[object detection] 구글 코랩에서 실행하기

# 텐서플로우 설치

import tensorflow as tf

tf.__version__

# COCO API installation :  pycocotools  를 설치 => 이미 설치되어있음.

! pip install pycocotools

# 깃허브에 있는 레파지토리를, 파이썬의 코드로 clone 하는 방법
import os
import pathlib

if "models" in pathlib.Path.cwd().parts : 
  while "models" in pathlib.Path.cwd().part :
    os.chdir('..')
elif not pathlib.Path('models').exists() :
  ! git clone --depth 1 https://github.com/tensorflow/models

### 여러분들의 레파지토리에서 pull 하고 싶을때 
# os.chdir('/content/models')
# ! git pull

# Object Detection API 설치하기

# ! 느낌표 없이, 리눅스의 명령어를 실행시키고 싶으면, %%bash 라고 쓴후, 아래에다가 
# 리눅스 명령어를 쓴다.

%%bash
cd models/research/
protoc object_detection/protos/*.proto --python_out=.
cp object_detection/packages/tf2/setup.py .
python -m pip install .

import tensorflow as tf
import os
import pathlib

import numpy as np
import zipfile

import matplotlib.pyplot as plt
from PIL import Image

from IPython.display import display

from object_detection.utils import ops as utils_ops
from object_detection.utils import label_map_util
from object_detection.utils import visualization_utils as viz_utils

# 내 로컬에 설치된 레이블 파일을, 인덱스와 연결시킨다.
PATH_TO_LABELS = '/content/models/research/object_detection/data/mscoco_label_map.pbtxt'
category_index = label_map_util.create_category_index_from_labelmap(PATH_TO_LABELS)

print(category_index)

def download_model(model_name, model_date):
    base_url = 'http://download.tensorflow.org/models/object_detection/tf2/'
    model_file = model_name + '.tar.gz'
    model_dir = tf.keras.utils.get_file(fname=model_name,
                                        origin=base_url + model_date + '/' + model_file,
                                        untar=True)
    return str(model_dir)

MODEL_DATE = '20200711'
MODEL_NAME = 'centernet_hg104_1024x1024_coco17_tpu-32'
PATH_TO_MODEL_DIR = download_model(MODEL_NAME, MODEL_DATE)

def load_model(model_dir) :
    model_full_dir = model_dir + "/saved_model"

    # Load saved model and build the detection function
    detection_model = tf.saved_model.load(model_full_dir)
    return detection_model

detection_model = load_model(PATH_TO_MODEL_DIR)

PATH_TO_MODEL_DIR

# 우리가 가지고 있는 이미지 경로에서 이미지를 가져오는 코드
PATH_TO_IMAGE_DIR = pathlib.Path('/content/models/research/object_detection/test_images')
IMAGE_PATHS = sorted(list( PATH_TO_IMAGE_DIR.glob('*.jpg') ))

IMAGE_PATHS


### Object Detection ###

def show_inference(detection_model, image_np) :
    # The input needs to be a tensor, convert it using `tf.convert_to_tensor`.
    input_tensor = tf.convert_to_tensor(image_np)
    # The model expects a batch of images, so add an axis with `tf.newaxis`.
    input_tensor = input_tensor[tf.newaxis, ...]

    # input_tensor = np.expand_dims(image_np, 0)
    detections = detection_model(input_tensor)
#     print(detections)
    # All outputs are batches tensors.
    # Convert to numpy arrays, and take index [0] to remove the batch dimension.
    # We're only interested in the first num_detections.
    num_detections = int(detections.pop('num_detections'))
    detections = {key: value[0, :num_detections].numpy()
                   for key, value in detections.items()}
    detections['num_detections'] = num_detections

    # detection_classes should be ints.
    detections['detection_classes'] = detections['detection_classes'].astype(np.int64)

    # print(detections)
    image_np_with_detections = image_np.copy()

    viz_utils.visualize_boxes_and_labels_on_image_array(
          image_np_with_detections,
          detections['detection_boxes'],
          detections['detection_classes'],
          detections['detection_scores'],
          category_index,
          use_normalized_coordinates=True,
          max_boxes_to_draw=200,
          min_score_thresh=.30,
          agnostic_mode=False)

    display(Image.fromarray(image_np_with_detections))

for image_path in IMAGE_PATHS :
    image_np = np.array(Image.open(image_path))

    show_inference(detection_model, image_np)

# 모델을 바꿔서 할 수 있다.

# /20200711/ssd_mobilenet_v1_fpn_640x640_coco17_tpu-8.tar.gz

MODEL_DATE = '20200711'
MODEL_NAME = 'ssd_mobilenet_v1_fpn_640x640_coco17_tpu-8'
PATH_TO_MODEL_DIR = download_model(MODEL_NAME, MODEL_DATE)

def load_model(model_dir) :
    model_full_dir = model_dir + "/saved_model"

    # Load saved model and build the detection function
    detection_model = tf.saved_model.load(model_full_dir)
    return detection_model

detection_model = load_model(PATH_TO_MODEL_DIR)
PATH_TO_MODEL_DIR

#### Object Detection ###########

def show_inference(detection_model, image_np) :
    # The input needs to be a tensor, convert it using `tf.convert_to_tensor`.
    input_tensor = tf.convert_to_tensor(image_np)
    # The model expects a batch of images, so add an axis with `tf.newaxis`.
    input_tensor = input_tensor[tf.newaxis, ...]

    # input_tensor = np.expand_dims(image_np, 0)
    detections = detection_model(input_tensor)
#     print(detections)
    # All outputs are batches tensors.
    # Convert to numpy arrays, and take index [0] to remove the batch dimension.
    # We're only interested in the first num_detections.
    num_detections = int(detections.pop('num_detections'))
    detections = {key: value[0, :num_detections].numpy()
                   for key, value in detections.items()}
    detections['num_detections'] = num_detections

    # detection_classes should be ints.
    detections['detection_classes'] = detections['detection_classes'].astype(np.int64)
    
    # print(detections)
    image_np_with_detections = image_np.copy()

    viz_utils.visualize_boxes_and_labels_on_image_array(
          image_np_with_detections,
          detections['detection_boxes'],
          detections['detection_classes'],
          detections['detection_scores'],
          category_index,
          use_normalized_coordinates=True,
          max_boxes_to_draw=100,
          min_score_thresh=.50,
          agnostic_mode=False)

    display(Image.fromarray(image_np_with_detections))

for image_path in IMAGE_PATHS :
  image_np = np.array( Image.open(image_path) )

  show_inference(detection_model, image_np)