在生成图片时，神经网络是冻结的，也就是网络的权重不再更新，只需要更新输入的图片。常用的预训练卷积网络包括Google的Inception、VGG网络和ResNet网络等。
DeePDream的基本步骤：
- 获取输入图片
- 将图片输入网络，得到所希望可视化的神经元的输出值
- 计算神经元输出值对图片各像素的梯度
- 使用梯度下降不断更新图片
重复第2、3、4步，直到满足所设定的条件
下面是使用Keras实现DeepDream的大致过程：
用Keras实现DeepDream
获取测试图片
In [1]:
# ---------------
from tensorflow import keras
import matplotlib.pyplot as plt
%matplotlib inline
base_image_path = keras.utils.get_file(
"coast.jpg",
origin="https://www.528045.com/file/upload/202411/30/kcits24l41e.jpg")
plt.axis("off")
plt.imshow(keras.utils.load_img(base_image_path))
plt.show()
上面是Keras自带的一张海岸线的图片。下面就是对这张图进行变化。
准备预训练模型InceptionV3
In [2]:
# 使用Inception V3实现
from keras.applications import inception_v3
# 使用预训练的ImageNet权重来加载模型
model = inception_v3.InceptionV3(weights="imagenet", # 构建不包含全连接层的Inceptino
include_top=False)
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/inception_v3/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5
87916544/87910968 [==============================] - 74s 1us/step
87924736/87910968 [==============================] - 74s 1us/step
In [3]:
model.summary()
设置DeepDream配置
In [4]:
# 层的名称 + 系数：该层对需要最大化的损失的贡献大小
layer_settings = {"mixed4":1.0,
"mixed5":1.5,
"mixed6":2.0,
"mixed7":2.5}
outputs_dict = dict(
[
(layer.name, layer.output) # 层的名字 + 该层的输出
for layer in [model.get_layer(name) for name in layer_settings.keys()]
]
)
outputs_dict
Out[4]:
{'mixed4': <KerasTensor: shape=(None, None, None, 768) dtype=float32 (created by layer 'mixed4')>,
'mixed5': <KerasTensor: shape=(None, None, None, 768) dtype=float32 (created by layer 'mixed5')>,
'mixed6': <KerasTensor: shape=(None, None, None, 768) dtype=float32 (created by layer 'mixed6')>,
'mixed7': <KerasTensor: shape=(None, None, None, 768) dtype=float32 (created by layer 'mixed7')>}
In [5]:
# 特征提取
feature_extractor = keras.Model(inputs=model.inputs, outputs=outputs_dict)
feature_extractor
Out[5]:
<keras.engine.functional.Functional at 0x15b5ff0d0>
计算损失
In [6]:
def compute_loss(image):
features = feature_extractor(image) # 特征提取
loss = tf.zeros(shape=()) # 损失初始化
for name in features.keys(): # 遍历层
coeff = layer_settings[name] # 某个层的系数
activation = features[name] # 某个层的激活函数
#为了避免出现边界伪影，损失中仅包含非边界的像素
loss += coeff * tf.reduce_mean(tf.square(activation[:, 2:-2, 2:-2, :])) # 将该层的L2范数添加到loss中；
return loss
梯度上升过程
In [7]:
import tensorflow as tf
@tf.function
def gradient_ascent_step(image, lr): # lr--->learning_rate 学习率
with tf.GradientTape() as tape:
tape.watch(image)
loss = compute_loss(image) # 调用计算损失方法
grads = tape.gradient(loss, image) # 梯度更新
grads = tf.math.l2_normalize(grads)
image += lr * grads
return loss, image
def gradient_ascent_loop(image, iterations, lr, max_loss=None):
for i in range(iterations):
loss, image = gradient_ascent_step(image, lr)
if max_loss is not None and loss > max_loss:
break
print(f"第{i}步的损失值是{loss:.2f}")
return image
图片生成
np.expand_dims用法（个人添加）
In [8]:
import numpy as np
array = np.array([[1,2,3],
[4,5,6]]
)
array
Out[8]:
array([[1, 2, 3],
[4, 5, 6]])
In [9]:
array.shapeOut[9]:
(2, 3)
In [10]:
array1 = np.expand_dims(array,axis=0)
array1
Out[10]:
array([[[1, 2, 3],
[4, 5, 6]]])
In [11]:
array1.shapeOut[11]:
(1, 2, 3)
In [12]:
array2 = np.expand_dims(array,axis=1)
array2
Out[12]:
array([[[1, 2, 3]],
[[4, 5, 6]]])
In [13]:
array2.shapeOut[13]:
(2, 1, 3)
In [14]:
array3 = np.expand_dims(array,axis=-1)
array3
Out[14]:
array([[[1],
[2],
[3]],
[[4],
[5],
[6]]])
In [15]:
array3.shapeOut[15]:
(2, 3, 1)
np.clip功能（个人添加）
np.clip(
array,
min(array),
max(array),
out=None):
In [16]:
array = np.array([1,2,3,4,5,6])
np.clip(array, 2, 5) # 输出长度和原数组相同
Out[16]:
array([2, 2, 3, 4, 5, 5])
In [17]:
array = np.arange(18).reshape((6,3))
array
Out[17]:
array([[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8],
[ 9, 10, 11],
[12, 13, 14],
[15, 16, 17]])
In [18]:
np.clip(array, 5, 15)
Out[18]:
array([[ 5, 5, 5],
[ 5, 5, 5],
[ 6, 7, 8],
[ 9, 10, 11],
[12, 13, 14],
[15, 15, 15]])
参数设置
In [19]:
step = 20. # 梯度上升的步长
num_octave = 3 # 运行梯度上升的尺度个数
octave_scale = 1.4 # 两个尺度间的比例大小
iterations = 30 # 在每个尺度上运行梯度上升的步数
max_loss = 15. # 损失值若大于15，则中断梯度上升过程
图片预处理
In [20]:
import numpy as np
def preprocess_image(image_path): # 预处理
img = keras.utils.load_img(image_path) # 导入图片
img = keras.utils.img_to_array(img) # 转成数组
img = np.expand_dims(img, axis=0) # 增加数组维度；见上面解释(x,y) ---->(1,x,y)
img = keras.applications.inception_v3.preprocess_input(img)
return img
def deprocess_image(img): # 图片压缩处理
img = img.reshape((img.shape[1], img.shape[2], 3))
img /= 2.0
img += 0.5
img *= 255.
# np.clip：截断功能，保证数组中的取值在0-255之间
img = np.clip(img, 0, 255).astype("uint8")
return img
生成图片
In [21]:
# step = 20. # 梯度上升的步长
# num_octave = 3 # 运行梯度上升的尺度个数
# octave_scale = 1.4 # 两个尺度间的比例大小
# iterations = 30 # 在每个尺度上运行梯度上升的步数
# max_loss = 15.0 # 损失值若大于15，则中断梯度上升过程
original_img = preprocess_image(base_image_path) # 预处理函数
original_shape = original_img.shape[1:3]
print(original_img.shape) # 四维图像
print(original_shape) # 第2和3维度的值
(1, 900, 1200, 3)
(900, 1200)
In [22]:
successive_shapes = [original_shape]
for i in range(1, num_octave):
shape = tuple([int(dim / (octave_scale ** i)) for dim in original_shape])
successive_shapes.append(shape)
successive_shapes = successive_shapes[::-1] # 翻转
shrunk_original_img = tf.image.resize(original_img, successive_shapes[0])
img = tf.identity(original_img)
for i, shape in enumerate(successive_shapes):
print(f"Processing octave {i} with shape {shape}")
# resize
img = tf.image.resize(img, shape)
img = gradient_ascent_loop( # 梯度上升函数调用
img,
iteratinotallow=iterations,
lr=step,
max_loss=max_loss
)
# resize
upscaled_shrunk_original_img = tf.image.resize(shrunk_original_img, shape)
same_size_original = tf.image.resize(original_img, shape)
lost_detail = same_size_original - upscaled_shrunk_original_img
img += lost_detail
shrunk_original_img = tf.image.resize(original_img, shape)
keras.utils.save_img("dream.png", deprocess_image(img.numpy()))
结果为：
Processing octave 0 with shape (459, 612)
第0步的损失值是0.80
第1步的损失值是1.07
第2步的损失值是1.44
第3步的损失值是1.82
......
第26步的损失值是11.44
第27步的损失值是11.72
第28步的损失值是12.03
第29步的损失值是12.49
同时在本地生成了新图片，看下效果：
再看一眼原图：相对比之下，新图有点梦幻的味道！
