备忘录||深度学习||输出模型特征图

导入相关包

import numpy as np
from keras.models import load_model
import matplotlib.pyplot as plt

import numpy as np
import matplotlib.pyplot as plt
from keras.preprocessing import image
from keras import layers
from keras import models
import math

载入模型

work_dir = 'D:/py/SAR'

# 载入模型
def read_model():
    model = load_model(work_dir + '/model/model_s5_nofilt.h5')
    return model
model = read_model()

输出模型每层的名字

layer_names = [layer.name for layer in model.layers]
print(layer_names)

输出特征图

def visible():
    model = load_model(work_dir + '/model/model_s5_nofilt.h5')

    layer_outputs = [layer.output for layer in model.layers[:5]]
    activation_model = models.Model(inputs=model.input, outputs=layer_outputs)

    img_path = 'D:/py/SAR/mstar/train/2S1/HB14931.jpg'
    img = image.load_img(img_path, target_size=(100, 100))
    img_tensor = image.img_to_array(img)
    img_tensor = img_tensor / 255.0
    img_tensor = np.expand_dims(img_tensor, axis=0)

    activations = activation_model.predict(img_tensor)
    for k in range(len(activations)):
        first_layer_activation = activations[k]
        print(first_layer_activation.shape)


        h = first_layer_activation.shape[1]
        w = first_layer_activation.shape[2]
        #num_channels = first_layer_activation.shape[3]
        num_channels = 6                      #若num_channels等于first_layer_activation.shape[3]，则相当于将全部通道特征图输出
        cols = 6
        rows = math.ceil(num_channels/cols)
        img_grid = np.zeros(((h*rows,w*cols)))
                for i in range(num_channels):
            channel_image = first_layer_activation[0, :, :, i]
            # col列。
            f_r = math.ceil((i+1)/cols)
            f_c = (i+1)if f_r==1 else ((i+1)-(f_r-1)*cols)
            '''
            对图像进行增强等操作，便于人眼观察
            channel_image -= channel_image.mean()
            channel_image /= channel_image.std()
            channel_image *= 64
            channel_image += 128
            channel_image = np.clip(channel_image, 0, 255).astype('uint8')
            '''
            img_grid[(f_r-1)*h:f_r*h,(f_c-1)*h:f_c*h]=channel_image
            #plt.xticks([])
            #plt.yticks([])
        # plt.savefig(str(k) + ".jpg", dpi=100)
        scale=20. / h
        plt.figure(figsize=(scale*h,scale))
        plt.imshow(img_grid,aspect='equal',cmap='viridis')
        plt.xticks([])
        plt.yticks([])
        plt.grid(False)
        plt.savefig(str(k) + "@.jpg", dpi=300,bbox_inches='tight')
    plt.show()