CNN返回相同的分类结果（keras）

Question

我正在使用CNN对两种类型的花粉进行分类：sugi和hinoki。当我将在可见光下拍摄的图像用作数据时，它预测所有测试图像的“ sugi”。另一方面，当我使用紫外线拍摄的图像作为数据时，它对测试集中的所有照片都预测为“ hinoki”。我多次更改了时期，过滤器大小，批处理大小，通道数，但结果是相同的。我该怎么办？

这是我的代码：

培训计划：

import os
from keras.applications.vgg16 import VGG16
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential, Model
from keras.layers import Input, Activation, Dropout, Flatten, Dense, Conv2D, MaxPool2D
#from keras.callbacks import EarlyStoppingByLossVal
from keras.preprocessing.image import ImageDataGenerator
from keras import optimizers
import numpy as np
import time
from PIL import Image 
import csv
import shutil
#import numpy.distutils.system_info as sysinfo
import scipy
import scipy.misc
import matplotlib.pyplot as plt 
import pandas as pd 
# kaneko
from keras.callbacks import TensorBoard


#sysinfo.get_info('lapack')
# 分類するクラス
classes = ['sugi', 'hinoki']
nb_classes = len(classes)


img_width, img_height = 100, 100

# トレーニング用とバリデーション用の画像格納先
train_data_dir = 'cut.kashi/train'
validation_data_dir = 'cut.kashi/validation'

# 今回はトレーニング用に200枚、バリデーション用に50枚の画像を用意した。
nb_train_samples = 1362
nb_validation_samples = 337
#nb_train_samples = 2171
#nb_validation_samples = 528

#batch_size = 64
nb_epoch = 50
gen_tr_batches = 4
folder = './output'
result_dir = 'results'
if not os.path.exists(result_dir):
    os.mkdir(result_dir)
train_imagelist = os.listdir(train_data_dir)

test_list = "./test.train"
font = cv2.FONT_HERSHEY_COMPLEX

def vgg_model_maker():

    model = Sequential()

    model.add(Conv2D(32,5,input_shape=(img_width, img_height,3)))
    model.add(Activation('relu'))
    #model.add(Conv2D(32,5))
    #model.add(Activation('relu'))
    model.add(MaxPool2D(pool_size=(2,2)))

    model.add(Conv2D(64,5))
    model.add(Activation('relu'))
    model.add(MaxPool2D(pool_size=(2,2)))

    model.add(Flatten())
    model.add(Dense(200))
    model.add(Activation('relu'))
    #model.add(Dropout(1.0))

    model.add(Dense(nb_classes, activation='softmax'))


    return model



def image_generator():
    """ ディレクトリ内の画像を読み込んでトレーニングデータとバリデーションデータの作成 """
    train_datagen = ImageDataGenerator(
        rescale=1.0 / 255,
        zoom_range=0.2,
        horizontal_flip=True,
        rotation_range = 180)


    validation_datagen = ImageDataGenerator(rescale=1.0 / 255)

    train_generator = train_datagen.flow_from_directory(
        train_data_dir,
        target_size=(img_width, img_height),
        color_mode='rgb',
        classes=classes,
        class_mode='categorical',
        batch_size=batch_size,
        shuffle=True)

    validation_generator = validation_datagen.flow_from_directory(
        validation_data_dir,
        target_size=(img_width, img_height),
        color_mode='rgb',
        classes=classes,
        class_mode='categorical',
        batch_size=batch_size,
        shuffle=True)

    return (train_generator,validation_generator)


def global_contrast_normalization(filename, s, lmda, epsilon):
    X = numpy.array(Image.open(filename))

    # replacement for the loop
    X_average = numpy.mean(X)
    print('Mean: ', X_average)
    X = X - X_average

    # `su` is here the mean, instead of the sum
    contrast = numpy.sqrt(lmda + numpy.mean(X**2))

    X = s * X / max(contrast, epsilon)

    # scipy can handle it
    scipy.misc.imsave('result.jpg', X)



# Generator for the network's training generator.



# Actual generator for the network's training.


if __name__ == '__main__':
    start = time.time()

    for the_file in os.listdir(folder):
        file_path = os.path.join(folder, the_file)
        try:
            if os.path.isfile(file_path):
                os.unlink(file_path)
        #elif os.path.isdir(file_path): shutil.rmtree(file_path)
        except Exception as e:
            print(e)
     # kaneko
    tensorboard = TensorBoard(log_dir="./kaneko", histogram_freq=0, batch_size= batch_size,write_graph=True)
    # モデル作成
    vgg_model = vgg_model_maker()

    # 最後のconv層の直前までの層をfreeze
    #for layer in vgg_model.layers[:15]:
        #layer.trainable = False

    # 多クラス分類を指定
    vgg_model.compile(loss='categorical_crossentropy',
              optimizer=optimizers.SGD(lr=1e-3, momentum=0.9),
              metrics=['accuracy'])

    # 画像のジェネレータ生成
    train_generator,validation_generator =  image_generator()



    # Fine-tuning
    history_callback = vgg_model.fit_generator(
        train_generator,
        samples_per_epoch=nb_train_samples,
        nb_epoch=nb_epoch,
        validation_data = validation_generator,
        nb_val_samples=nb_validation_samples,
        callbacks=[tensorboard])

    loss_history = history_callback.history["loss"]
    accuracy_history = history_callback.history["acc"]
    val_loss_history = history_callback.history["val_loss"]
    val_accuracy_history = history_callback.history["val_acc"]
    numpy_loss_history = np.array(loss_history)
    numpy_accuracy_history = np.array(accuracy_history)
    numpy_val_loss_history = np.array(val_loss_history)
    numpy_val_accuracy_history = np.array(val_accuracy_history)

    f = open("results/result.csv","w")
    writer = csv.writer(f)
    writer.writerow(["loss","accuracy","validation loss","validation accuracy"])
    for j in range(len(numpy_loss_history)):
        writer.writerow([numpy_loss_history[j],numpy_accuracy_history[j],numpy_val_loss_history[j],numpy_val_accuracy_history[j]])
    epochnum = range(len(numpy_loss_history))
    print(len(epochnum))
    #plt.plot(epochnum,numpy_loss_history, label = "loss")
    #plt.legend()
    plt.plot(loss_history)
    plt.plot(val_loss_history)
    plt.legend(['loss', 'val_loss'])
    plt.show()
    #plt.savefig("./Documents/Ghi1/shigaisen_loss.png")
    plt.clf()
    plt.plot(epochnum,numpy_accuracy_history, label = "accuracy")
    plt.show()
    #plt.savefig(".../Documents/Ghi1/shigaisen_accuracy.png")
    plt.clf()
    vgg_model.save_weights(os.path.join(result_dir, 'finetuning.h5'))

    process_time = (time.time() - start) / 60
    print(u'学習終了。かかった時間は', process_time, u'分です。')

测试程序：

import os, sys
import numpy as np
import cv2

from keras.applications.vgg16 import VGG16
from keras.models import Sequential, Model
from keras.layers import Input, Activation, Dropout, Flatten, Dense, Conv2D,MaxPool2D
from keras.preprocessing import image
from keras.preprocessing.image import ImageDataGenerator
from keras import optimizers
from datetime import datetime

classes = ['sugi', 'hinoki']
nb_classes = len(classes)
img_width, img_height = 100, 100
DataShape = (100,100,3)
result_dir = 'results'
#test_list = "./testfile"
test_list = "./test.train"
font = cv2.FONT_HERSHEY_COMPLEX
# このディレクトリにテストしたい画像を格納しておく
test_data_dir = 'cut/test'
folder = './output'
def model_load():
    # VGG16, FC層は不要なので include_top=False
    model = Sequential()

    model.add(Conv2D(32,5,input_shape=(img_width, img_height,3)))
    model.add(Activation('relu'))
    #model.add(Conv2D(32,5))
    #model.add(Activation('relu'))
    model.add(MaxPool2D(pool_size=(2,2)))

    model.add(Conv2D(64,5))
    model.add(Activation('relu'))
    model.add(MaxPool2D(pool_size=(2,2)))

    model.add(Flatten())
    model.add(Dense(200))
    model.add(Activation('relu'))
    #model.add(Dropout(1.0))

    model.add(Dense(nb_classes, activation='softmax'))

    #adam = Adam(lr=1e-4)

    # 学習済みの重みをロード
    model.load_weights(os.path.join(result_dir, 'finetuning.h5'))

    # 多クラス分類を指定
    model.compile(loss='categorical_crossentropy',
              optimizer=optimizers.SGD(lr=1e-3, momentum=0.9),
              metrics=['accuracy'])

    return model

def image_generator():
    """ ディレクトリ内の画像を読み込んでトレーニングデータとバリデーションデータの作成 """
    test_datagen = ImageDataGenerator(
        rescale=1.0 / 255,
        zoom_range=0.2,
        horizontal_flip=True,
        rotation_range = 180)

    #validation_datagen = ImageDataGenerator(rescale=1.0 / 255)

    test_generator = test_datagen.flow_from_directory(
        test_data_dir,
        target_size=(img_width, img_height),
        color_mode='rgb',
        classes=classes,
        class_mode='categorical',
        batch_size=batch_size,
        shuffle=True)

def image_resize(image, width = None, height = None, inter = cv2.INTER_AREA):
    # initialize the dimensions of the image to be resized and
    # grab the image size
    dim = None
    (h, w) = image.shape[:2]

    # if both the width and height are None, then return the
    # original image
    if width is None and height is None:
        return image

    # check to see if the width is None
    if width is None:
        # calculate the ratio of the height and construct the
        # dimensions
        r = height / float(h)
        dim = (int(w * r), height)

    # otherwise, the height is None
    else:
        # calculate the ratio of the width and construct the
        # dimensions
        r = width / float(w)
        dim = (width, int(h * r))

    # resize the image
    resized = cv2.resize(image, dim, interpolation = inter)

    # return the resized image
    return resized

def test(model,path,filename,sugi):
    test_imagelist = []

    # テスト用画像取得
    #test_imagelist = os.listdir(test_data_dir)

    #test_imagelist = os.listdir(test_data_dir)

    iml = cv2.imread(path,cv2.IMREAD_COLOR)


    img = image_resize(iml,height=960)
    img_array = np.array(img)

    cimg = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
    cimg = cv2.medianBlur(cimg,5)

    #_,cimg = cv2.threshold(cimg,0,255,cv2.THRESH_BINARY| cv2.THRESH_OTSU)
    #cv2.imwrite(datetime.now().strftime('%s')+"binary.jpg",cimg)
    #sys.exit()

    circles = cv2.HoughCircles(cimg,cv2.HOUGH_GRADIENT,1,10,param1=15,param2=20,minRadius=10,maxRadius=25)
    circles = np.uint16(np.around(circles))[0,:]
    print (len(circles))
    center = []
    predict = []
    for i in circles:
        half = DataShape[0]//2
        zoom_data = img_array[i[1]-half:i[1]+half,i[0]-half:i[0]+half,:]
        if zoom_data.shape!=DataShape : continue
        czoom = cv2.cvtColor(zoom_data, cv2.COLOR_BGR2GRAY)
        czoomarr = np.array(zoom_data)
        cen = czoom[half,half]
        #edge = czoom[0,0]
        if cen != 0:
        #if cen < 255:
        #if czoom[30,30] < 80:
            test_imagelist.append(zoom_data)
            center.append(i)
        label_num = len(test_imagelist)

    print(len(center))
    print(label_num)

    for im in test_imagelist:
        x = image.img_to_array(im)
        x = np.expand_dims(x, axis=0)
        # 学習時に正規化してるので、ここでも正規化
        x = x / 255
        pred = model.predict(x)[0]
        print(pred)
        predict.append(pred)
    TP = 0
    TN = 0
    FN = 0
    FP = 0
    for j in range(label_num):
        if predict[j][0] > predict[j][1]:
            if sugi == 1:
                #TP+=1
                TN+=1
            else:
                #FP+=1
                FN+=1
            #cv2.circle(img,(center[j][0],center[j][1]),center[j][2],(0,255,0),2)
            cv2.putText(img,'S',(center[j][0],center[j][1]), font, 0.5,(0,255,0),1,cv2.LINE_AA)
        if predict[j][0] < predict[j][1]:
            #cv2.circle(img,(center[j][0],center[j][1]),center[j][2],(0,0,255),2)
            if sugi == 1:
                #FN+=1
                FP+=1
            else:
                #TN+=1
                TP+=1
            cv2.putText(img,'H',(center[j][0],center[j][1]), font,0.5,(0,0,255),1,cv2.LINE_AA)

    cv2.imwrite("output/"+"output"+filename,img)

    return TP, FP, FN, TN


if __name__ == '__main__':

    # モデルのロード
    TP,FP,FN,TN = 0,0,0,0
    print(TP,FP,FN,TN) 
    sugi = 0
    c = "ス"
    model = model_load()
    for the_file in os.listdir(folder):
        file_path = os.path.join(folder, the_file)
        try:
            if os.path.isfile(file_path):
                os.unlink(file_path)
        #elif os.path.isdir(file_path): shutil.rmtree(file_path)
        except Exception as e:
            print(e)

    for the_file in os.listdir(test_list):
        #print(the_file)
        if c in the_file:
            sugi = 1
        else:
            sugi = 0
        file_path = os.path.join(test_list, the_file)
        tp1,fp1,fn1,tn1 = test(model,file_path,the_file,sugi)
        TP += tp1
        FP += fp1
        FN += fn1
        TN += tn1


    precision = TP/(TP + FP)
    recall = TP/(TP + FN)
    F = (2*recall*precision)/(recall + precision)
    #cv2.imwrite("output/" + "result.jpg",img)

    print("TP = %lf, TN = %lf, FN = %lf, FP = %lf" %(TP,TN,FN,FP))
    print("precision = %lf, recall = %lf" %(precision,recall))
    print("F measure = %lf" %(F))

Answer 1

我可以看到的一个问题是x = x / 255方法中的test。您需要获取float值以进行适当的标准化。我遇到了同样的问题，适当的缩放使其正常工作。这是link

我希望这会有所帮助。

编辑：我的答案正在考虑使用python2。

Answer 2

我怀疑您的文件夹结构错误。

ImageDataGenerator将根据您使用的文件夹结构创建类。

您应该在“ datadir”中：

• 一个包含所有sugi图像的“ sugi”文件夹
• 一个带有所有hinoki图像的“ hinoki”文件夹

但是似乎您有：

• 一个“可见”文件夹
• 一个“紫外线”文件夹

这肯定会使生成器认为“ visible = sugi”和“ ultraviolet = hinoki”。