让我给你一些背景
数据集：
https://www.kaggle.com/dionyshsmiaris/xrays 包括类的 X 射线：
0/"normal" (1227 xrays)
1/"pneumonia_cause_viral_infection" (2238)
?/"pneumonia_cause_bacteria" (1207)
导入数据：
train_dir = "/content/gdrive/MyDrive/Xraydataset/train_images/"
test_dir = "/content/gdrive/MyDrive/Xraydataset/test_images/"
def get_data(folder):
X = []
y = []
for image_filename in tqdm(os.listdir(folder)):#/content/gdrive/MyDrive/Xraydataset/train_images/ == folder
img_file = tf.keras.preprocessing.image.load_img(folder + '/' + image_filename,color_mode="grayscale")
if img_file is not None:
img_file=img_file.resize((224,224),1)
#img_file = tf.keras.preprocessing.image.smart_resize(img_file, (224, 224)) #256 256 1?
#img_file = scipy.misc.imresize(arr=img_file, size=(150, 150, 3))
img_arr = np.asarray(img_file)
X.append(img_arr)
label=pd.read_csv('/content/gdrive/MyDrive/Xraydataset/labels_train.csv',usecols = ['class_id'])
# y.append(label)
X = np.asarray(X)
y = np.asarray(label)
return X,y
X_train, y_train = get_data(train_dir)
X_test, y_test= get_data(test_dir)
预处理：
X_train=np.expand_dims(X_train, axis=3)
X_test=np.expand_dims(X_test, axis=3)
X_train = X_train.astype('float32') / 255
X_test = X_test.astype('float32') / 255
X_train_mean = np.mean(X_train, axis=0)
X_train -= X_train_mean
X_test -= X_train_mean
print('x_train shape:', X_train.shape)
print(X_train.shape[0], 'train samples')
print(X_test.shape[0], 'test samples')
xtrain=X_train[:3700]
ytrain=y_train[:3700]
#split data to train and test
xtest=X_train[3700:]
ytest=y_train[3700:]
t_train = keras.utils.to_categorical(ytrain, 3)
t_test = keras.utils.to_categorical(ytest, 3)
数据生成：
datagen = ImageDataGenerator(
featurewise_center=False, # set input mean to 0 over the dataset
samplewise_center=False, # set each sample mean to 0
featurewise_std_normalization=False, # divide inputs by std of the dataset
samplewise_std_normalization=False, # divide each input by its std
zca_whitening=False, # apply ZCA whitening
rotation_range = 30, # randomly rotate images in the range (degrees, 0 to 180)
zoom_range = 0.2, # Randomly zoom image
width_shift_range=0.1, # randomly shift images horizontally (fraction of total width)
height_shift_range=0.1, # randomly shift images vertically (fraction of total height)
horizontal_flip = True, # randomly flip images
vertical_flip=False) # randomly flip images
datagen.fit(xtrain)
编译：
model = Sequential()
model.add(Conv2D(32 , (3,3) , strides = 1 , padding = 'same' , activation = 'relu' , input_shape = (224,224,1)))
model.add(BatchNormalization())
model.add(MaxPool2D((2,2) , strides = 2 , padding = 'same'))
model.add(Conv2D(64 , (3,3) , strides = 1 , padding = 'same' , activation = 'relu'))
model.add(Dropout(0.1))
model.add(BatchNormalization())
model.add(MaxPool2D((2,2) , strides = 2 , padding = 'same'))
model.add(Conv2D(64 , (3,3) , strides = 1 , padding = 'same' , activation = 'relu'))
model.add(BatchNormalization())
model.add(MaxPool2D((2,2) , strides = 2 , padding = 'same'))
model.add(Conv2D(128 , (3,3) , strides = 1 , padding = 'same' , activation = 'relu'))
model.add(Dropout(0.2))
model.add(BatchNormalization())
model.add(MaxPool2D((2,2) , strides = 2 , padding = 'same'))
model.add(Conv2D(256 , (3,3) , strides = 1 , padding = 'same' , activation = 'relu'))
model.add(Dropout(0.2))
model.add(BatchNormalization())
model.add(MaxPool2D((2,2) , strides = 2 , padding = 'same'))
model.add(Flatten())
model.add(Dense(units = 128 , activation = 'relu'))
model.add(Dropout(0.2))
model.add(Dense(units = 3 , activation = 'softmax'))
model.compile(optimizer = "Adam" , loss = 'categorical_crossentropy' , metrics = ['accuracy'])
学习率、训练
learning_rate_reduction = ReduceLROnPlateau(monitor='val_accuracy',
patience = 2,
verbose=1,
factor=0.3,
min_lr=0.000001)
history = model.fit(datagen.flow(xtrain,t_train, batch_size = 32) ,shuffle=True,
epochs = 20 , validation_data =datagen.flow(xtest,t_test),
callbacks = [learning_rate_reduction])
输出：
[[0.276 0.455 0.27 ]
[0.291 0.423 0.286]
[0.286 0.434 0.28 ]
...
[0.288 0.43 0.282]
[0.292 0.429 0.28 ]
[0.289 0.427 0.284]]
典型的 acc 约为 46%，验证 acc 为 49%，但那是因为它预测频率最高的类别
损失也几乎相同，所以我认为它没有训练任何提示/帮助
编辑:(class_weights = {0: .9837, 1: .5394, 2: 1.0})
Epoch 1/20
116/116 [==============================] - 19s 154ms/step - loss: 0.8551 - accuracy: 0.3216 - val_loss: 1.1174 - val_accuracy: 0.3302
Epoch 2/20
116/116 [==============================] - 18s 155ms/step - loss: 0.8587 - accuracy: 0.3265 - val_loss: 1.1073 - val_accuracy: 0.3261
Epoch 3/20
116/116 [==============================] - 18s 158ms/step - loss: 0.8554 - accuracy: 0.3330 - val_loss: 1.1114 - val_accuracy: 0.3241
.
.
.
Epoch 18/20
116/116 [==============================] - 18s 155ms/step - loss: 0.8585 - accuracy: 0.3208 - val_loss: 1.1097 - val_accuracy: 0.3251
Epoch 19/20
116/116 [==============================] - 18s 154ms/step - loss: 0.8556 - accuracy: 0.3378 - val_loss: 1.1084 - val_accuracy: 0.3395
Epoch 20/20
116/116 [==============================] - 18s 156ms/step - loss: 0.8562 - accuracy: 0.3422 - val_loss: 1.1127 - val_accuracy: 0.3364
输出：
[[0.333 0.345 0.322]
[0.361 0.283 0.356]
[0.319 0.343 0.338]
...
[0.335 0.329 0.335]
[0.439 0.179 0.383]
[0.338 0.311 0.351]]
从我现在得到的情况来看，我的模型能够猜测所有三个类别，但如果你看损失，它不会学习，它保持不变，acc 约为 30-33%，这会导致随机猜测任何提示？< /p>
答案 0 :(得分：1)
我复制了您的代码并运行了它，并得到了与您相同的结果。为了查看模型是否是问题，我修改了代码以将迁移学习与 Resnet 结合使用。得到了基本相同的结果。我去 Kaggle 看看有没有结果更好的笔记本。没有笔记本。我认为问题出在数据集上。您可以在 https://www.kaggle.com/bachrr/covid-chest-xray 的 Kaggle 上使用替代数据集。还有这个数据集https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia