kares提取中层特征及其可视化

kares代码如下：（哈哈，我的代码有点繁琐（乱），部分函数就不具体展开啦）

1. input = Input(shape=(-1，15, 15, 200, 1))
2. conv1 = _conv_bn_relu(nb_filter=64, kernel_dim1=3, kernel_dim2=3, kernel_dim3=input._keras_shape[3], subsample=(1, 1, 1))(input) 
3. spa1_output = _residual_block1(nb_filter=nb_filter, repetitions=1, is_first_layer=True, name = 'spa1')(conv1) 
4. spa2_output = _residual_block2(nb_filter=nb_filter, repetitions=1, is_first_layer=True, name = 'spa2')(spa1_output)
5. spa3_output = _residual_block3(nb_filter=nb_filter, repetitions=1, is_first_layer=True, name = 'spa3')(spa2_output)
6. block = _bn_relu(spa3_output)
7. block_norm = BatchNormalization(axis=CHANNEL_AXIS)(block)
8. block_output = Activation("relu")(block_norm)
9. pool = AveragePooling3D(pool_size=(block._keras_shape[CONV_DIM1],                     block._keras_shape[CONV_DIM2],block._keras_shape[CONV_DIM3],), strides=(1, 1, 1))(block_output)
10. flatten1 = Flatten()(pool)
11. drop1 = Dropout(0.5)(flatten1)
12. dense = Dense(units=num_outputs, activation="softmax", kernel_initializer="he_normal")(drop1)
13. model1 = Model(inputs=input, outputs=dense)
14. model1.compile(loss='categorical_crossentropy', optimizer=RMSprop(lr=0.0003), metrics=['accuracy'])
15. best_weights_RES_path_ss4 = 'models/Indian_best_1.hdf5'
16. earlyStopping6 = kcallbacks.EarlyStopping(monitor='val_loss', patience=patience, verbose=1, mode='auto')
17. saveBestModel6 = kcallbacks.ModelCheckpoint(best_weights_RES_path_ss4, monitor='val_loss', verbose=1,save_best_only=True, mode='auto')
18. model1.fit(x= x_train2.reshape(x_train2.shape[0], x_train2.shape[1], x_train2.shape[2], x_train2.shape[3], 1),                      y= y_train, validation_data=([x_val1.reshape(x_val1.shape[0], x_val1.shape[1], x_val1.shape[2], x_val1.shape[3], 1),         x_val2.reshape(x_val2.shape[0], x_val2.shape[1], x_val2.shape[2], x_val2.shape[3], 1)], y_val),     batch_size=batch_size, nb_epoch=nb_epoch, shuffle=True, callbacks=[earlyStopping6, saveBestModel6])

下载模型及提取中间3层的特征：

1. model1_out = load_model('models/Indian_best_1.hdf5')
2. spa1_model = Model(inputs=model1_out.inputs, outputs=model1_out.get_layer('spa1').output)
3. spa2_model = Model(inputs=model1_out.inputs, outputs=model1_out.get_layer('spa2').output)
4. spa3_model = Model(inputs=model1_out.inputs, outputs=model1_out.get_layer('spa3').output)
5. spa1_output = spa1_model.predict(x_test2[1].reshape(1, x_test2[1].shape[0], x_test2[1].shape[1], x_test2[1].shape[2], 1))
6. spa2_output = spa2_model.predict(x_test2[1].reshape(1, x_test2[1].shape[0], x_test2[1].shape[1], x_test2[1].shape[2], 1))
7. spa3_output = spa3_model.predict(x_test2[1].reshape(1, x_test2[1].shape[0], x_test2[1].shape[1], x_test2[1].shape[2], 1))

中间3层的可视化，每层展示3个feature map：

1. plt.figure(figsize=(3, 3))
2. for i in range(3):
3.       plt.subplot(3, 6, 6 * i + 1)
4.       spa1_img = spa1_output[0, :, :, 0, i]
5.       spa1_img = (spa1_img - np.min(spa1_img)) / (np.max(spa1_img) - np.min(spa1_img))
6.       plt.title("spa1")
7.       plt.imshow(spa1_img)
8.       plt.subplot(3, 6, 6 * i + 2)
9.       spa2_img = spa2_output[0, :, :, 0, i]
10.       spa2_img = (spa2_img - np.min(spa2_img)) / (np.max(spa2_img) - np.min(spa2_img))
11.       plt.title("spa2")
12.       plt.imshow(spa2_img)
13.       plt.subplot(3, 6, 6 * i + 3)
14.       spa3_img = spa3_output[0, :, :, 0, i]
15.       spa3_img = (spa3_img - np.min(spa3_img)) / (np.max(spa3_img) - np.min(spa3_img))
16.       plt.title("spa3")
17.       plt.imshow(spa3_img)
18.   plt.show()