python - Keras:使用 fit_generator 时出现 notImplementedError/RuntimeError

Question

我在使用 keras 和 tensorflow 时遇到问题，使用以下代码:

from tensorflow.keras.layers import Activation, Conv2D
from tensorflow.keras import Model

from data import DataGenerator
from config import train_datapath, test_datapath


training_generator = DataGenerator(train_datapath)
validation_generator = DataGenerator(test_datapath)

class model(Model):
    def __init__(self):
        super(model, self).__init__()
        self.conv1 = Conv2D(filters=2, kernel_size=1, strides=1, padding='same', input_shape=(256, 256, 1))
        self.act1 = Activation('relu')

    def call(self, input):
        """Run the model."""
        return self.act1(self.conv1(input))

model = model()
model.compile(optimizer='adam', loss='mean_squared_error')
history = model.fit_generator(training_generator, epochs=5000, verbose=1, validation_data=(validation_generator),
                              use_multiprocessing=False)

运行该代码会出现以下错误:

Using TensorFlow backend.
Traceback (most recent call last):
  File "C:/Users/...py", line 23, in <module>
    use_multiprocessing=False)
  File "C:\Users\...\AppData\Local\Programs\Python\Python36\lib\site-packages\tensorflow\python\keras\engine\training.py", line 2161, in fit_generator
    '`fit_generator` is not yet enabled for unbuilt Model subclasses')
NotImplementedError: `fit_generator` is not yet enabled for unbuilt Model subclasses

我试图找到一些提示，如果我真的尝试做某事，那不是故意的，但对我来说这似乎不合逻辑，因为这种方式对于定义更复杂的网络来说非常舒服，而且我已经准备好了与 pytorch 一起工作，我很确定，这也应该与 tf 和 keras 一起工作。

有一个thread有同样的错误，但对于顺序网络实现，这显然不是我的目标。

如果我对 keras 使用直接导入，我会收到不同的错误:

Using TensorFlow backend.
Traceback (most recent call last):
  File "C:/Users/...py", line 23, in <module>
    use_multiprocessing=False)
  File "C:\Users\...\AppData\Local\Programs\Python\Python36\lib\site-packages\keras\legacy\interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "C:\Users\...\AppData\Local\Programs\Python\Python36\lib\site-packages\keras\engine\training.py", line 1418, in fit_generator
    initial_epoch=initial_epoch)
  File "C:\Users\...\AppData\Local\Programs\Python\Python36\lib\site-packages\keras\engine\training_generator.py", line 40, in fit_generator
    model._make_train_function()
  File "C:\Users\...\AppData\Local\Programs\Python\Python36\lib\site-packages\keras\engine\training.py", line 496, in _make_train_function
    raise RuntimeError('You must compile your model before using it.')
RuntimeError: You must compile your model before using it.

Process finished with exit code 1

但是我正在编译上面一行的模型，调用 model.fit_generator() 函数...当我没有完全错的时候，我也给出了正确的输入形状，所以这也不应该是问题......

使用 windows、python 3.6、tensorflow-gpu 1.12。

Answer 1

您不是在任何地方创建模型。

你需要:

model = Model(inputTensors, outputTensors)

或者至少在你的类中的某个时刻super(model,self).__init__(inputTensors,outputTensors)。

理想情况下:

def createModel():
    inputs = Input((256,256,1))
    outputs = Conv2D(filters=2, kernel_size=1, strides=1, 
                     padding='same', activation='relu')(inputs)
    return Model(inputs,outputs)

如果出于某种原因你确实想要一个子类:

class MyModel(Model):
    def init(self):
        inputs = Input((256,256,1))
        self.conv1 = Conv2D(filters=2, kernel_size=1, strides=1, padding='same')
        self.act1 = Activation('relu')
        outputs = self.act1(self.conv1(inputs))

        super(MyModel,self).__init__(inputs,outputs)