【飞桨模型复现计划】SRCNN网络-超分辨率重建
项目简介
本项目是paperweekly paddlepaddle复现活动的第23篇论文《Single Image Super-Resolution Using Deep Learning》的复现代码。 论文主要介绍了图像超分辨率重建的DNN方法,SRCNN网络。本代码复现了基础模型、修改网络大小的结果和3通道RGB模型的结果，并展示经过超分辨率重建后的图片。
In[1]
# 查看当前挂载的数据集目录
!ls /home/aistudio/data/data863 data904
In[2]
# 查看个人持久化工作区文件
!rm -rf /home/aistudio/work/*
!ls /home/aistudio/work/解压数据集
In[3]
# 这个cell的代码只有第一次运行时需要执行
import os
os.mkdir('/home/aistudio/work/model') # 用于存储训练好的模型文件
os.mkdir('/home/aistudio/work/model_3d') # 用于存储训练好的3d模型文件
os.mkdir('/home/aistudio/work/model_fz') # 用于存储训练好的fz模型文件
os.mkdir('/home/aistudio/work/dataset') # 用于存放所有的数据集
os.mkdir('/home/aistudio/work/dataset/timofte') # 91图训练集
# 解压训练集图片
os.chdir('/home/aistudio/work/dataset/timofte')
os.system('tar -xvf /home/aistudio/data/data863/input_images.tar')
# 解压测试集图片
os.chdir('/home/aistudio/work/dataset')
os.system('tar -xvf /home/aistudio/data/data904/val_dataset.tar')
# 回到工作路径
os.chdir('/home/aistudio')In[4]
# import 要用的库
import paddle.fluid as fluid
import paddle.v2 as paddle
import numpy as np
import cv2
import os
from matplotlib import pyplot as plt
%matplotlib inlinebaseline model
In[6]
class SRCNN(object):
# 只对Y通道做超分辨率重构
def __init__(self, lr, lr_f, batch_size, iter_num):
self.lr = lr # 学习率
self.lr_f = lr_f # 最后一层学习率
self.batch_size = batch_size
self.iter_num = iter_num # 总共训练多少次
def net(self, X, Y):
# 搭建模型
conv1 = fluid.layers.conv2d(X, 64, 9,act='relu', name='conv1' ,
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='conv1_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='conv1_b'))
conv2 = fluid.layers.conv2d(conv1, 32, 1, act='relu', name='conv2' ,
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='conv2_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='conv2_b'))
pred = fluid.layers.conv2d(conv2, 1, 5, name='pred',
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='pred_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='pred_b'))
loss = fluid.layers.reduce_mean(fluid.layers.square(pred - Y))
return pred, loss
def train(self):
# 模型训练
X_train = fluid.layers.data(shape=[1, 33, 33], dtype='float32', name='image')
Y_train = fluid.layers.data(shape=[1, 21, 21], dtype='float32', name='gdt')
y_predict, y_loss = self.net(X_train, Y_train)
Optimizer = fluid.optimizer.AdamOptimizer(learning_rate=self.lr)
Optimizer_f = fluid.optimizer.AdamOptimizer(learning_rate=self.lr_f)
Optimizer.minimize(y_loss, parameter_list=['conv1_w','conv1_b', 'conv2_w', 'conv2_b'])
Optimizer_f.minimize(y_loss, parameter_list=['pred_w', 'pred_b'])
# 读取训练集数据
train_reader = paddle.batch(self.read_data('work/dataset/timofte'), batch_size=self.batch_size)
# 定义执行器
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
def train_loop(main_program):
feeder = fluid.DataFeeder(place=place, feed_list=[X_train, Y_train])
exe.run(fluid.default_startup_program())
backprops_cnt = 0 # 论文中作图的横坐标
self.backprops_cnts = [] # 绑定为类的一个属性，用于画图
self.psnr = [] # psnr的值
for epoch in range(self.iter_num):
for batch_id, data in enumerate(train_reader()):
loss = exe.run(
fluid.framework.default_main_program(),
feed=feeder.feed(data),
fetch_list=[y_loss])
if batch_id == 0: # 每个epoch算一下psnr，画图用的
# # 算psnr要在测试集上面
fluid.io.save_inference_model('work/model/', ['image'], [y_predict], exe)
val_loss, val_psnr = self.validation()
self.backprops_cnts.append(backprops_cnt * self.batch_size)
self.psnr.append(val_psnr)
print("%i\tEpoch: %d \tCur Cost : %f\t Val Cost: %f\t PSNR :%f" % (backprops_cnt, epoch, np.array(loss[0])[0], val_loss, val_psnr))
backprops_cnt += 1
fluid.io.save_inference_model('work/model/', ['image'], [y_predict], exe)
train_loop(fluid.default_main_program())
def validation(self):
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
test_set = 'work/dataset/set5/'
scale_factor = 3
for img_name in os.listdir(test_set):
img_val = cv2.imread(os.path.join(test_set, img_name))
yuv = cv2.cvtColor(img_val, cv2.COLOR_BGR2YCrCb)
img_y, img_u, img_v = cv2.split(yuv)
img_h, img_w = img_y.shape
img_blur = cv2.GaussianBlur(img_y, (5, 5), 0)
img_subsample = cv2.resize(img_blur, (img_w/scale_factor, img_h/scale_factor))
img_input = cv2.resize(img_blur, (img_w, img_h), interpolation=cv2.INTER_CUBIC)
img_input = np.reshape(img_input, [1,1,img_h, img_w]).astype("float32") # h,w
losses = []
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model('work/model/', exe))
results = exe.run(inference_program,
feed={feed_target_names[0]: img_input},
fetch_list=fetch_targets)[0]
loss = np.mean(np.square(results[0,0]-img_y[6:-6, 6:-6]))
losses.append(loss)
avg_loss = np.sum(np.array(losses))/len(losses)
psnr = 10 * np.log10(255*255/avg_loss)
return avg_loss,psnr
def generate_reconstruct_img(self, img_name):
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
img_test = cv2.imread('work/dataset/set5/%s' % img_name)
yuv_test = cv2.cvtColor(img_test, cv2.COLOR_BGR2YCrCb)
img_h, img_w, img_c = img_test.shape
print "=====原始图片========="
b,g,r = cv2.split(img_test) # AI studio 不支持cv2.imshow，所以用plt.imshow输出，两者rgb顺序不一样
img_test = cv2.merge([r,g,b])
plt.imshow(img_test)
plt.show()
# 图像模糊+cubic插值
img_blur = cv2.GaussianBlur(yuv_test.copy(), (5, 5), 0)
img_subsample = cv2.resize(img_blur, (img_w/3, img_h/3)) #这里注意cv2.resize里面的shape是w,h的顺序
img_cubic = cv2.resize(img_blur, (img_w, img_h), interpolation=cv2.INTER_CUBIC)
img_y, img_u, img_v = cv2.split(img_cubic)
img_input = np.reshape(img_y, [1,1,img_h, img_w]).astype("float32") # 把y通道作为输入
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model('work/model/', exe))
results = exe.run(inference_program,
feed={feed_target_names[0]: img_input},
fetch_list=fetch_targets)[0]
result_img = np.array(results)
result_img[result_img < 0] = 0
result_img[result_img >255] = 255
gap_y = (img_y.shape[0]-result_img.shape[2])/2
gap_x = (img_y.shape[1]-result_img.shape[3])/2
print "=====Y通道输入========="
plt.imshow(img_y, cmap='gray')
plt.show()
img_y[gap_y: gap_y + result_img.shape[2],
gap_x: gap_x + result_img.shape[3]]=result_img
img_test_r = cv2.merge([img_y, img_u, img_v])
img_test_r = cv2.cvtColor(img_test_r, cv2.COLOR_YCrCb2BGR)
print "=====Y通道输出========="
plt.imshow(img_y, cmap='gray')
plt.show()
print "=====彩图结果========="
b,g,r = cv2.split(img_test_r)
img_test_show = cv2.merge([r,g,b])
plt.imshow(img_test_show)
plt.show()
def read_data(self, data_path):
def data_reader():
for image in os.listdir(data_path):
if image.endswith('.bmp'):
img = cv2.imread(os.path.join(data_path, image))
yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YCrCb)
img_y, img_u, img_v = cv2.split(yuv)
# 下面是切图的步骤
j = 0
count = 0
while j+33 < len(img_y):
i = 0
while i+33 < len(img_y[0]):
img_patch = img_y[j:j+33, i:i+33]
img_gth = img_patch[6:27, 6:27].copy()
img_blur = cv2.GaussianBlur(img_patch, (5, 5), 0)
img_sumsample = cv2.resize(img_blur, (11, 11))
img_input = cv2.resize(img_blur, (33, 33), interpolation=cv2.INTER_CUBIC)
yield img_input, img_gth
i+=14
j+= 14
return data_readerIn[7]
model = SRCNN(0.0001, 0.00001, 100, 150)
model.train()0 Epoch: 0 Cur Cost : 18164.226562 Val Cost: 21233.191406 PSNR :4.860651 219 Epoch: 1 Cur Cost : 284.385284 Val Cost: 417.904175 PSNR :21.920037 438 Epoch: 2 Cur Cost : 281.428558 Val Cost: 413.779419 PSNR :21.963115 657 Epoch: 3 Cur Cost : 277.070312 Val Cost: 406.204071 PSNR :22.043361 876 Epoch: 4 Cur Cost : 253.116455 Val Cost: 372.079315 PSNR :22.424448 1095 Epoch: 5 Cur Cost : 207.459930 Val Cost: 303.296204 PSNR :23.312134 1314 Epoch: 6 Cur Cost : 154.225296 Val Cost: 221.624435 PSNR :24.674627 1533 Epoch: 7 Cur Cost : 117.966713 Val Cost: 165.380890 PSNR :25.945950 1752 Epoch: 8 Cur Cost : 114.436462 Val Cost: 158.598419 PSNR :26.127815 1971 Epoch: 9 Cur Cost : 112.926453 Val Cost: 155.753204 PSNR :26.206434 2190 Epoch: 10 Cur Cost : 111.557121 Val Cost: 153.391846 PSNR :26.272781 2409 Epoch: 11 Cur Cost : 110.163071 Val Cost: 151.123138 PSNR :26.337494 2628 Epoch: 12 Cur Cost : 108.720184 Val Cost: 148.844711 PSNR :26.403470 2847 Epoch: 13 Cur Cost : 107.220848 Val Cost: 146.517197 PSNR :26.471918 3066 Epoch: 14 Cur Cost : 105.661781 Val Cost: 144.118179 PSNR :26.543616 3285 Epoch: 15 Cur Cost : 104.040611 Val Cost: 141.630585 PSNR :26.619233 3504 Epoch: 16 Cur Cost : 102.351303 Val Cost: 139.039047 PSNR :26.699436 3723 Epoch: 17 Cur Cost : 100.585503 Val Cost: 136.326965 PSNR :26.784986 3942 Epoch: 18 Cur Cost : 98.731277 Val Cost: 133.479095 PSNR :26.876671 4161 Epoch: 19 Cur Cost : 96.780418 Val Cost: 130.485748 PSNR :26.975173 4380 Epoch: 20 Cur Cost : 94.736946 Val Cost: 127.345490 PSNR :27.080968 4599 Epoch: 21 Cur Cost : 92.607079 Val Cost: 124.054619 PSNR :27.194674 4818 Epoch: 22 Cur Cost : 90.392410 Val Cost: 120.612396 PSNR :27.316884 5037 Epoch: 23 Cur Cost : 88.097008 Val Cost: 117.033394 PSNR :27.447706 5256 Epoch: 24 Cur Cost : 85.727638 Val Cost: 113.339539 PSNR :27.586989 5475 Epoch: 25 Cur Cost : 83.291718 Val Cost: 109.558945 PSNR :27.734325 5694 Epoch: 26 Cur Cost : 80.803673 Val Cost: 105.716370 PSNR :27.889381 5913 Epoch: 27 Cur Cost : 78.289711 Val Cost: 101.814842 PSNR :28.052693 6132 Epoch: 28 Cur Cost : 75.769890 Val Cost: 97.854851 PSNR :28.224980 6351 Epoch: 29 Cur Cost : 73.263191 Val Cost: 93.865204 PSNR :28.405757 6570 Epoch: 30 Cur Cost : 70.802010 Val Cost: 89.902550 PSNR :28.593084 6789 Epoch: 31 Cur Cost : 68.425529 Val Cost: 86.056076 PSNR :28.782988 7008 Epoch: 32 Cur Cost : 66.165245 Val Cost: 82.378616 PSNR :28.972659 7227 Epoch: 33 Cur Cost : 64.029266 Val Cost: 78.877708 PSNR :29.161261 7446 Epoch: 34 Cur Cost : 61.938168 Val Cost: 75.441162 PSNR :29.354720 7665 Epoch: 35 Cur Cost : 59.845867 Val Cost: 71.962860 PSNR :29.559719 7884 Epoch: 36 Cur Cost : 57.924450 Val Cost: 68.783943 PSNR :29.755933 8103 Epoch: 37 Cur Cost : 56.233124 Val Cost: 65.943611 PSNR :29.939076 8322 Epoch: 38 Cur Cost : 54.610748 Val Cost: 63.266205 PSNR :30.119086 8541 Epoch: 39 Cur Cost : 53.087059 Val Cost: 60.819660 PSNR :30.290364 8760 Epoch: 40 Cur Cost : 51.752823 Val Cost: 58.767235 PSNR :30.439451 8979 Epoch: 41 Cur Cost : 50.548962 Val Cost: 56.994938 PSNR :30.572441 9198 Epoch: 42 Cur Cost : 49.409630 Val Cost: 55.392437 PSNR :30.696299 9417 Epoch: 43 Cur Cost : 48.292534 Val Cost: 53.872486 PSNR :30.817133 9636 Epoch: 44 Cur Cost : 47.169697 Val Cost: 52.356304 PSNR :30.941114 9855 Epoch: 45 Cur Cost : 46.026646 Val Cost: 50.804356 PSNR :31.071794 10074 Epoch: 46 Cur Cost : 44.855438 Val Cost: 49.224430 PSNR :31.208997 10293 Epoch: 47 Cur Cost : 43.651192 Val Cost: 47.612514 PSNR :31.353592 10512 Epoch: 48 Cur Cost : 42.413403 Val Cost: 45.968498 PSNR :31.506200 10731 Epoch: 49 Cur Cost : 41.161961 Val Cost: 44.319637 PSNR :31.664842 10950 Epoch: 50 Cur Cost : 39.941639 Val Cost: 42.723743 PSNR :31.824111 11169 Epoch: 51 Cur Cost : 38.763580 Val Cost: 41.210186 PSNR :31.980758 11388 Epoch: 52 Cur Cost : 37.628029 Val Cost: 39.770855 PSNR :32.135154 11607 Epoch: 53 Cur Cost : 36.646980 Val Cost: 38.539745 PSNR :32.271715 11826 Epoch: 54 Cur Cost : 35.795269 Val Cost: 37.488255 PSNR :32.391851 12045 Epoch: 55 Cur Cost : 35.093742 Val Cost: 36.636829 PSNR :32.491625 12264 Epoch: 56 Cur Cost : 34.488441 Val Cost: 35.905704 PSNR :32.579169 12483 Epoch: 57 Cur Cost : 33.963333 Val Cost: 35.274933 PSNR :32.656142 12702 Epoch: 58 Cur Cost : 33.505180 Val Cost: 34.726208 PSNR :32.724230 12921 Epoch: 59 Cur Cost : 33.103012 Val Cost: 34.243057 PSNR :32.785078 13140 Epoch: 60 Cur Cost : 32.750072 Val Cost: 33.818142 PSNR :32.839306 13359 Epoch: 61 Cur Cost : 32.436031 Val Cost: 33.438747 PSNR :32.888304 13578 Epoch: 62 Cur Cost : 32.150639 Val Cost: 33.090755 PSNR :32.933737 13797 Epoch: 63 Cur Cost : 31.886650 Val Cost: 32.765667 PSNR :32.976613 14016 Epoch: 64 Cur Cost : 31.639585 Val Cost: 32.457993 PSNR :33.017587 14235 Epoch: 65 Cur Cost : 31.405382 Val Cost: 32.163200 PSNR :33.057211 14454 Epoch: 66 Cur Cost : 31.181067 Val Cost: 31.877956 PSNR :33.095899 14673 Epoch: 67 Cur Cost : 30.963821 Val Cost: 31.595957 PSNR :33.134488 14892 Epoch: 68 Cur Cost : 30.745806 Val Cost: 31.313644 PSNR :33.173467 15111 Epoch: 69 Cur Cost : 30.530462 Val Cost: 31.029354 PSNR :33.213076 15330 Epoch: 70 Cur Cost : 30.316664 Val Cost: 30.746445 PSNR :33.252855 15549 Epoch: 71 Cur Cost : 30.103735 Val Cost: 30.461548 PSNR :33.293284 15768 Epoch: 72 Cur Cost : 29.889498 Val Cost: 30.172472 PSNR :33.334695 15987 Epoch: 73 Cur Cost : 29.668657 Val Cost: 29.870338 PSNR :33.378402 16206 Epoch: 74 Cur Cost : 29.430288 Val Cost: 29.451651 PSNR :33.439707 16425 Epoch: 75 Cur Cost : 29.207491 Val Cost: 29.195040 PSNR :33.477713 16644 Epoch: 76 Cur Cost : 29.008448 Val Cost: 28.942610 PSNR :33.515427 16863 Epoch: 77 Cur Cost : 28.805315 Val Cost: 28.677612 PSNR :33.555374 17082 Epoch: 78 Cur Cost : 28.597673 Val Cost: 28.398453 PSNR :33.597857 17301 Epoch: 79 Cur Cost : 28.384197 Val Cost: 28.109222 PSNR :33.642315 17520 Epoch: 80 Cur Cost : 28.165312 Val Cost: 27.810051 PSNR :33.688786 17739 Epoch: 81 Cur Cost : 27.941040 Val Cost: 27.501369 PSNR :33.737260 17958 Epoch: 82 Cur Cost : 27.711397 Val Cost: 27.183748 PSNR :33.787710 18177 Epoch: 83 Cur Cost : 27.476480 Val Cost: 26.856071 PSNR :33.840379 18396 Epoch: 84 Cur Cost : 27.235754 Val Cost: 26.519581 PSNR :33.895137 18615 Epoch: 85 Cur Cost : 26.991680 Val Cost: 26.172489 PSNR :33.952353 18834 Epoch: 86 Cur Cost : 26.743433 Val Cost: 25.811661 PSNR :34.012644 19053 Epoch: 87 Cur Cost : 26.480677 Val Cost: 25.444807 PSNR :34.074812 19272 Epoch: 88 Cur Cost : 26.224283 Val Cost: 25.074034 PSNR :34.138562 19491 Epoch: 89 Cur Cost : 25.963892 Val Cost: 24.700274 PSNR :34.203786 19710 Epoch: 90 Cur Cost : 25.710884 Val Cost: 24.330662 PSNR :34.269264 19929 Epoch: 91 Cur Cost : 25.457993 Val Cost: 23.969437 PSNR :34.334225 20148 Epoch: 92 Cur Cost : 25.210745 Val Cost: 23.615768 PSNR :34.398783 20367 Epoch: 93 Cur Cost : 24.963886 Val Cost: 23.267546 PSNR :34.463298 20586 Epoch: 94 Cur Cost : 24.722895 Val Cost: 22.925728 PSNR :34.527572 20805 Epoch: 95 Cur Cost : 24.486914 Val Cost: 22.588749 PSNR :34.591882 21024 Epoch: 96 Cur Cost : 24.253819 Val Cost: 22.257490 PSNR :34.656042 21243 Epoch: 97 Cur Cost : 24.023243 Val Cost: 21.936876 PSNR :34.719056 21462 Epoch: 98 Cur Cost : 23.806646 Val Cost: 21.630701 PSNR :34.780098 21681 Epoch: 99 Cur Cost : 23.600349 Val Cost: 21.338070 PSNR :34.839252 21900 Epoch: 100 Cur Cost : 23.404703 Val Cost: 21.059601 PSNR :34.896302 22119 Epoch: 101 Cur Cost : 23.220484 Val Cost: 20.796125 PSNR :34.950979 22338 Epoch: 102 Cur Cost : 23.047068 Val Cost: 20.546223 PSNR :35.003484 22557 Epoch: 103 Cur Cost : 22.884609 Val Cost: 20.312155 PSNR :35.053244 22776 Epoch: 104 Cur Cost : 22.732927 Val Cost: 20.091757 PSNR :35.100624 22995 Epoch: 105 Cur Cost : 22.590960 Val Cost: 19.883532 PSNR :35.145868 23214 Epoch: 106 Cur Cost : 22.457445 Val Cost: 19.684614 PSNR :35.189535 23433 Epoch: 107 Cur Cost : 22.331661 Val Cost: 19.496771 PSNR :35.231177
23652 Epoch: 108 Cur Cost : 22.212849 Val Cost: 19.317987 PSNR :35.271185 23871 Epoch: 109 Cur Cost : 22.100451 Val Cost: 19.148111 PSNR :35.309544 24090 Epoch: 110 Cur Cost : 21.994497 Val Cost: 18.987167 PSNR :35.346202 24309 Epoch: 111 Cur Cost : 21.894119 Val Cost: 18.834215 PSNR :35.381328 24528 Epoch: 112 Cur Cost : 21.798851 Val Cost: 18.687769 PSNR :35.415229 24747 Epoch: 113 Cur Cost : 21.708071 Val Cost: 18.547321 PSNR :35.447992 24966 Epoch: 114 Cur Cost : 21.621136 Val Cost: 18.413755 PSNR :35.479380 25185 Epoch: 115 Cur Cost : 21.537928 Val Cost: 18.285975 PSNR :35.509623 25404 Epoch: 116 Cur Cost : 21.458000 Val Cost: 18.164145 PSNR :35.538654 25623 Epoch: 117 Cur Cost : 21.381271 Val Cost: 18.046951 PSNR :35.566765 25842 Epoch: 118 Cur Cost : 21.307785 Val Cost: 17.934608 PSNR :35.593885 26061 Epoch: 119 Cur Cost : 21.236940 Val Cost: 17.827406 PSNR :35.619922 26280 Epoch: 120 Cur Cost : 21.168789 Val Cost: 17.724836 PSNR :35.644981 26499 Epoch: 121 Cur Cost : 21.103100 Val Cost: 17.627266 PSNR :35.668954 26718 Epoch: 122 Cur Cost : 21.039707 Val Cost: 17.533913 PSNR :35.692015 26937 Epoch: 123 Cur Cost : 20.978670 Val Cost: 17.444506 PSNR :35.714217 27156 Epoch: 124 Cur Cost : 20.919706 Val Cost: 17.358051 PSNR :35.735794 27375 Epoch: 125 Cur Cost : 20.862633 Val Cost: 17.275480 PSNR :35.756502 27594 Epoch: 126 Cur Cost : 20.807596 Val Cost: 17.195841 PSNR :35.776569 27813 Epoch: 127 Cur Cost : 20.754372 Val Cost: 17.119400 PSNR :35.795918 28032 Epoch: 128 Cur Cost : 20.702871 Val Cost: 17.045691 PSNR :35.814658 28251 Epoch: 129 Cur Cost : 20.652927 Val Cost: 16.974653 PSNR :35.832794 28470 Epoch: 130 Cur Cost : 20.604521 Val Cost: 16.906082 PSNR :35.850374 28689 Epoch: 131 Cur Cost : 20.557560 Val Cost: 16.840302 PSNR :35.867305 28908 Epoch: 132 Cur Cost : 20.511902 Val Cost: 16.776218 PSNR :35.883863 29127 Epoch: 133 Cur Cost : 20.467619 Val Cost: 16.714396 PSNR :35.899897 29346 Epoch: 134 Cur Cost : 20.424440 Val Cost: 16.654495 PSNR :35.915489 29565 Epoch: 135 Cur Cost : 20.382488 Val Cost: 16.596294 PSNR :35.930692 29784 Epoch: 136 Cur Cost : 20.341431 Val Cost: 16.539652 PSNR :35.945540 30003 Epoch: 137 Cur Cost : 20.301138 Val Cost: 16.484528 PSNR :35.960039 30222 Epoch: 138 Cur Cost : 20.261866 Val Cost: 16.430748 PSNR :35.974230 30441 Epoch: 139 Cur Cost : 20.223341 Val Cost: 16.378759 PSNR :35.987994 30660 Epoch: 140 Cur Cost : 20.185614 Val Cost: 16.328671 PSNR :36.001295 30879 Epoch: 141 Cur Cost : 20.148609 Val Cost: 16.278326 PSNR :36.014706 31098 Epoch: 142 Cur Cost : 20.112509 Val Cost: 16.229719 PSNR :36.027694 31317 Epoch: 143 Cur Cost : 20.076233 Val Cost: 16.180443 PSNR :36.040900 31536 Epoch: 144 Cur Cost : 20.039700 Val Cost: 16.131332 PSNR :36.054101 31755 Epoch: 145 Cur Cost : 20.001612 Val Cost: 16.085133 PSNR :36.066557 31974 Epoch: 146 Cur Cost : 19.969093 Val Cost: 16.049982 PSNR :36.076058 32193 Epoch: 147 Cur Cost : 19.926191 Val Cost: 15.993620 PSNR :36.091336 32412 Epoch: 148 Cur Cost : 19.890646 Val Cost: 15.950057 PSNR :36.103181 32631 Epoch: 149 Cur Cost : 19.856085 Val Cost: 15.906347 PSNR :36.115099
In[8]
# 绘制训练曲线
plt.plot(model.backprops_cnts,model.psnr)[]
In[9]
model.generate_reconstruct_img('butterfly_GT.bmp')=====原始图片=========
=====Y通道输入=========
=====Y通道输出=========
=====彩图结果=========
In[10]
model.generate_reconstruct_img('baby_GT.bmp')=====原始图片=========
=====Y通道输入=========
=====Y通道输出=========
=====彩图结果=========
In[11]
class SRCNN_fz(object):
# 只对Y通道做超分辨率重构
def __init__(self, lr, lr_f, batch_size, iter_num):
self.lr = lr # 学习率
self.lr_f = lr_f
self.batch_size = batch_size
self.iter_num = iter_num # 总共训练多少次
def net(self, X, Y):
# 搭建模型
conv1 = fluid.layers.conv2d(X, 64, 9,act='relu', name='conv1' ,
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='conv1_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='conv1_b'))
conv2 = fluid.layers.conv2d(conv1, 32, 3, act='relu', name='conv2' ,
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='conv2_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='conv2_b'))
pred = fluid.layers.conv2d(conv2, 1, 5, name='pred',
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='pred_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='pred_b'))
loss = fluid.layers.reduce_mean(fluid.layers.square(pred - Y))
return pred, loss
def train(self):
# 模型训练
X_train = fluid.layers.data(shape=[1, 33, 33], dtype='float32', name='image')
Y_train = fluid.layers.data(shape=[1, 19, 19], dtype='float32', name='gdt')
y_predict, y_loss = self.net(X_train, Y_train)
Optimizer = fluid.optimizer.AdamOptimizer(learning_rate=self.lr)
Optimizer_f = fluid.optimizer.AdamOptimizer(learning_rate=self.lr_f)
Optimizer.minimize(y_loss, parameter_list=['conv1_w','conv1_b', 'conv2_w', 'conv2_b'])
Optimizer_f.minimize(y_loss, parameter_list=['pred_w', 'pred_b'])
# 读取训练集数据
train_reader = paddle.batch(self.read_data('work/dataset/timofte'), batch_size=self.batch_size)
# 定义执行器
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
def train_loop(main_program):
feeder = fluid.DataFeeder(place=place, feed_list=[X_train, Y_train])
exe.run(fluid.default_startup_program())
backprops_cnt = 0 # 论文中作图的横坐标
self.backprops_cnts = [] # 绑定为类的一个属性，用于画图
self.psnr = [] # psnr的值
for epoch in range(self.iter_num):
for batch_id, data in enumerate(train_reader()):
loss = exe.run(
fluid.framework.default_main_program(),
feed=feeder.feed(data),
fetch_list=[y_loss])
if batch_id == 0: # 每个epoch算一下psnr，画图用的
# # 算psnr要在测试集上面
fluid.io.save_inference_model('work/model_fz/', ['image'], [y_predict], exe)
val_loss, val_psnr = self.validation()
self.backprops_cnts.append(backprops_cnt * self.batch_size)
self.psnr.append(val_psnr)
print("%i\tEpoch: %d \tCur Cost : %f\t Val Cost: %f\t PSNR :%f" % (backprops_cnt, epoch, np.array(loss[0])[0], val_loss, val_psnr))
backprops_cnt += 1
fluid.io.save_inference_model('work/model_fz/', ['image'], [y_predict], exe)
train_loop(fluid.default_main_program())
def validation(self):
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
test_set = 'work/dataset/set5/'
scale_factor = 3
for img_name in os.listdir(test_set):
img_val = cv2.imread(os.path.join(test_set, img_name))
yuv = cv2.cvtColor(img_val, cv2.COLOR_BGR2YCrCb)
img_y, img_u, img_v = cv2.split(yuv)
img_h, img_w = img_y.shape
img_blur = cv2.GaussianBlur(img_y, (5, 5), 0)
img_subsample = cv2.resize(img_blur, (img_w/scale_factor, img_h/scale_factor))
img_input = cv2.resize(img_blur, (img_w, img_h), interpolation=cv2.INTER_CUBIC)
img_input = np.reshape(img_input, [1,1,img_h, img_w]).astype("float32") # h,w
losses = []
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model('work/model_fz/', exe))
results = exe.run(inference_program,
feed={feed_target_names[0]: img_input},
fetch_list=fetch_targets)[0]
loss = np.mean(np.square(results[0,0]-img_y[7:-7, 7:-7]))
losses.append(loss)
avg_loss = np.sum(np.array(losses))/len(losses)
psnr = 10 * np.log10(255*255/avg_loss)
return avg_loss,psnr
def generate_reconstruct_img(self, img_name):
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
img_test = cv2.imread('work/dataset/set5/%s' % img_name)
yuv_test = cv2.cvtColor(img_test, cv2.COLOR_BGR2YCrCb)
img_h, img_w, img_c = img_test.shape
print "=====原始图片========="
b,g,r = cv2.split(img_test) # AI studio 不支持cv2.imshow，所以用plt.imshow输出，两者rgb顺序不一样
img_test = cv2.merge([r,g,b])
plt.imshow(img_test)
plt.show()
# 图像模糊+cubic插值
img_blur = cv2.GaussianBlur(yuv_test.copy(), (5, 5), 0)
img_subsample = cv2.resize(img_blur, (img_w/3, img_h/3)) #这里注意cv2.resize里面的shape是w,h的顺序
img_cubic = cv2.resize(img_blur, (img_w, img_h), interpolation=cv2.INTER_CUBIC)
img_y, img_u, img_v = cv2.split(img_cubic)
img_input = np.reshape(img_y, [1,1,img_h, img_w]).astype("float32") # 把y通道作为输入
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model('work/model_fz/', exe))
results = exe.run(inference_program,
feed={feed_target_names[0]: img_input},
fetch_list=fetch_targets)[0]
result_img = np.array(results)
result_img[result_img < 0] = 0
result_img[result_img >255] = 255
gap_y = (img_y.shape[0]-result_img.shape[2])/2
gap_x = (img_y.shape[1]-result_img.shape[3])/2
print "=====Y通道输入========="
plt.imshow(img_y, cmap='gray')
plt.show()
img_y[gap_y: gap_y + result_img.shape[2],
gap_x: gap_x + result_img.shape[3]]=result_img
img_test_r = cv2.merge([img_y, img_u, img_v])
img_test_r = cv2.cvtColor(img_test_r, cv2.COLOR_YCrCb2BGR)
print "=====Y通道输出========="
plt.imshow(img_y, cmap='gray')
plt.show()
print "=====彩图结果========="
b,g,r = cv2.split(img_test_r)
img_test_show = cv2.merge([r,g,b])
plt.imshow(img_test_show)
plt.show()
def read_data(self, data_path):
def data_reader():
for image in os.listdir(data_path):
if image.endswith('.bmp'):
img = cv2.imread(os.path.join(data_path, image))
yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YCrCb)
img_y, img_u, img_v = cv2.split(yuv)
# 下面是切图的步骤
j = 0
count = 0
while j+33 < len(img_y):
i = 0
while i+33 < len(img_y[0]):
img_patch = img_y[j:j+33, i:i+33]
img_gth = img_patch[7:-7, 7:-7].copy()
img_blur = cv2.GaussianBlur(img_patch, (5, 5), 0)
img_sumsample = cv2.resize(img_blur, (11, 11))
img_input = cv2.resize(img_blur, (33, 33), interpolation=cv2.INTER_CUBIC)
yield img_input, img_gth
i+=14
j+= 14
return data_reader改变filter_size的训练结果
需要重启一下kernel清空定义的模型
In[12]
model = SRCNN_fz(0.0001, 0.00001, 100, 150)
model.train()0 Epoch: 0 Cur Cost : 18187.878906 Val Cost: 21415.037109 PSNR :4.823615 219 Epoch: 1 Cur Cost : 293.490234 Val Cost: 423.864105 PSNR :21.858537 438 Epoch: 2 Cur Cost : 286.713898 Val Cost: 413.923096 PSNR :21.961607 657 Epoch: 3 Cur Cost : 277.595795 Val Cost: 400.915894 PSNR :22.100271 876 Epoch: 4 Cur Cost : 266.524780 Val Cost: 385.093658 PSNR :22.275140 1095 Epoch: 5 Cur Cost : 253.052292 Val Cost: 365.947205 PSNR :22.496619 1314 Epoch: 6 Cur Cost : 236.651108 Val Cost: 342.572601 PSNR :22.783277 1533 Epoch: 7 Cur Cost : 216.273224 Val Cost: 312.513641 PSNR :23.182114 1752 Epoch: 8 Cur Cost : 189.500977 Val Cost: 272.160767 PSNR :23.782548 1971 Epoch: 9 Cur Cost : 157.191208 Val Cost: 224.049789 PSNR :24.627358 2190 Epoch: 10 Cur Cost : 129.589066 Val Cost: 182.939468 PSNR :25.507729 2409 Epoch: 11 Cur Cost : 113.882393 Val Cost: 155.549286 PSNR :26.212123 2628 Epoch: 12 Cur Cost : 109.084671 Val Cost: 146.027054 PSNR :26.486470 2847 Epoch: 13 Cur Cost : 104.128502 Val Cost: 137.899292 PSNR :26.735183 3066 Epoch: 14 Cur Cost : 97.342422 Val Cost: 127.754402 PSNR :27.067045 3285 Epoch: 15 Cur Cost : 91.355812 Val Cost: 118.736206 PSNR :27.384972 3504 Epoch: 16 Cur Cost : 85.422371 Val Cost: 109.795456 PSNR :27.724960 3723 Epoch: 17 Cur Cost : 79.140945 Val Cost: 100.095528 PSNR :28.126657 3942 Epoch: 18 Cur Cost : 73.169403 Val Cost: 90.881409 PSNR :28.546053 4161 Epoch: 19 Cur Cost : 67.917305 Val Cost: 82.804932 PSNR :28.950242 4380 Epoch: 20 Cur Cost : 63.625111 Val Cost: 76.324135 PSNR :29.304185 4599 Epoch: 21 Cur Cost : 59.751232 Val Cost: 70.921104 PSNR :29.623049 4818 Epoch: 22 Cur Cost : 55.927074 Val Cost: 65.637527 PSNR :29.959281 5037 Epoch: 23 Cur Cost : 52.361134 Val Cost: 60.628269 PSNR :30.304052 5256 Epoch: 24 Cur Cost : 49.219032 Val Cost: 56.010677 PSNR :30.648095 5475 Epoch: 25 Cur Cost : 46.496906 Val Cost: 51.900589 PSNR :30.979081 5694 Epoch: 26 Cur Cost : 44.209206 Val Cost: 48.567101 PSNR :31.267382 5913 Epoch: 27 Cur Cost : 42.444748 Val Cost: 46.098759 PSNR :31.493911 6132 Epoch: 28 Cur Cost : 40.951580 Val Cost: 44.224030 PSNR :31.674220 6351 Epoch: 29 Cur Cost : 39.668636 Val Cost: 42.721249 PSNR :31.824364 6570 Epoch: 30 Cur Cost : 38.468319 Val Cost: 41.372662 PSNR :31.963669 6789 Epoch: 31 Cur Cost : 37.305447 Val Cost: 40.067513 PSNR :32.102880 7008 Epoch: 32 Cur Cost : 36.248657 Val Cost: 38.869499 PSNR :32.234714 7227 Epoch: 33 Cur Cost : 35.280151 Val Cost: 37.761169 PSNR :32.360349 7446 Epoch: 34 Cur Cost : 34.358017 Val Cost: 36.653786 PSNR :32.489615 7665 Epoch: 35 Cur Cost : 33.456329 Val Cost: 35.482918 PSNR :32.630610 7884 Epoch: 36 Cur Cost : 32.594936 Val Cost: 34.364208 PSNR :32.769740 8103 Epoch: 37 Cur Cost : 31.807510 Val Cost: 33.297546 PSNR :32.906681 8322 Epoch: 38 Cur Cost : 31.098452 Val Cost: 32.322258 PSNR :33.035787 8541 Epoch: 39 Cur Cost : 30.458292 Val Cost: 31.449808 PSNR :33.154624 8760 Epoch: 40 Cur Cost : 29.896788 Val Cost: 30.646156 PSNR :33.267043 8979 Epoch: 41 Cur Cost : 29.400312 Val Cost: 29.952234 PSNR :33.366511 9198 Epoch: 42 Cur Cost : 28.963032 Val Cost: 29.316887 PSNR :33.459625 9417 Epoch: 43 Cur Cost : 28.554939 Val Cost: 28.730633 PSNR :33.547352 9636 Epoch: 44 Cur Cost : 28.187880 Val Cost: 28.178003 PSNR :33.631701 9855 Epoch: 45 Cur Cost : 27.815355 Val Cost: 27.621120 PSNR :33.718391 10074 Epoch: 46 Cur Cost : 27.450251 Val Cost: 27.054892 PSNR :33.808346 10293 Epoch: 47 Cur Cost : 27.095476 Val Cost: 26.506290 PSNR :33.897314 10512 Epoch: 48 Cur Cost : 26.726200 Val Cost: 25.993595 PSNR :33.982140 10731 Epoch: 49 Cur Cost : 26.418917 Val Cost: 25.480820 PSNR :34.068670 10950 Epoch: 50 Cur Cost : 26.102079 Val Cost: 25.046286 PSNR :34.143370 11169 Epoch: 51 Cur Cost : 25.814594 Val Cost: 24.607016 PSNR :34.220214 11388 Epoch: 52 Cur Cost : 25.541672 Val Cost: 24.180517 PSNR :34.296148 11607 Epoch: 53 Cur Cost : 25.276190 Val Cost: 23.795362 PSNR :34.365880 11826 Epoch: 54 Cur Cost : 25.032814 Val Cost: 23.423958 PSNR :34.434201 12045 Epoch: 55 Cur Cost : 24.795341 Val Cost: 23.076561 PSNR :34.499093 12264 Epoch: 56 Cur Cost : 24.567106 Val Cost: 22.729460 PSNR :34.564912 12483 Epoch: 57 Cur Cost : 24.350388 Val Cost: 22.400122 PSNR :34.628300 12702 Epoch: 58 Cur Cost : 24.139818 Val Cost: 22.082649 PSNR :34.690292 12921 Epoch: 59 Cur Cost : 23.937288 Val Cost: 21.775942 PSNR :34.751034 13140 Epoch: 60 Cur Cost : 23.740856 Val Cost: 21.478758 PSNR :34.810712 13359 Epoch: 61 Cur Cost : 23.550505 Val Cost: 21.190653 PSNR :34.869360 13578 Epoch: 62 Cur Cost : 23.365761 Val Cost: 20.912657 PSNR :34.926712 13797 Epoch: 63 Cur Cost : 23.186628 Val Cost: 20.643282 PSNR :34.983016 14016 Epoch: 64 Cur Cost : 23.012552 Val Cost: 20.381088 PSNR :35.038530 14235 Epoch: 65 Cur Cost : 22.843679 Val Cost: 20.126986 PSNR :35.093016 14454 Epoch: 66 Cur Cost : 22.680412 Val Cost: 19.880857 PSNR :35.146452 14673 Epoch: 67 Cur Cost : 22.522045 Val Cost: 19.643797 PSNR :35.198549 14892 Epoch: 68 Cur Cost : 22.368488 Val Cost: 19.414352 PSNR :35.249575 15111 Epoch: 69 Cur Cost : 22.227560 Val Cost: 19.191557 PSNR :35.299702 15330 Epoch: 70 Cur Cost : 22.084782 Val Cost: 18.978037 PSNR :35.348291 15549 Epoch: 71 Cur Cost : 21.936750 Val Cost: 18.766928 PSNR :35.396872 15768 Epoch: 72 Cur Cost : 21.809185 Val Cost: 18.569225 PSNR :35.442866 15987 Epoch: 73 Cur Cost : 21.701038 Val Cost: 18.389750 PSNR :35.485045 16206 Epoch: 74 Cur Cost : 21.573944 Val Cost: 18.225378 PSNR :35.524038 16425 Epoch: 75 Cur Cost : 21.466406 Val Cost: 18.062704 PSNR :35.562976 16644 Epoch: 76 Cur Cost : 21.362574 Val Cost: 17.919756 PSNR :35.597483 16863 Epoch: 77 Cur Cost : 21.265091 Val Cost: 17.759726 PSNR :35.636441 17082 Epoch: 78 Cur Cost : 21.167427 Val Cost: 17.617479 PSNR :35.671366 17301 Epoch: 79 Cur Cost : 21.079617 Val Cost: 17.467531 PSNR :35.708488 17520 Epoch: 80 Cur Cost : 20.975424 Val Cost: 17.328568 PSNR :35.743177 17739 Epoch: 81 Cur Cost : 20.885769 Val Cost: 17.211353 PSNR :35.772653 17958 Epoch: 82 Cur Cost : 20.804918 Val Cost: 17.072248 PSNR :35.807896 18177 Epoch: 83 Cur Cost : 20.729193 Val Cost: 16.936848 PSNR :35.842478 18396 Epoch: 84 Cur Cost : 20.651018 Val Cost: 16.827904 PSNR :35.870503 18615 Epoch: 85 Cur Cost : 20.577908 Val Cost: 16.708214 PSNR :35.901503 18834 Epoch: 86 Cur Cost : 20.506536 Val Cost: 16.616051 PSNR :35.925526 19053 Epoch: 87 Cur Cost : 20.444763 Val Cost: 16.530025 PSNR :35.948068 19272 Epoch: 88 Cur Cost : 20.384291 Val Cost: 16.441322 PSNR :35.971436 19491 Epoch: 89 Cur Cost : 20.331388 Val Cost: 16.359850 PSNR :35.993010 19710 Epoch: 90 Cur Cost : 20.274405 Val Cost: 16.283634 PSNR :36.013290 19929 Epoch: 91 Cur Cost : 20.223141 Val Cost: 16.202314 PSNR :36.035033 20148 Epoch: 92 Cur Cost : 20.172297 Val Cost: 16.124725 PSNR :36.055880 20367 Epoch: 93 Cur Cost : 20.120977 Val Cost: 16.051615 PSNR :36.075616 20586 Epoch: 94 Cur Cost : 20.073071 Val Cost: 15.968109 PSNR :36.098269 20805 Epoch: 95 Cur Cost : 20.027214 Val Cost: 15.915732 PSNR :36.112537 21024 Epoch: 96 Cur Cost : 19.984432 Val Cost: 15.839420 PSNR :36.133411 21243 Epoch: 97 Cur Cost : 19.940655 Val Cost: 15.775101 PSNR :36.151082 21462 Epoch: 98 Cur Cost : 19.899282 Val Cost: 15.708839 PSNR :36.169363 21681 Epoch: 99 Cur Cost : 19.859011 Val Cost: 15.640730 PSNR :36.188233 21900 Epoch: 100 Cur Cost : 19.821726 Val Cost: 15.585388 PSNR :36.203627 22119 Epoch: 101 Cur Cost : 19.786446 Val Cost: 15.522793 PSNR :36.221105 22338 Epoch: 102 Cur Cost : 19.748981 Val Cost: 15.472646 PSNR :36.235158 22557 Epoch: 103 Cur Cost : 19.715595 Val Cost: 15.406190 PSNR :36.253851 22776 Epoch: 104 Cur Cost : 19.679775 Val Cost: 15.351366 PSNR :36.269333 22995 Epoch: 105 Cur Cost : 19.644049 Val Cost: 15.300534 PSNR :36.283738 23214 Epoch: 106 Cur Cost : 19.611214 Val Cost: 15.245579 PSNR :36.299364 23433 Epoch: 107 Cur Cost : 19.583275 Val Cost: 15.178621 PSNR :36.318480 23652 Epoch: 108 Cur Cost : 19.554924 Val Cost: 15.142241 PSNR :36.328902
23871 Epoch: 109 Cur Cost : 19.527620 Val Cost: 15.100776 PSNR :36.340811 24090 Epoch: 110 Cur Cost : 19.499495 Val Cost: 15.060397 PSNR :36.352439 24309 Epoch: 111 Cur Cost : 19.472237 Val Cost: 15.014300 PSNR :36.365753 24528 Epoch: 112 Cur Cost : 19.443647 Val Cost: 14.971251 PSNR :36.378223 24747 Epoch: 113 Cur Cost : 19.417927 Val Cost: 14.922546 PSNR :36.392374 24966 Epoch: 114 Cur Cost : 19.388918 Val Cost: 14.871901 PSNR :36.407139 25185 Epoch: 115 Cur Cost : 19.362228 Val Cost: 14.819825 PSNR :36.422373 25404 Epoch: 116 Cur Cost : 19.338102 Val Cost: 14.777211 PSNR :36.434879 25623 Epoch: 117 Cur Cost : 19.306536 Val Cost: 14.740740 PSNR :36.445611 25842 Epoch: 118 Cur Cost : 19.279057 Val Cost: 14.691047 PSNR :36.460276 26061 Epoch: 119 Cur Cost : 19.254765 Val Cost: 14.637246 PSNR :36.476210 26280 Epoch: 120 Cur Cost : 19.223206 Val Cost: 14.604916 PSNR :36.485813 26499 Epoch: 121 Cur Cost : 19.197382 Val Cost: 14.570190 PSNR :36.496151 26718 Epoch: 122 Cur Cost : 19.166422 Val Cost: 14.527482 PSNR :36.508900 26937 Epoch: 123 Cur Cost : 19.141092 Val Cost: 14.476220 PSNR :36.524252 27156 Epoch: 124 Cur Cost : 19.113115 Val Cost: 14.421227 PSNR :36.540782 27375 Epoch: 125 Cur Cost : 19.084183 Val Cost: 14.397281 PSNR :36.547999 27594 Epoch: 126 Cur Cost : 19.063801 Val Cost: 14.360329 PSNR :36.559160 27813 Epoch: 127 Cur Cost : 19.037209 Val Cost: 14.328735 PSNR :36.568725 28032 Epoch: 128 Cur Cost : 19.020245 Val Cost: 14.305547 PSNR :36.575759 28251 Epoch: 129 Cur Cost : 18.989275 Val Cost: 14.252780 PSNR :36.591808 28470 Epoch: 130 Cur Cost : 18.959911 Val Cost: 14.243009 PSNR :36.594786 28689 Epoch: 131 Cur Cost : 18.954180 Val Cost: 14.218270 PSNR :36.602336 28908 Epoch: 132 Cur Cost : 18.909533 Val Cost: 14.155800 PSNR :36.621459 29127 Epoch: 133 Cur Cost : 18.883560 Val Cost: 14.137563 PSNR :36.627058 29346 Epoch: 134 Cur Cost : 18.869696 Val Cost: 14.116048 PSNR :36.633672 29565 Epoch: 135 Cur Cost : 18.832462 Val Cost: 14.054341 PSNR :36.652699 29784 Epoch: 136 Cur Cost : 18.814505 Val Cost: 14.029723 PSNR :36.660313 30003 Epoch: 137 Cur Cost : 18.805397 Val Cost: 14.000203 PSNR :36.669460 30222 Epoch: 138 Cur Cost : 18.781073 Val Cost: 13.994488 PSNR :36.671234 30441 Epoch: 139 Cur Cost : 18.746599 Val Cost: 13.965572 PSNR :36.680216 30660 Epoch: 140 Cur Cost : 18.732073 Val Cost: 13.914907 PSNR :36.696000 30879 Epoch: 141 Cur Cost : 18.704937 Val Cost: 13.902081 PSNR :36.700005 31098 Epoch: 142 Cur Cost : 18.685591 Val Cost: 13.889931 PSNR :36.703803 31317 Epoch: 143 Cur Cost : 18.670761 Val Cost: 13.844592 PSNR :36.718002 31536 Epoch: 144 Cur Cost : 18.646767 Val Cost: 13.832062 PSNR :36.721934 31755 Epoch: 145 Cur Cost : 18.633806 Val Cost: 13.826068 PSNR :36.723817 31974 Epoch: 146 Cur Cost : 18.607412 Val Cost: 13.794584 PSNR :36.733717 32193 Epoch: 147 Cur Cost : 18.596941 Val Cost: 13.769987 PSNR :36.741468 32412 Epoch: 148 Cur Cost : 18.576069 Val Cost: 13.756820 PSNR :36.745623 32631 Epoch: 149 Cur Cost : 18.549088 Val Cost: 13.727606 PSNR :36.754856
In[ 13 ]
plt.plot(model.backprops_cnts,model.psnr)In[ 14 ]
model.generate_reconstruct_img('butterfly_GT.bmp')In[ 15 ]
model.generate_reconstruct_img('baby_GT.bmp')3通道彩色模型的训练结果
需要重启kernel
In[ 16 ]
# 先读取单通道训练结果，作为pre_train结果
def get_w(img_name):
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
img_test = cv2.imread('work/dataset/set5/%s' % img_name)
yuv_test = cv2.cvtColor(img_test, cv2.COLOR_BGR2YCrCb)
img_h, img_w, img_c = img_test.shape
# 图像模糊+cubic插值
img_blur = cv2.GaussianBlur(yuv_test.copy(), (5, 5), 0)
img_subsample = cv2.resize(img_blur, (img_w/3, img_h/3)) #这里注意cv2.resize里面的shape是w,h的顺序
img_cubic = cv2.resize(img_blur, (img_w, img_h), interpolation=cv2.INTER_CUBIC)
img_y, img_u, img_v = cv2.split(img_cubic)
img_input = np.reshape(img_y, [1,1,img_h, img_w]).astype("float32") # 把y通道作为输入
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model('work/model/', exe))
results = exe.run(inference_program,
feed={feed_target_names[0]: img_input},
fetch_list=fetch_targets)[0]
with fluid.program_guard(inference_program):
conv1_w_v = fluid.fetch_var('conv1_w')
conv1_b_v = fluid.fetch_var('conv1_b')
conv2_w_v = fluid.fetch_var('conv2_w')
conv2_b_v = fluid.fetch_var('conv2_b')
pred_w_v = fluid.fetch_var('pred_w')
pred_b_v = fluid.fetch_var('pred_b')
return conv1_w_v, conv1_b_v, conv2_w_v, conv2_b_v, pred_w_v, pred_b_v
conv1_w_v, conv1_b_v, conv2_w_v, conv2_b_v, pred_w_v, pred_b_v = get_w('butterfly_GT.bmp')
conv1_w_v.dump('work/model/conv1_w_v')
conv1_b_v.dump('work/model/conv1_b_v')
conv2_w_v.dump('work/model/conv2_w_v')
conv2_b_v.dump('work/model/conv2_b_v')
pred_w_v.dump('work/model/pred_w_v')
pred_b_v.dump('work/model/pred_b_v')
In[ 17 ]
class SRCNN_3dim(object):
def __init__(self, lr, lr_f, batch_size, iter_num):
self.lr = lr # 学习率
self.lr_f = lr_f # 最后一层学习率
self.batch_size = batch_size
self.iter_num = iter_num # 总共训练多少次
def net(self, X, Y):
# 搭建模型
conv1 = fluid.layers.conv2d(X, 64, 9,act='relu', name='conv1' ,
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='conv1_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='conv1_b'))
conv2 = fluid.layers.conv2d(conv1, 32, 1, act='relu', name='conv2' ,
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='conv2_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='conv2_b'))
pred = fluid.layers.conv2d(conv2, 3, 5, name='pred',
param_attr= fluid.ParamAttr(initializer=fluid.initializer.NormalInitializer(scale=0.001),
name='pred_w'),
bias_attr=fluid.ParamAttr(initializer=fluid.initializer.ConstantInitializer(value=0.),
name='pred_b'))
loss = fluid.layers.reduce_mean(fluid.layers.square(pred - Y))
return pred, loss
def train(self):
# 模型训练
X_train = fluid.layers.data(shape=[3, 33, 33], dtype='float32', name='image')
Y_train = fluid.layers.data(shape=[3, 21, 21], dtype='float32', name='gdt')
y_predict, y_loss = self.net(X_train, Y_train)
Optimizer = fluid.optimizer.AdamOptimizer(learning_rate=self.lr)
Optimizer_f = fluid.optimizer.AdamOptimizer(learning_rate=self.lr_f)
Optimizer.minimize(y_loss, parameter_list=['conv1_w','conv1_b', 'conv2_w', 'conv2_b'])
Optimizer_f.minimize(y_loss, parameter_list=['pred_w', 'pred_b'])
# 读取训练集数据
train_reader = paddle.batch(self.read_data('work/dataset/timofte'), batch_size=self.batch_size)
# 定义执行器
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
def train_loop(main_program):
feeder = fluid.DataFeeder(place=place, feed_list=[X_train, Y_train])
exe.run(fluid.default_startup_program())
# 用预训练的结果赋值
conv1_w = fluid.global_scope().find_var('conv1_w').get_tensor()
conv1_b = fluid.global_scope().find_var('conv1_b').get_tensor()
conv2_w = fluid.global_scope().find_var('conv2_w').get_tensor()
conv2_b = fluid.global_scope().find_var('conv2_b').get_tensor()
pred_w = fluid.global_scope().find_var('pred_w').get_tensor()
pred_b = fluid.global_scope().find_var('pred_b').get_tensor()
conv1_w.set(np.tile(conv1_w_v, (1,3,1,1)),place)
conv1_b.set(conv1_b_v,place)
conv2_w.set(conv2_w_v,place)
conv2_b.set(conv2_b_v,place)
pred_w.set(np.tile(pred_w_v,(3,1,1,1)),place)
pred_b.set(np.tile(pred_b_v,(3)),place)
backprops_cnt = 0 # 论文中作图的横坐标
self.backprops_cnts = [] # 绑定为类的一个属性，用于画图
self.psnr = [] # psnr的值
for epoch in range(self.iter_num):
for batch_id, data in enumerate(train_reader()):
loss = exe.run(
fluid.framework.default_main_program(),
feed=feeder.feed(data),
fetch_list=[y_loss])
if batch_id == 0: # 每个epoch算一下psnr，画图用的
# # 算psnr要在测试集上面
fluid.io.save_inference_model('work/model_3d/', ['image'], [y_predict], exe)
val_loss, val_psnr = self.validation()
self.backprops_cnts.append(backprops_cnt * self.batch_size)
self.psnr.append(val_psnr)
print("%i\tEpoch: %d \tCur Cost : %f\t Val Cost: %f\t PSNR :%f" % (backprops_cnt, epoch, np.array(loss[0])[0], val_loss, val_psnr))
backprops_cnt += 1
fluid.io.save_inference_model('work/model_3d/', ['image'], [y_predict], exe)
train_loop(fluid.default_main_program())
def validation(self):
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
test_set = 'work/dataset/set5/'
scale_factor = 3
for img_name in os.listdir(test_set):
img_val = cv2.imread(os.path.join(test_set, img_name))
img_h, img_w, _ = img_val.shape
img_blur = cv2.GaussianBlur(img_val, (5, 5), 0)
img_subsample = cv2.resize(img_blur, (img_w/scale_factor, img_h/scale_factor))
img_input = cv2.resize(img_blur, (img_w, img_h), interpolation=cv2.INTER_CUBIC)
img_input = np.swapaxes(img_input, 1, 2) # HWC->CHW
img_input = np.swapaxes(img_input, 0, 1)
img_input = np.reshape(img_input, [1,3, img_h, img_w]).astype("float32") # h,w
losses = []
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model('work/model_3d/', exe))
results = exe.run(inference_program,
feed={feed_target_names[0]: img_input},
fetch_list=fetch_targets)[0]
img_val = np.swapaxes(img_val, 1, 2) # HWC->CHW
img_val = np.swapaxes(img_val, 0, 1)
loss = np.mean(np.square(results[0]-img_val[:,6:-6, 6:-6]))
losses.append(loss)
avg_loss = np.sum(np.array(losses))/len(losses)
psnr = 10 * np.log10(255*255/avg_loss)
return avg_loss,psnr
def generate_reconstruct_img(self, img_name):
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
img_test = cv2.imread('work/dataset/set5/%s' % img_name)
print "=====原始图片========="
b,g,r = cv2.split(img_test) # AI studio 不支持cv2.imshow，所以用plt.imshow输出，两者rgb顺序不一样
img_show = cv2.merge([r,g,b])
plt.imshow(img_show)
plt.show()
# 图像模糊+cubic插值
img_h, img_w, _ = img_test.shape
img_blur = cv2.GaussianBlur(img_test, (5, 5), 0)
img_subsample = cv2.resize(img_blur, (img_w/3, img_h/3)) #这里注意cv2.resize里面的shape是w,h的顺序
img_cubic = cv2.resize(img_blur, (img_w, img_h), interpolation=cv2.INTER_CUBIC)
print "=====输入图片========="
b,g,r = cv2.split(img_cubic) # AI studio 不支持cv2.imshow，所以用plt.imshow输出，两者rgb顺序不一样
img_show = cv2.merge([r,g,b])
plt.imshow(img_show)
plt.show()
img_cubic = np.swapaxes(img_cubic, 1, 2) # HWC->CHW
img_cubic = np.swapaxes(img_cubic, 0, 1)
img_input = np.reshape(img_cubic, [1,3,img_h, img_w]).astype("float32") # 把RGB3通道作为输入
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model('work/model_3d/', exe))
results = exe.run(inference_program,
feed={feed_target_names[0]: img_input},
fetch_list=fetch_targets)[0]
result_img = np.array(results)
result_img[result_img < 0] = 0
result_img[result_img >255] = 255
gap_y = (img_test.shape[0]-result_img.shape[2])/2
gap_x = (img_test.shape[1]-result_img.shape[3])/2
result = np.swapaxes(result_img[0].copy(), 0,1) # CHW_>HWC
result = np.swapaxes(result, 1,2)
img_test[gap_y: gap_y + result_img.shape[2],
gap_x: gap_x + result_img.shape[3]]=result
print "=====彩图结果========="
b,g,r = cv2.split(img_test)
img_test_show = cv2.merge([r,g,b])
plt.imshow(img_test_show)
plt.show()
def read_data(self, data_path):
def data_reader():
for image in os.listdir(data_path):
if image.endswith('.bmp'):
img = cv2.imread(os.path.join(data_path, image))
# 下面是切图的步骤
j = 0
count = 0
while j+33 < len(img):
i = 0
while i+33 < len(img[0]):
img_patch = img[j:j+33, i:i+33, :]
img_gth = img_patch[6:27, 6:27].copy()
img_blur = cv2.GaussianBlur(img_patch, (5, 5), 0)
img_sumsample = cv2.resize(img_blur, (11, 11))
img_input = cv2.resize(img_blur, (33, 33), interpolation=cv2.INTER_CUBIC)
img_input = np.swapaxes(img_input, 1, 2) # HWC->CHW
img_input = np.swapaxes(img_input, 0, 1)
img_gth = np.swapaxes(img_gth, 1, 2) # HWC->CHW
img_gth = np.swapaxes(img_gth, 0, 1)
yield img_input, img_gth
i+=14
j+= 14
return data_readerIn[ 18 ]
conv1_w_v=np.load('work/model/conv1_w_v')
conv1_b_v=np.load('work/model/conv1_b_v')
conv2_w_v=np.load('work/model/conv2_w_v')
conv2_b_v=np.load('work/model/conv2_b_v')
pred_w_v=np.load('work/model/pred_w_v')
pred_b_v=np.load('work/model/pred_b_v')In[ 19 ]
model = SRCNN_3dim(0.0001, 0.00001, 100, 200)
model.train()In[ 20 ]
plt.plot(model.backprops_cnts,model.psnr)In[ 21 ]
model.generate_reconstruct_img('butterfly_GT.bmp')In[ 22 ]
model.generate_reconstruct_img('baby_GT.bmp')