Top_5_Correctness

MobileNetV2를 활용한 ImageNet 분류

 

1. Tensorflow Hub

import tensorflow
import numpy as np
import tensorflow_hub as hub

url = 'https://tfhub.dev/google/tf2-preview/mobilenet_v2/classification/2'

#사전 훈련된 MobileNetV2
MbNetV2 = tensorflow.keras.Sequential([hub.KerasLayer(handle = url,
                                                      input_shape = (224, 224, 3), 
                                                      trainable = False)])

MbNetV2.summary()

from tensorflow.keras import utils

utils.plot_model(MbNetV2,
                 show_shapes = True,
                 show_dtype = True)

---

2. ImageNetV2 TopImages

2-1. 이미지(ImageNetV2) 다운로드

%%time

import os

image_url = 'https://s3-us-west-2.amazonaws.com/imagenetv2public/imagenetv2-top-images.tar.gz'
data_dir = '/content/'

data_root_orig = tensorflow.keras.utils.get_file('imagenetV2', 
                                                 image_url, 
                                                 cache_dir = data_dir, 
                                                 extract = True)

!ls /content/datasets/imagenetv2-top-images-format-val

이미지별 디렉토리 확인

---

2-2. Image Label 다운로드

import pandas as pd

url = 'https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt'
DF = pd.read_table(url, header = None)

label_text = DF.values.reshape(-1)

print(len(label_text))
print(label_text[:5])
print(label_text[-5:])

1001개 주의!!!

---

2-3. Image(X) vs. Label(y) 확인

import random

all_image_paths = list(data_root.glob('*/*'))
all_image_paths = [str(path) for path in all_image_paths]

random.choice(all_image_paths)

import PIL.Image as Image
import matplotlib.pyplot as plt

random.shuffle(all_image_paths)

image_count = len(all_image_paths)
print('image_count:', image_count)

plt.figure(figsize = (12, 12))
for n in range(9):
    image_path = random.choice(all_image_paths)
    plt.subplot(3, 3 , n + 1)
    plt.imshow(plt.imread(image_path))
    idx = int(image_path.split('/')[-2]) + 1
    plt.title(str(idx) + ' : ' + label_text[idx])
    plt.axis('off')
plt.show()

---

3. MobileNet의 분류 라벨 확인

import cv2

plt.figure(figsize = (16, 16))

def softmax(x):
    m = np.max(x)
    sm = np.exp(x - m)/np.sum(np.exp(x - m))
    return sm
  
for n in range(3):
    # 랜던 이미지 경로 추출
    image_path = random.choice(all_image_paths)
    
    # 이미지 출력
    plt.subplot(3, 2, n * 2 + 1)
    plt.imshow(plt.imread(image_path))
    idx = int(image_path.split('/')[-2]) + 1
    plt.title(str(idx) + ' : ' + label_text[idx])
    plt.axis('off')
    
    # Top-5 예측값 출력 공간 
    plt.subplot(3, 2 , n * 2 + 2)

    # 이미지 크기 변경 및 정규화
    img = cv2.imread(image_path)
    img = cv2.resize(img, dsize = (224, 224))
    img = img / 255.0
    img = np.expand_dims(img, axis = 0)

    # MobileNetV2 이미지 분류
    logits = MbNetV2.predict(img)[0]
    prediction = softmax(logits)
    
    # Top-5 예측값
    top_5_predict = prediction.argsort()[::-1][:5]
    labels = [label_text[index] for index in top_5_predict]  

    # Top-5 막대그래프 색상 지정
    color = ['gray'] * 5
    if idx in top_5_predict:
        color[top_5_predict.tolist().index(idx)] = 'green'
    color = color[::-1]

    # Top-5 막대그래프 출력
    plt.barh(range(5), prediction[top_5_predict][::-1] * 100, color = color)
    plt.yticks(range(5), labels[::-1])

AR = np.array([5555, 55, 5, 555])
print('오름차순 정렬 :', np.sort(AR), '\n')

arg = np.argsort(AR)
print('오름차순 정렬 후 Index :', arg, '\n')

print('argsort 결과 적용 :', AR[arg])

---

4. MobileNetV2의 분류 성능 확인

%%time

import cv2

top_1 = 0
top_5 = 0

for image_path in all_image_paths:
    img = cv2.imread(image_path)
    img = cv2.resize(img, dsize = (224, 224))
    img = img / 255.0
    img = np.expand_dims(img, axis = 0)

    top_5_predict = MbNetV2.predict(img)[0].argsort()[::-1][:5]
    idx = int(image_path.split('/')[-2]) + 1

    if idx in top_5_predict:
        top_5 =  top_5 + 1
        if top_5_predict[0] == idx:
            top_1 = top_1 + 1

print('Top-5 Correctness:', top_5 / len(all_image_paths) * 100, '%')
print('Top-1 Correctness:', top_1 / len(all_image_paths) * 100, '%')