尝试在ResNet50（notop）上添加Flatten图层并出现错误

Question

I am trying to add a Flatten layer, a Dense layer(relu) and a Dense layer(softmax) on ResNet50 for a multi-classification task using Keras 2.0.2 Theano 0.9.0 py2.7 on Win10.Here is my code:

def create_model():
    base_model = ResNet50(include_top=False, weights=None,
                            input_tensor=None, input_shape=(3,224,224),
                            pooling=None)

    base_model.load_weights(weight_path+'/resnet50_weights_th_dim_ordering_th_kernels_notop.h5')
    x = base_model.output
    x = Flatten()(x)
    x = Dense(128,activation='relu',kernel_initializer='random_normal',
            kernel_regularizer=regularizers.l2(0.1),
            activity_regularizer=regularizers.l2(0.1))(x)

    x=Dropout(0.3)(x)
    y = Dense(8, activation='softmax')(x)
    model = Model(base_model.input, y)
    for layer in base_model.layers:
        layer.trainable = False
    model.compile(optimizer='adadelta',
    loss='categorical_crossentropy')
    return model

I have set image_dim_ordering:

from keras import backend as K
K.set_image_dim_ordering('th')

And here is my Keras.json file:

{

"backend": "theano", ``"image_data_format": "channels_first", ``"epsilon": 1e-07, ``"floatx": "float32" }

Here is the error message:

ValueError: The shape of the input to "Flatten" is not fully defined (got (2048, None, None). Make sure to pass a complete "input_shape" or "batch_input_shape" argument to the first layer in your model.

Answer 1

You should

Pass an input_shape argument to the first layer. This is a shape tuple (a tuple of integers or None entries, where None indicates that any positive integer may be expected). In input_shape, the batch dimension is not included.

In your case, the first layer is Flatten() layer. It should be like

your_input = Input(shape=output_shape_of_resnet)
x = Flatten(your_input)

As for feeding the output of resnet50 into your own layers, consider defining a new model that incorporates your own layers and resnet, like

 new_model = Sequential()
 new_model.add(resnet_model) #Of course you need the definition and weights of resnet
 resnet_model.trainable = False #I guess?
 new_model.add(your_own_layers_model)

Answer 2

I had some error in case, when size of input image was too small for model of network. If the size of output data of layer become 0, this error appear. Your may use model.summary() to see how your network looks. This is example of model.summary() output:

Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_78 (Conv2D)           (None, 16, 21, 21)        160       
_________________________________________________________________
max_pooling2d_62 (MaxPooling (None, 16, 5, 5)          0         
_________________________________________________________________
...
flatten_25 (Flatten)         (None, 32)                0         
_________________________________________________________________
dense_28 (Dense)             (None, 2)                 1026      
=================================================================
Total params: 31,970
Trainable params: 31,970
Non-trainable params: 0
_________________________________________________________________