합성곱 계층 출력 분석

import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import matplotlib.pyplot as plt

tf.config.set_visible_devices([], 'GPU')

model = keras.applications.vgg16.VGG16(weights='imagenet', include_top=True)
model.summary()

img = keras.preprocessing.image.load_img('../data/mozzi.jpg', target_size=(224, 224))
img

# collect the outputs of the intermediate layers
layer_outputs = [layer.output for layer in model.layers]
# create a new model with the intermediate layers as outputs
activation_model = keras.Model(inputs=model.input, outputs=layer_outputs)

# preprocess the image
x1 = keras.preprocessing.image.img_to_array(img)
x1 = keras.applications.vgg16.preprocess_input(x1)
# convert the image to a numpy array
X = np.array([x1])
# get the activations
activations = activation_model.predict(X)

# print out the output shapes with the layer names
for layer, activation in zip(model.layers, activations):
    print(layer.name, activation.shape)

# normalize the image
def normalize_image(img):
    if img.sum() == 0:
        return img

    # image values are standardized to have zero mean and unit variance
    img -= img.mean()
    img /= img.std()
    # amplify the image to make it more visible
    img *= 64
    img += 128
    img = np.clip(img, 0, 255).astype('uint8')
    return img

# visualize a layer output
def visualize_layer(activation, images_per_row=16, output_file=None):
    # plot all outputs
    output_size = activation.shape[-1]
    # calculate the number of columns and rows
    cols = output_size // images_per_row
    fig, subplots = plt.subplots(cols, images_per_row, figsize=(images_per_row, cols))
    plt.setp(subplots, xticks=[], yticks=[])
    for i in range(output_size):            
        channel_image = activation[0, :, :, i].copy()
        if channel_image.sum() != 0:
            channel_image = normalize_image(channel_image)
        # plot the channel
        subplots.ravel()[i].imshow(channel_image, cmap='viridis')
    
    if output_file:
        fig.savefig(output_file)
        # close the figure to free up memory
        plt.close(fig)
    else:
        plt.show()

# visualize all layers
for layer, activation in zip(model.layers, activations):
    # only visualize convolutional layers
    if 'conv' in layer.name:
        print(layer.name, activation.shape)
        visualize_layer(activation, output_file=f'../outputs/{layer.name}.png')