Keras CNN只能预测一个班级

Question

I've been trying to do some basic facial recognition using Keras, but I always get the same label for some reason. I am training it on 2 classes, Person A and Person B.

After its own validation testing, I input images that were neither in the training nor the validation set but it always gives me the same label. I have reduced the number of validation samples in this case for the sake of time, but even with more validation samples it doesn't appear to work. I have tried it with binary and sigmoid, but that doesn't help me either.

My accuracy often jumps up to 90% and even 100% at times, and my loss is around 0.05 to 0.1. Person A has 809 images while Person B has 777 images for training.

The end goal I need is to output probabilities of the image belonging to a certain class, and I would greatly appreciate any help since I am new to this field. Thanks!

from keras.preprocessing.image import ImageDataGenerator
from keras import backend as K
import keras
from keras.models import Sequential
from keras.layers import Convolution2D, MaxPooling2D
from keras.layers import Activation, Dropout, Flatten, Dense
K.set_image_dim_ordering('th')
import tensorflow as tf
config = tf.ConfigProto( device_count = {'GPU': 1 , 'CPU': 56} )
sess = tf.Session(config=config)
keras.backend.set_session(sess)
K.tensorflow_backend._get_available_gpus()

# Setting the image size
img_width, img_height = 250, 250

train_data_dir = 'data/train2'
validation_data_dir = 'data/validation2'
datagen = ImageDataGenerator(rescale=1./255)

# Train and Test Generators
train_generator = datagen.flow_from_directory(
        train_data_dir,
        target_size=(img_width, img_height),
        batch_size=32,
        class_mode = 'categorical')

validation_generator = datagen.flow_from_directory(
        validation_data_dir,
        target_size=(img_width, img_height),
        batch_size=32,
        class_mode = 'categorical')


#The actual CNN
model = Sequential()
model.add(Convolution2D(32, 3, 3, input_shape=(3, img_width, img_height)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Convolution2D(32, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2), data_format="channels_first", strides=(2, 2)))


model.add(Convolution2D(64, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2), data_format="channels_first", strides=(2, 2)))

model.add(Flatten())
model.add(Dropout(0.5))
model.add(Dense(2))
model.add(Activation('softmax'))

model.compile(loss='categorical_crossentropy',
              optimizer='adam',
              metrics=['categorical_accuracy'])

#Setting train conditions
nb_epoch = 50
nb_train_samples = 1586
nb_validation_samples = 122
print(validation_generator.class_indices)

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)

#Saving of the model
print(model.evaluate_generator(validation_generator, nb_validation_samples))
model.save_weights('my_ML_model3_weights.h5')
model.save('my_ML_model3.h5')
print("Done")

This is how I feed extra images to it for testing.

import cv2
from keras.models import load_model
import numpy as np

img_width, img_height = 250, 250
x = 1

while x < 6:
        img_width, img_height = img_width, img_height
        img = cv2.imread('D:\FaceRecog\Pictures\Person1 (%s).jpg' %(x))
        img = cv2.resize(img, (img_width, img_height))
        img = img.reshape(3, img_width, img_height)
        model = load_model("my_ML_model3.h5")
        model = model.predict(img[None, :, :, :])
        print(model)
        print('-----------')
        x += 1

Answer 1

You use scaling in your training and validation data:

datagen = ImageDataGenerator(rescale=1./255)

But I cannot find any scaling of your test data. That might be an issue, because the network would be trained on completely different input ranges. In general I would advice you to construct a similar image generator for your test images as for your training and validation data and see whether the problem persists.