尝试创建GAN：InvalidArgumentError：矩阵大小不兼容

Question

I am quite new to the field but I am trying to create a Generative Adversarial Network to generate Music. I have a model that is a combination of the Generator and Discriminator but when I train it, it gives me an error. Something about the output it doesn't like. I am using Keras sequential. Any help would be much appreciated.

As I understand it the input and output should have the same dimensions in any Keras model. My input shape - (300, 30, 1). Output shape - (300, 1). And when I train them separately they don't cause an error. But when I combine them in a separate model they start to give an error - particularly in the discriminator's last line -> Dense(1, activation='sigmoid')

    def __generator(self):
        """ Declare generator """
        model = Sequential()
        model.add(LSTM(256, input_shape=(self.n_prev, 1), return_sequences=True))
        model.add(Dropout(0.6))
        model.add(LSTM(128, input_shape=(self.n_prev, 1), return_sequences=True))
        model.add(Dropout(0.6))
        model.add(LSTM(64, input_shape=(self.n_prev, 1), return_sequences=False))
        model.add(Dropout(0.6))
        model.add(Dense(1))

        print(model.summary())
        return model


    def __discriminator1b (self, width=300, height=30, channels=1):

        shape = (width, height, channels)
        model = Sequential()
        model.add(Flatten(input_shape=((30, 1))))
        model.add(Dense((height * channels), input_shape=(30, 1)))
        model.add(LeakyReLU(alpha=0.2))
        model.add(Dense(np.int64((height * channels)/2)))
        model.add(LeakyReLU(alpha=0.2))

        model.add(Dense(1, activation='sigmoid'))   

        print(model.summary())
        return model


    def __gen_disc (self):
        model = Sequential()
        model.add(self.G)
        model.add(self.D)

        return model

Training:
self.G.train_on_batch(np.array(gen_noiseX), np.array(genY))
self.D.train_on_batch(np.array(gen_noiseX), disc_label)
self.GD.train_on_batch(np.array(gen_noiseX), y_mislabled)

Model Summaries:
Generator:
Layer (type)                 Output Shape              Param #   
=================================================================
lstm_28 (LSTM)               (None, 30, 256)           264192    
_________________________________________________________________
dropout_52 (Dropout)         (None, 30, 256)           0         
_________________________________________________________________
lstm_29 (LSTM)               (None, 30, 128)           197120    
_________________________________________________________________
dropout_53 (Dropout)         (None, 30, 128)           0         
_________________________________________________________________
lstm_30 (LSTM)               (None, 64)                49408     
_________________________________________________________________
dropout_54 (Dropout)         (None, 64)                0         
_________________________________________________________________
dense_36 (Dense)             (None, 1)                 65        
=================================================================
Total params: 510,785
Trainable params: 510,785
Non-trainable params: 0
_________________________________________________________________
None


Discriminator:

Layer (type)                 Output Shape              Param #   
=================================================================
flatten_8 (Flatten)          (None, 30)                0         
_________________________________________________________________
dense_37 (Dense)             (None, 30)                930       
_________________________________________________________________
leaky_re_lu_15 (LeakyReLU)   (None, 30)                0         
_________________________________________________________________
dense_38 (Dense)             (None, 15)                465       
_________________________________________________________________
leaky_re_lu_16 (LeakyReLU)   (None, 15)                0         
_________________________________________________________________
dense_39 (Dense)             (None, 1)                 16        
=================================================================
Total params: 1,411
Trainable params: 1,411
Non-trainable params: 0
_________________________________________________________________
None

So the error itself is

InvalidArgumentError: Matrix size-incompatible: In[0]: [300,1], In[1]: [30,30]   [[{{node sequential_22/dense_37/MatMul}}]]

Whenever I remove the Dense(1, sigmoid) layer of the discriminator it works but I need that layer for binary classification. Maybe I need to rebuild the model or just do a small fix but anyway all suggestions are appreciated.

Answer 1

Welcome to stackoverflow. This error occurs because, your model needs (30,30) but you are feeding it (300,1) .
There are a few changes that would be better here:

1. In discriminator model.add(Flatten(input_shape=((30, 1)))) is at wrong place. It should be below before dense layer. Or as you are building a RNN, I would argue you don't need a Flatten layer.
2. Use bidirectional LSTMs .
3. In generator use batchnormalization .

There are a few other changes needed for the whole network, you can see this great article for music generation . Hope this helps!!