用更多类别重新训练现有构建的 keras cnn 顺序模型（预测 16 个类别）

Question

以下是创建模型并将其保存在本地目录的代码。这里所有的图像都放在本地标记的文件夹中。现在我想在不同的标签文件夹中添加更多图像并将它们包含到这个模型中。所以总的来说，我不想从头开始重新训练模型，而是增加新标签

from keras.layers import Conv2D, Activation, MaxPooling2D, Flatten, Dense
from keras.models import Sequential
from keras.optimizers import Adam

def readTestData(testDir):
data = []
filenames = []
# loop over the input images
images = os.listdir(testDir)
for imageFileName in images:
    # load the image, pre-process it, and store it in the data list
    imageFullPath = os.path.join(testDir, imageFileName)
    #print(imageFullPath)
    img = load_img(imageFullPath)
    arr = img_to_array(img)  # Numpy array with shape (...,..,3)
    arr = cv2.resize(arr, (HEIGHT,WIDTH)) 
    data.append(arr)
    filenames.append(imageFileName)
    return data, filenames
 def createModel():
    #model = Sequential()
    #model.add(Conv2D(20, (5, 5), padding="same", input_shape=inputShape))
    #model.add(Activation("relu"))
    #model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    #model.add(Conv2D(50, (5, 5), padding="same"))
    #model.add(Activation("relu"))
    #model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    #model.add(Flatten())
    #model.add(Dense(500))
    #model.add(Activation("relu"))
    #model.add(Dense(output_dim=22))
    #model.add(Activation("softmax"))
    model = load_model('test')
    model.pop()
    model.pop()
    for layer in model.layers:
     layer.trainable = False
    model.add(Dense(output_dim=24,name='new_Dense',activation='softmax'))

    opt = Adam(lr=INIT_LR, decay=INIT_LR / EPOCHS)
    model.compile(loss="categorical_crossentropy", optimizer=opt, metrics= 
["accuracy"])
    return model    

random.seed(10)

X, Y = readTrainData("labelled images directory path")
# scale the raw pixel intensities to the range [0, 1]
X = np.array(X, dtype="float") / 255.0
Y = np.array(Y)
# convert the labels from integers to vectors
Y =  to_categorical(Y, num_classes=22)

(trainX, valX, trainY, valY) = train_test_split(X,Y,test_size=0.10, 
 random_state=10)

aug = ImageDataGenerator(rotation_range=30, width_shift_range=0.1, \
height_shift_range=0.1, shear_range=0.2, zoom_range=0.2,\
horizontal_flip=True, fill_mode="nearest")

# initialize the model
model = createModel()

# train the network
H = model.fit_generator(aug.flow(trainX, trainY, batch_size=BS), \
validation_data=(valX, valY), \
steps_per_epoch=len(trainX) // BS, samples_per_epoch=len(trainX) * 
5,epochs=EPOCHS, verbose=1)

# save the model to disk
model.save("test_new")

Answer 1

所以你可能想要做的是删除最后 2 层，它们对应于输出维度 22，并添加对应于新输出维度的 2 个新层（相同但 @ 的维度不同987654322@层）。

然后，如果您只想进行良好的初始化，您可以根据新数据重新调整您的模型。但是，如果您想“冻结”模型的权重并仅微调最后一层，则需要将模型的所有层设置为不可训练，然后重新编译模型：

# these lines will remove the last 2 layers
model.pop()
model.pop() 
# do the following 2 lines only if you want to keep the weights from the first training
for layer in model.layers:
  layer.trainable = False
model.add(Dense(output_dim=new_output_dim))
model.add(Activation("softmax"))
# do the following 2 lines only if you want to keep the weights from the first training
opt = Adam(lr=INIT_LR, decay=INIT_LR / EPOCHS)
model.compile(loss="categorical_crossentropy", optimizer=opt, metrics= 
["accuracy"])