无法将chexnet预训练权重文件加载到Densenet121

Question

我正在尝试将Keras chexNet权重文件加载到Densenet121，https://www.kaggle.com/theewok/chexnet-keras-weights

我正在获取ValueError:您正在尝试将包含242层的权重文件加载到具有241层的模型中。如果我调用densenet121

densenet = tf.keras.applications.DenseNet121(
include_top=False,
weights="CheXNet_Keras_0.3.0_weights.h5",
input_shape=(224,224,3)
)

如果我试一下：-

densenet = tf.keras.applications.DenseNet121(
include_top=True,
weights="CheXNet_Keras_0.3.0_weights.h5",
input_shape=(224,224,3)
)

我将获取ValueError: Shape (1024,1000)和(1024，14)不兼容

Answer 1

他们在没有正确的输出层的情况下保存了模型，修复方法如下：

base_model = densenet.DenseNet121(weights=None,
                            include_top=False,
                            input_shape=(224,224,3), pooling="avg")

predictions = tf.keras.layers.Dense(14, activation='sigmoid', name='predictions')(base_model.output)
base_model = tf.keras.Model(inputs=base_model.input, outputs=predictions)
base_model.load_weights("./temp/CheXNet_Keras_0.3.0_weights.h5")
base_model.layers.pop()

print("CheXNet loaded")

Answer 2

弹出最后一层的答案现在不再有效，pop只返回最后一层，但模型保持不变。

我推荐这样的东西：

densenet = DenseNet121(weights=None, include_top=False, 
                       input_shape=(224, 224, 3), pooling="avg")
output = tf.keras.layers.Dense(14, activation='sigmoid', name='output')(densenet.layers[-1].output)
model = tf.keras.Model(inputs=[densenet.input], outputs=[output])
model.load_weights("./CheXNet_weights.h5")