devbox
我家自动化
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

theano lasagna nolearn 实践,显著性检测深度学习_shallow and deep convolutional networks for salien

作者：我家自动化 | 2024-05-31 02:29:47

赞

踩

shallow and deep convolutional networks for saliency predictio

七小时入门theano

官方文档还是很值得参考的，

lasagne

theano


上面的源码是参考这个git的 

Shallow and Deep Convolutional Networks for Saliency Prediction

不过我改了训练集目前为止还没训练成功。。。就当是学习一下相关框架吧，以后如果训练成功了在更新

shallow-fig

数据预处理 
  1. import cv2
  2. import numpy as np
  3. import cPickle as pickle
  4.  
  5. #PIC_PATH='/home/zcb/datasets/trainSet/Stimuli/Alldata/'
  6. #SALMAP_PATH='/home/zcb/datasets/trainSet/FIXATIONMAPS/Alldata/'
  7.  
  8. PIC_PATH='/home/zcb/datasets/ecssd/images/'
  9. SALMAP_PATH='/home/zcb/datasets/ecssd/ground_truth_mask/'
  10.  
  11. datalist = open(PIC_PATH+'list.txt','r')
  12. namelist=[l.strip('\n') for l in datalist.readlines()]
  13.  
  14. sallist = open(SALMAP_PATH+'list.txt','r')
  15. sallist=[l.strip('\n') for l in sallist.readlines()]
  16.  
  17. input_h=48
  18. input_w=48
  19. output_h=24
  20. output_w=24
  21. #NumSample=len(namelist)
  22. NumSample=32
  23. X1 = np.zeros((NumSample, 3,input_h, input_w), dtype='float32')
  24. Y1 = np.zeros((NumSample, output_w*output_h), dtype='float32')
  25. print NumSample
  26.  
  27. for i in range(NumSample):
  28. 	img = cv2.imread(PIC_PATH+namelist[i], cv2.IMREAD_COLOR)
  29. 	img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
  30. 	#print img.shape
  31. 	img = cv2.resize(img,(input_w,input_h),interpolation=cv2.INTER_CUBIC)
  32. 	#print img.shape
  33. 	#cv2.imshow('show',img)
  34. 	img=img.astype(np.float32)
  35. 	img = img /255. 
  36. 	#img -= 1.
  37. 	if(cmp(img.shape , (input_h,input_w,3)) == 0):
  38. 		img = img.transpose(2,0,1).reshape(3, input_h, input_w)
  39. 		X1[i]=img
  40. 	else:
  41. 		print 'error'
  42. 	
  43. 	#np.set_printoptions(threshold='nan') 
  44. 	label = cv2.imread(SALMAP_PATH+sallist[i],cv2.IMREAD_GRAYSCALE)
  45. 	#label = loadSaliencyMapSUN(names[i])
  46. 	label = cv2.resize(label,(output_w,output_h),interpolation=cv2.INTER_CUBIC)
  47. 	#cv2.imshow('label',label)
  48. 	#cv2.waitKey(0)
  49. 	label = label.astype(np.float32)/255.
  50. 	#label = misc.imresize(label,(48,48)) / 
  51. 	#label = label -1.
  52. #	print 'data',X1[i]
  53. #	print 'label',label
  54. 	#Y1.append(label.reshape(1,48*48))
  55. 	Y1[i]=label.reshape(1,output_h*output_w)
  56.  
  57. data_to_save = (X1, Y1)
  58. f = file('data_ecssd_T.cPickle', 'wb')
  59. pickle.dump(data_to_save, f, protocol=pickle.HIGHEST_PROTOCOL)
  60. f.close()


训练代码
代码如下: 

  1. # add to kfkd.py
  2. import lasagne
  3. from lasagne.layers import *
  4. from lasagne.updates import nesterov_momentum
  5. from nolearn.lasagne import NeuralNet,BatchIterator
  6. import os
  7. import numpy as np
  8. from sklearn.utils import shuffle
  9. import cPickle as pickle
  10. import matplotlib.pyplot as plt
  11. import Image
  12. import ImageOps
  13. from scipy import misc
  14. import scipy.io
  15. import theano
  16. import sys
  17. sys.setrecursionlimit(10**8)  # set the maximum depth 
  18.  
  19. def load():
  20.     f = file('data_ecssd_T.cPickle', 'rb')
  21.     loaded_obj = pickle.load(f)
  22.     f.close()
  23.     X, y = loaded_obj
  24.     return X, y
  25.  
  26. def float32(k):
  27.     return np.cast['float32'](k)
  28.  
  29. class AdjustVariable(object):
  30.     def __init__(self, name, start=0.03, stop=0.001):
  31.         self.name = name
  32.         self.start, self.stop = start, stop
  33.         self.ls = None
  34.  
  35.     def __call__(self, nn, train_history):
  36.         if self.ls is None:
  37.             self.ls = np.linspace(self.start, self.stop, nn.max_epochs)
  38.         
  39.         epoch = train_history[-1]['epoch']
  40.         #if epoch > 200:
  41.         #    getattr(nn, self.name).set_value(0.01)
  42.         #if epoch > 800:
  43.         #    getattr(nn, self.name).set_value(0.005)
  44.         new_value = float32(self.ls[epoch - 1])
  45.         getattr(nn, self.name).set_value(new_value)
  46.  
  47. input_h=48
  48. input_w=48
  49. output_h=24
  50. output_w=24
  51.  
  52. class FlipBatchIterator(BatchIterator):
  53.     def transform(self, Xb, yb):
  54.         Xb, yb = super(FlipBatchIterator, self).transform(Xb, yb)
  55.  
  56.         # Flip half of the images in this batch at random:
  57.         bs = Xb.shape[0]
  58.         indices = np.random.choice(bs, bs / 2, replace=False)
  59.         Xb[indices] = Xb[indices, :, :, ::-1]
  60.         tmp =  yb[indices].reshape(bs/2,1,output_h,output_w)
  61.         mirror = tmp[ :,:,:, ::-1]
  62.         yb[indices] =  mirror.reshape(bs/2,output_w*output_h)
  63.         return Xb, yb
  64.  
  65. class EarlyStopping(object):
  66.     def __init__(self, patience=100):
  67.         self.patience = patience
  68.         self.best_valid = np.inf
  69.         self.best_valid_epoch = 0
  70.         self.best_weights = None
  71.  
  72.     def __call__(self, nn, train_history):
  73.         current_valid = train_history[-1]['valid_loss']
  74.         current_epoch = train_history[-1]['epoch']
  75.         if current_valid < self.best_valid:
  76.             self.best_valid = current_valid
  77.             self.best_valid_epoch = current_epoch
  78.             self.best_weights = nn.get_all_params_values()
  79.         elif self.best_valid_epoch + self.patience < current_epoch:
  80.             print("Early stopping.")
  81.             print("Best valid loss was {:.6f} at epoch {}.".format(
  82.                 self.best_valid, self.best_valid_epoch))
  83.             nn.load_params_from(self.best_weights)
  84.             raise StopIteration()
  85. def my_loss(a,b):
  86.     return theano.tensor.abs_(a-b)
  87.  
  88. layers0 = [
  89.     # layer dealing with the input data
  90.     (InputLayer, {'shape':(None, 3, input_h, input_w)}),
  91.  
  92.     # first stage of our convolutional layers
  93.     (Conv2DLayer, {'num_filters':32 , 'filter_size': 5,'pad':1,'W':lasagne.init.Uniform()}),
  94.     #(Conv2DLayer, {'num_filters':32 , 'filter_size': 7,'pad':1,'W':lasagne.init.Uniform()}),
  95.     (MaxPool2DLayer, {'pool_size': 2}),
  96.  
  97.     # second stage of our convolutional layers
  98.     (Conv2DLayer, {'num_filters':64 , 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  99.     #(Conv2DLayer, {'num_filters':16 , 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  100.     (MaxPool2DLayer, {'pool_size': 2}),
  101.  
  102.     # third stage of our convolutional layers
  103.     (Conv2DLayer, {'num_filters': 128, 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  104.     #(Conv2DLayer, {'num_filters': 32, 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  105.     (MaxPool2DLayer, {'pool_size': 2}),
  106.  
  107.         
  108.     #(Conv2DLayer, {'num_filters': 32, 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  109.     #(Conv2DLayer, {'num_filters': 32, 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  110.     #(Upscale2DLayer,{'scale_factor':2}),
  111.     
  112.     #(Conv2DLayer, {'num_filters': 16, 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  113.     #(Conv2DLayer, {'num_filters': 16, 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  114.     #(Upscale2DLayer,{'scale_factor':2}),
  115. 	#(Conv2DLayer, {'num_filters': 8, 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  116. 	#(Conv2DLayer, {'num_filters': 8, 'filter_size': 3,'pad':1,'W':lasagne.init.Uniform()}),
  117. 		
  118. 	#(Conv2DLayer, {'num_filters': 1, 'filter_size': 3,'pad':1,'nonlinearity':lasagne.nonlinearities.sigmoid}),
  119. 	#(Conv2DLayer, {'num_filters': 1, 'filter_size': 3,'pad':1,'nonlinearity':None}),
  120. 	#(FlattenLayer,{}),	
  121. 	
  122. 	# two dense layers with dropout
  123. 	#(DropoutLayer, {}),
  124.     (DenseLayer, {'num_units': output_w*output_h*2}),
  125.     (FeaturePoolLayer, {'pool_size':2}),
  126. 	# the output layer
  127.     #(DenseLayer, {'num_units': output_w*output_h,'nonlinearity':lasagne.nonlinearities.sigmoid}),
  128.     (DenseLayer, {'num_units': output_w*output_h,'nonlinearity':None}),
  129.     #(DenseLayer, {'num_units': 10, 'nonlinearity': softmax}),
  130. ]
  131.  
  132.  
  133.  
  134. net2 = NeuralNet(
  135.  
  136.  #    layers=[
  137.  #        ('input', layers.InputLayer),
  138.  #        ('conv1', layers.Conv2DLayer),
  139. 	# 	('pool1', layers.MaxPool2DLayer),
  140.  #        ('conv2', layers.Conv2DLayer),
  141.  #        ('pool2', layers.MaxPool2DLayer),
  142.  #        ('conv3', layers.Conv2DLayer),
  143.  #        ('pool3', layers.MaxPool2DLayer),
  144.  #        ('hidden4', layers.DenseLayer),
  145.  #        ('maxout6',layers.FeaturePoolLayer),
  146.  #        ('output', layers.DenseLayer),
  147.  
  148.  #        ],
  149.  #    input_shape=(None, 3, input_h, input_w),
  150.  #    #conv1
  151. 	# conv1_num_filters=16, conv1_filter_size=(5, 5),
  152. 	# conv1_nonlinearity=lasagne.nonlinearities.rectify,
  153. 	# conv1_W=lasagne.init.GlorotUniform(),
  154. 	# conv1_pad=1,
  155. 	# conv1_stride=1,
  156. 	# conv4_num_filters=16, conv4_filter_size=(3, 3),
  157. 	# conv4_nonlinearity=lasagne.nonlinearities.rectify,
  158. 	# conv4_W=lasagne.init.GlorotUniform(),
  159. 	# conv4_pad=1,
  160. 	# conv4_stride=1,
  161.  
  162. 	# #pool1
  163. 	# pool1_pool_size=(2, 2),
  164. 	# #conv2
  165.  #    conv2_num_filters=32, conv2_filter_size=(3, 3), 
  166. 	# conv2_nonlinearity=lasagne.nonlinearities.rectify,
  167. 	# conv2_W=lasagne.init.GlorotUniform(),
  168. 	# conv2_pad=1,
  169. 	# conv2_stride=1,
  170. 	# #pool2
  171. 	# pool2_pool_size=(2, 2),
  172.  #    #conv3
  173. 	# conv3_num_filters=32, conv3_filter_size=(3, 3), 
  174. 	# conv3_nonlinearity=lasagne.nonlinearities.rectify,
  175. 	# conv3_W=lasagne.init.GlorotUniform(),
  176. 	# conv3_pad=1,
  177. 	# #pool3
  178. 	# pool3_pool_size=(2, 2),
  179.     
  180. 	# hidden4_num_units=output_w*output_h*2, 
  181.  #    maxout6_pool_size=2,output_num_units=output_h*output_w,output_nonlinearity=None,
  182. 	layers=layers0,
  183. 	#update=lasagne.updates.rmsprop,
  184.     update_learning_rate=theano.shared(float32(0.05)),
  185.     update_momentum=theano.shared(float32(0.9)),
  186. 	
  187. 	#objective_loss_function=my_loss,#lasagne.objectives.squared_error,
  188.     regression=True,
  189.     on_epoch_finished=[
  190.         AdjustVariable('update_learning_rate', start=0.03, stop=0.0001),
  191.         AdjustVariable('update_momentum', start=0.9, stop=0.999),
  192.         #EarlyStopping(patience=150),
  193. 		],
  194.     batch_iterator_train=FlipBatchIterator(batch_size=128),
  195.     max_epochs=800,
  196. 	verbose=1,
  197.     )
  198.  
  199. X, y = load() 
  200. print net2.objective_loss_function,net2.train_split.eval_size
  201. #print("X.shape == {}; X.min == {:.3f}; X.max == {:.3f}".format(X.shape, X.min(), X.max()))
  202. #print("y.shape == {}; y.min == {:.3f}; y.max == {:.3f}".format(y.shape, y.min(), y.max()))
  203.  
  204. X = X.astype(np.float32)
  205. y = y.astype(np.float32)
  206. net2.fit(X, y)
  207.  
  208. with open('JuntingNet_SALICON.pickle', 'wb') as f:
  209.    pickle.dump(net2, f, -1)



声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/我家自动化/article/detail/649892
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号