我正在尝试使用来自bvlc_alexnet.npy的预训练权重来实现AlexNet:
#load the weight data
weights_dic = numpy.load('bvlc_alexnet.npy', encoding='bytes').item()
conv1W = weights_dic["conv1"][0] # <class 'numpy.ndarray'> (11, 11, 3, 96)
conv1b = weights_dic["conv1"][1] # <class 'numpy.ndarray'> (96,)
model = Sequential()
model.add(Conv2D(96, kernel_size=[11, 11], kernel_initializer = <???>,
bias_initializer = <???>, dtype=np.ndarray), activation='relu', strides=4, padding="same")
在这里,我陷入了如何为conv1W
和conv1b
属性分配这些权重(kernel_initializer
和bias_initializer
)的问题。
答案 0 :(得分:2)
首先构造模型,而无需设置任何初始化程序。然后按照图层在模型中出现的顺序将所有权重放入列表中(例如conv1_weights,conv1_biases,conv2_weights,conv2_biases等),然后调用模型的set_weights
方法:
model.set_weights(weights)
或者,您可以分别设置每个图层的权重:
model.layers[layer_index].set_weights([layer_weights, layer_biases])
# or using layer's name if you have specified names for them
model.get_layer(layer_name).set_weights([layer_weights, layer_biases])