if clip:
    clip_coords(x, (h - eps, w - eps))  # warning: inplace clip
y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
y[:, 0] = ((x[:, 0] + x[:, 2]) / 2) / w  # x center
y[:, 1] = ((x[:, 1] + x[:, 3]) / 2) / h  # y center
y[:, 2] = (x[:, 2] - x[:, 0]) / w  # width
y[:, 3] = (x[:, 3] - x[:, 1]) / h  # height
return y

  • 預測結果打包成Json格式回傳

    • {
      "0": { 圖片內第幾個ROI
      "index": 0, 物體類別的index
      "label": "a man", 物體類別
      "x": 0.7348242402076721, ROI中心點x座標/圖片寬度
      "y": 0.7646925449371338, ROI中心點y座標/圖片高度
      "w": 0.16020514070987701, ROI寬度/圖片寬度
      "h": 0.2502952814102173 ROI高度/圖片高度
      },
      "1": {
      "index": 0,
      "label": "a man",
      "x": 0.7718421816825867,
      "y": 0.7066231966018677,
      "w": 0.04969634860754013,
      "h": 0.08035018295049667
      }
      }
    def predict2json(image,caption):
    image = np.array(image)[:,:,::-1]
    predictions = glip_demo.compute_prediction(image, caption)
    glip_demo.confidence_threshold = 0.5
    top_predictions = glip_demo._post_process_fixed_thresh(predictions)
    boxs = top_predictions.bbox
    index = top_predictions.get_field("labels")
    h,w,_ = image.shape
    xywhs = xyxy2xywhn(x=boxs,w=w,h=h)

    res = {}
    for c, (i,loc) in enumerate(zip(index,xywhs)):
        x,y,w,h = loc
        res[c] = {}
        res[c]['index'] = int(i) -1
        res[c]['label'] = glip_demo.entities[int(i) -1]
        res[c]['x'] = float(x)
        res[c]['y'] = float(y)
        res[c]['w'] = float(w)
        res[c]['h'] = float(h)
    return res

  • 啟動GLIP model

    config_file = "configs/pretrain/glip_Swin_T_O365_GoldG.yaml"
weight_file = "MODEL/glip_tiny_model_o365_goldg_cc_sbu.pth"

cfg.local_rank = 0
cfg.num_gpus = 1
cfg.merge_from_file(config_file)
cfg.merge_from_list(["MODEL.WEIGHT", weight_file])
cfg.merge_from_list(["MODEL.DEVICE", "cuda"])

glip_demo = GLIPDemo(
    cfg,
    min_image_size=800,
    confidence_threshold=0.5,
    show_mask_heatmaps=False
)

  • 架設FastAPI

    • uvicorn.run(app, host="0.0.0.0", port=5000) 可自由設定port
        app = FastAPI()


    class Item(BaseModel):
        name: str
        price: float
        is_offer: Union[bool, None] = None


    @app.get("/")
    def read_root():
        return {"Hello": "World"}


    @app.post("/upload")
    def upload(base64_str: str = Form(...), caption: str = Form(...)):
        try:
            image = base64_to_image(base64_str)
            res = predict2json(image,caption)
        except Exception as e:
            return {"message": f"{e}"}

        return res

    if __name__ == "__main__":
        uvicorn.run(app, host="0.0.0.0", port=5000)

About

License:MIT License

Languages

Language:Python 90.4%Language:Cuda 7.5%Language:C++ 1.5%Language:C 0.6%Language:Shell 0.0%