手写数字
from tensorflow.keras.layers import Conv2D,MaxPooling2D,Input,Conv2DTranspose,Flatten,Dense
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.models import Sequential,Model,load_model
import tensorflow as tf
import numpy as np
import cv2
IMG_H = 28
IMG_W = 28
IMG_C = 1
def tf_generate(x_train,y_train):
dataset = tf.data.Dataset.from_tensor_slices((x_train,y_train))
dataset = dataset.repeat()
dataset = dataset.shuffle(100).batch(1).repeat()
return dataset
def train_dataset(dataset):
latent_dim = IMG_H//2*IMG_W//2*IMG_C
input_img = Input(shape =(IMG_H,IMG_W,IMG_C))
x = Conv2D(32,3,padding='same',activation='relu')(input_img)
x = Conv2D(64,3,padding='same',activation='relu',strides=(2,2))(x)
x = Conv2D(64,3, padding='same',activation='relu')(x)
x = Conv2D(64,3,padding='same',activation='relu')(x)
x = Flatten()(x)
out = Dense(latent_dim)(x)
y = Flatten()(out)
y = tf.reshape(out,(1,IMG_H//2,IMG_W//2,IMG_C))
y = Conv2DTranspose(64,3,padding="same",use_bias=False)(y)
y = Conv2DTranspose(64,3,padding="same",use_bias=False)(y)
y = Conv2DTranspose(64,3,strides=(2,2),padding="same",use_bias=False)(y)
y = Conv2DTranspose(32,3,padding="same",use_bias=False)(y)
y = Conv2DTranspose(1,1,padding="same",use_bias=False)(y)
model = Model(input_img,y)
model.summary()
model.compile(loss="mse", optimizer="adam",metrics=["accuracy"])
history = model.fit(dataset, batch_size=1,epochs=20,steps_per_epoch=1000)
model.save("auto_mnist.h5")
def predict(test):
model = load_model("auto_mnist.h5")
cv2.imwrite("test.png",test)
x_test = cv2.resize(test,(IMG_H, IMG_W))/255.0
x_test = x_test.reshape(1,IMG_H,IMG_W,1)
result = model.predict(x_test)
print (result[0].shape)
result = result[0].reshape(IMG_H,IMG_W,1)
cv2.imwrite("result.png",result*255.0)
def main():
def train():
mnist=tf.keras.datasets.mnist
(x_train,y_train),(x_test,y_test)=mnist.load_data()
x_train = tf.expand_dims(x_train,axis=-1)/255
x_test = tf.expand_dims(x_test,axis=-1)
print(x_train.shape)
dataset = tf_generate(x_train,x_train)
train_dataset(dataset)
def test():
mnist=tf.keras.datasets.mnist
(x_train,y_train),(x_test,y_test)=mnist.load_data()
predict(x_test[0])
train()
#test()
main()
原图
模型输出
彩色图片
from tensorflow.keras.layers import Conv2D,MaxPooling2D,Input,Conv2DTranspose,Flatten,Dense
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.models import Sequential,Model,load_model
import tensorflow as tf
import numpy as np
import cv2
IMG_H = 512
IMG_W = 512
IMG_C = 3 ## Change this to 1 for grayscale.
def tf_dataset(image_path1,image_path2):
X_train = np.empty((len(image_path1), IMG_H, IMG_W, IMG_C))
Y_train = np.empty((len(image_path2), IMG_H, IMG_W, IMG_C))
for index,src in enumerate(image_path1):
img = cv2.imread(src)
image = cv2.resize(img,(IMG_H, IMG_W))/255.0
X_train[index] = image
for index,src in enumerate(image_path2):
img = cv2.imread(src)
image = cv2.resize(img,(IMG_H, IMG_W))/255.0
Y_train[index] = image
return X_train,Y_train
def tf_generate(x_train,y_train):
dataset = tf.data.Dataset.from_tensor_slices((x_train,y_train))
dataset = dataset.repeat()
dataset = dataset.shuffle(100).batch(1).repeat()
return dataset
def train_dataset(dataset):
latent_dim = 4*4*512
input_img = Input(shape =(224,224,3))
x = Conv2D(32,3,padding='same',activation='relu')(input_img)
print (x.shape)
x = Conv2D(64,3,padding='same',activation='relu',strides=(2,2))(x)
print (x.shape)
x = x+Conv2D(64,3, padding='same',activation='relu',strides=(1,1))(x)
print (x.shape)
x = Conv2D(128,3, padding='same',activation='relu',strides=(2,2))(x)
print (x.shape)
x = x+Conv2D(128,3, padding='same',activation='relu')(x)
print (x.shape)
x = Conv2D(256,3, padding='same',activation='relu',strides=(2,2))(x)
print (x.shape)
x = x+Conv2D(256,3, padding='same',activation='relu')(x)
print (x.shape)
x = Conv2D(512,3, padding='same',activation='relu',strides=(2,2))(x)
print (x.shape)
x = x+Conv2D(512,3, padding='same',activation='relu')(x)
print (x.shape)
y = x+Conv2DTranspose(512,3,padding="same",use_bias=False)(x)
print (y.shape)
y = Conv2DTranspose(256,3,padding="same",strides=(2,2),use_bias=False)(y)
print (y.shape)
y = y+Conv2DTranspose(256,3,padding="same",use_bias=False)(y)
print (y.shape)
y = Conv2DTranspose(128,3,padding="same",strides=(2,2),use_bias=False)(y)
print (y.shape)
y = y+Conv2DTranspose(128,3,padding="same",use_bias=False)(y)
print (y.shape)
y = Conv2DTranspose(64,3,padding="same",strides=(2,2),use_bias=False)(y)
print (y.shape)
y = y+Conv2DTranspose(64,3,padding="same",use_bias=False)(y)
print (y.shape)
y = Conv2DTranspose(32,3,padding="same",strides=(2,2),use_bias=False)(y)
print (y.shape)
out = Conv2DTranspose(3,3,padding="same",use_bias=False)(y)
print (y.shape)
model = Model(input_img,out)
model.summary()
model.compile(loss="mse", optimizer="adam",metrics=["accuracy"])
history = model.fit(dataset, batch_size=1,epochs=4,steps_per_epoch=100)
model.save("auto_new_style_a_b.h5")
def predict():
src = "./0.png"
model = load_model("auto_style_a_b.h5")
img = cv2.imread(src)
cv2.imwrite("test.png",img)
x_test = cv2.resize(img,(IMG_H, IMG_W))/255.0
x_test = x_test.reshape(1,IMG_H,IMG_W,3)
result = model.predict(x_test)
print (result[0].shape)
result = result[0].reshape(IMG_H,IMG_W,3)
cv2.imwrite("output5.png",result*255.0)
def main():
def train():
images_path1 = glob("./style/a/*.png")
images_path2 = glob("./style/b/*.png")
count_dataset = len(images_path1),len(images_path2)
print (count_dataset)
x_train,y_train = tf_dataset(images_path1,images_path2)
dataset = tf_generate(x_train,x_train)
train_dataset(dataset)
#train()
predict()
main()
原图
模型输出
说明：本想从真人到动漫图像转换输出，结果不好。
图片较小（64*64），转换结果不错，越大，结果越差。
不同图片大小转换结果：（真人-动漫，动漫-真人）
