这是过拟合的情况吗？ 我该怎么做才能修复我的 model？

Question

我对 CNN 很陌生，我一直在研究 CNN model，它在 506 个图像的小数据集上进行训练。 （类别 0 为 306，类别 1 为 200）。 我在 Google Colab 上完成了所有工作。 我从网上可以找到的许多示例中挑选、混合和玩弄 model。 我设法使模型的训练精度达到 0.8022 和val_acc 0.8431 model.evaluate 给出：损失：0.6061 - 精度：0.8755

现在，我的问题是，我尝试用我的 model 做 model.predict。 预测很糟糕，每个类别 0 和 1 的每个图像，我的 model 将其分类为类别 0。所有的预测结果都只是 0。（我喂给 model 等量的类别 0 和 1）

这是我的代码：

import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np
import cv2
import os
import glob
from tensorflow import keras
from keras import layers
from keras.models import Sequential, model_from_json
from keras.layers import Dense, Dropout, Flatten, BatchNormalization, Activation
from keras.layers import Conv2D, MaxPooling2D
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten, Conv2D, MaxPooling2D
from keras.preprocessing import image
from keras.regularizers import l2 


DATADIR = r"/content/gdrive/MyDrive/Dataset/Training"
CATEGORIES = ['Abnormal', 'Normal']


training_data = []

def create_training_data():
    for category in CATEGORIES:
        path = os.path.join(DATADIR, category)
        class_num = CATEGORIES.index(category)
        for img in os.listdir(path):
            try:
                img_array = cv2.imread(os.path.join(path,img), cv2.IMREAD_GRAYSCALE)
                new_array = cv2.resize(img_array, (1000, 400))
                training_data.append([new_array, class_num])
            except Exception as e:
                pass
                
create_training_data()

print(len(training_data))

import random
random.shuffle(training_data)

for sample in training_data:
    print(sample[1])

X = []
y = []

for features, label in training_data:
    X.append(features)
    y.append(label)
X = np.array(X).reshape(-1, 1000, 400, 1)
y = np.array(y)

import pickle

pickle_out = open("X.pickle","wb")
pickle.dump(X, pickle_out)
pickle_out.close()

pickle_out = open("y.pickle","wb")
pickle.dump(y, pickle_out)
pickle_out.close()
pickle_in = open("X.pickle", "rb")
X = pickle.load(pickle_in)

X = pickle.load(open("X.pickle", "rb"))
y = pickle.load(open("y.pickle", "rb"))

X = X/255.0



model = Sequential()


model.add(Conv2D(128, (3,3), activation = 'relu', input_shape=(1000,400,1), kernel_regularizer=l2(0.001)))
model.add(Dropout(0.5))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(64, (3, 3),activation = 'relu',kernel_regularizer=l2(0.001)))
model.add(Dropout(0.5))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(32, (3, 3),activation = 'relu',kernel_regularizer=l2(0.001)))
model.add(Dropout(0.5))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Flatten())
model.add(Dense(128 , activation='relu',kernel_regularizer=l2(0.001)))
model.add(Dropout(0.4))
model.add(Dense(1, activation='sigmoid',kernel_regularizer=l2(0.001)))
model.summary()

model.compile(loss='binary_crossentropy',

              optimizer=keras.optimizers.Adam(learning_rate=0.0001),

              metrics=['accuracy'])

callback = tf.keras.callbacks.EarlyStopping(monitor='loss', patience=2)

model.fit(X, y, shuffle=True, batch_size = 7, epochs = 20, validation_split = 0.1, callbacks=[callback])
model.evaluate(X, y)

这里是model总结和培训：

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d (Conv2D)             (None, 998, 398, 128)     1280      
                                                                 
 dropout (Dropout)           (None, 998, 398, 128)     0         
                                                                 
 max_pooling2d (MaxPooling2D  (None, 499, 199, 128)    0         
 )                                                               
                                                                 
 conv2d_1 (Conv2D)           (None, 497, 197, 64)      73792     
                                                                 
 dropout_1 (Dropout)         (None, 497, 197, 64)      0         
                                                                 
 max_pooling2d_1 (MaxPooling  (None, 248, 98, 64)      0         
 2D)                                                             
                                                                 
 conv2d_2 (Conv2D)           (None, 246, 96, 32)       18464     
                                                                 
 dropout_2 (Dropout)         (None, 246, 96, 32)       0         
                                                                 
 max_pooling2d_2 (MaxPooling  (None, 123, 48, 32)      0         
 2D)                                                             
                                                                 
 flatten (Flatten)           (None, 188928)            0         
                                                                 
 dense (Dense)               (None, 128)               24182912  
                                                                 
 dropout_3 (Dropout)         (None, 128)               0         
                                                                 
 dense_1 (Dense)             (None, 1)                 129       
                                                                 
=================================================================
Total params: 24,276,577
Trainable params: 24,276,577
Non-trainable params: 0
_________________________________________________________________
Epoch 1/20
65/65 [==============================] - 35s 338ms/step - loss: 0.9851 - accuracy: 0.5692 - val_loss: 0.9020 - val_accuracy: 0.6275
Epoch 2/20
65/65 [==============================] - 21s 328ms/step - loss: 0.8785 - accuracy: 0.5956 - val_loss: 0.8668 - val_accuracy: 0.6275
Epoch 3/20
65/65 [==============================] - 21s 331ms/step - loss: 0.8428 - accuracy: 0.5956 - val_loss: 0.8425 - val_accuracy: 0.6275
Epoch 4/20
65/65 [==============================] - 22s 333ms/step - loss: 0.8229 - accuracy: 0.5934 - val_loss: 0.8219 - val_accuracy: 0.6275
Epoch 5/20
65/65 [==============================] - 22s 335ms/step - loss: 0.7974 - accuracy: 0.6154 - val_loss: 0.8033 - val_accuracy: 0.6275
Epoch 6/20
65/65 [==============================] - 22s 334ms/step - loss: 0.7646 - accuracy: 0.5956 - val_loss: 0.7968 - val_accuracy: 0.6275
Epoch 7/20
65/65 [==============================] - 22s 334ms/step - loss: 0.7629 - accuracy: 0.6264 - val_loss: 0.7898 - val_accuracy: 0.6275
Epoch 8/20
65/65 [==============================] - 22s 334ms/step - loss: 0.7467 - accuracy: 0.6176 - val_loss: 0.7761 - val_accuracy: 0.6667
Epoch 9/20
65/65 [==============================] - 22s 335ms/step - loss: 0.7191 - accuracy: 0.6286 - val_loss: 0.7667 - val_accuracy: 0.6667
Epoch 10/20
65/65 [==============================] - 22s 334ms/step - loss: 0.7222 - accuracy: 0.6527 - val_loss: 0.7674 - val_accuracy: 0.6863
Epoch 11/20
65/65 [==============================] - 22s 335ms/step - loss: 0.6664 - accuracy: 0.6769 - val_loss: 0.7440 - val_accuracy: 0.6863
Epoch 12/20
65/65 [==============================] - 22s 334ms/step - loss: 0.6688 - accuracy: 0.6769 - val_loss: 0.7550 - val_accuracy: 0.6471
Epoch 13/20
65/65 [==============================] - 22s 334ms/step - loss: 0.6535 - accuracy: 0.7077 - val_loss: 0.7611 - val_accuracy: 0.7059
Epoch 14/20
65/65 [==============================] - 22s 334ms/step - loss: 0.6221 - accuracy: 0.7187 - val_loss: 0.7487 - val_accuracy: 0.7255
Epoch 15/20
65/65 [==============================] - 22s 334ms/step - loss: 0.5909 - accuracy: 0.7473 - val_loss: 0.7286 - val_accuracy: 0.7451
Epoch 16/20
65/65 [==============================] - 22s 335ms/step - loss: 0.5814 - accuracy: 0.7516 - val_loss: 0.7235 - val_accuracy: 0.7647
Epoch 17/20
65/65 [==============================] - 22s 334ms/step - loss: 0.5653 - accuracy: 0.7714 - val_loss: 0.7264 - val_accuracy: 0.8039
Epoch 18/20
65/65 [==============================] - 22s 334ms/step - loss: 0.5367 - accuracy: 0.8000 - val_loss: 0.6934 - val_accuracy: 0.8039
Epoch 19/20
65/65 [==============================] - 22s 334ms/step - loss: 0.5147 - accuracy: 0.8022 - val_loss: 0.7242 - val_accuracy: 0.7255
Epoch 20/20
65/65 [==============================] - 22s 335ms/step - loss: 0.4832 - accuracy: 0.7824 - val_loss: 0.6765 - val_accuracy: 0.8431
16/16 [==============================] - 15s 747ms/step - loss: 0.6061 - accuracy: 0.8755
[0.6060804724693298, 0.8754940629005432]

此外，这就是我使用 model.predict 的方式：

for i in os.listdir(testdata):
    img = image.load_img(testtest+'//'+ i, color_mode = "grayscale", target_size=(1000,400))
    xtest = image.img_to_array(img)
    xtest = np.expand_dims(xtest, axis = 0)
    images = np.vstack([xtest])
    val = model.predict(images)
    val2 = model.evaluate(images, verbose = 0)
    print (np.argmax(val))

Answer 1

预测很可怕，每个类别 0 和 1 的每个图像，我的 model 将其分类为类别 0。

这实际上在二元分类中很常见。 你的训练集中的分布是什么？ 鉴于您的训练精度为 0.8022，听起来您的训练数据的 80.22% 是 class 0，而 model 只是学会了选择大多数 ZA2F2ED4F8EBC2CBB4C21A29DC40AB61