def run_batch(datasetName, strategies, neurons, neuronLayersCount, maxpoolings,
fullLayers, batch_size, learning, train_decay, epoches,
following_strategy, trees, learning_sample_ratios,
dataset_format):
#for time_counts in range(experiment_times):
mix_model_svm_ratio = 0
#print strategies
print learning_sample_ratios
file_name, neighbors, raws_size, lines_size = data_util.prepare(
learning_sample_ratios, datasetName, int(strategies), 2)
neighbors = neighbors + 1
print "the neighbors strategy is: " + str(neighbors)
print "starting ..."
HSICNN.run_network(file_name, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, batch_size, learning,
train_decay, epoches)
print "the training of the network have done."
if following_strategy == 1:
print "now processing the cnn + svm joint framework..."
cnnsvm.run(file_name, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, batch_size, learning, train_decay)
elif following_strategy == 2:
print "now processing the cnn + rf joint framework..."
cnnrf.run(file_name, trees, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, raws_size, lines_size)
elif following_strategy == 3:
print "now processing the cnn + svm joint framework..."
cnnsvm.run(file_name, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, batch_size, learning, train_decay)
print "now processing the cnn + rf joint framework..."
cnnrf.run(file_name, trees, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, raws_size, lines_size, -1)
file = open(file_name + "_experiment_description.txt", 'w')
file.write("-------------Experiment Description-------------\n")
file.write("Data set:" + file_name + "#\n")
file.write("neighbor strategy:" + str(neighbors) + "#\n")
file.write("Convolutional Neurons:" + str(neurons) + "#\n")
file.write("Each Convolutional Neuron operate " + str(neuronLayersCount) +
'#\n')
file.write("Max Polling Kernel Size:" + str(maxpoolings) + "#\n")
file.write("Full Layer Neuron number:" + str(fullLayers) + "#\n")
file.write("Batch size of SGD training:" + str(batch_size) + "#\n")
file.write("Training epoches of deep CNN:" + str(epoches) + "#\n")
file.write("Learning ratio:" + str(learning) + "#\n")
file.write("Train decay:" + str(train_decay) + "#\n")
if following_strategy == 2 or following_strategy == 3:
file.write("Number of trees in random forest: " + str(trees) + "#\n")
file.write("===============================================\n")
file.close()
return file_name
def run_batch(datasetName,strategies, neurons, neuronLayersCount, maxpoolings, fullLayers, batch_size, learning, train_decay, epoches, following_strategy, trees, learning_sample_ratios, dataset_format):
#for time_counts in range(experiment_times):
mix_model_svm_ratio = 0
#print strategies
print learning_sample_ratios
file_name, neighbors, raws_size, lines_size = data_util.prepare(learning_sample_ratios, datasetName, int(strategies), 2)
neighbors = neighbors + 1
print "the neighbors strategy is: " + str(neighbors)
print "starting ..."
HSICNN.run_network(file_name, neurons,neuronLayersCount, neighbors, maxpoolings, fullLayers,batch_size, learning, train_decay, epoches)
print "the training of the network have done."
if following_strategy == 1:
print "now processing the cnn + svm joint framework..."
cnnsvm.run(file_name, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, batch_size, learning, train_decay)
elif following_strategy == 2:
print "now processing the cnn + rf joint framework..."
cnnrf.run(file_name,trees, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, raws_size, lines_size)
elif following_strategy == 3:
print "now processing the cnn + svm joint framework..."
cnnsvm.run(file_name, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, batch_size, learning, train_decay)
print "now processing the cnn + rf joint framework..."
cnnrf.run(file_name, trees, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, raws_size,lines_size,-1)
file = open(file_name + "_experiment_description.txt", 'w')
file.write("-------------Experiment Description-------------\n")
file.write("Data set:" + file_name + "#\n")
file.write("neighbor strategy:" + str(neighbors) + "#\n")
file.write("Convolutional Neurons:" + str(neurons) + "#\n")
file.write("Each Convolutional Neuron operate " + str(neuronLayersCount) + '#\n')
file.write("Max Polling Kernel Size:" + str(maxpoolings) + "#\n")
file.write("Full Layer Neuron number:" + str(fullLayers) + "#\n")
file.write("Batch size of SGD training:" + str(batch_size) + "#\n")
file.write("Training epoches of deep CNN:" + str(epoches) + "#\n")
file.write("Learning ratio:" + str(learning) + "#\n")
file.write("Train decay:" + str(train_decay) +"#\n")
if following_strategy == 2 or following_strategy == 3:
file.write("Number of trees in random forest: " + str(trees) + "#\n")
file.write("===============================================\n")
file.close()
return file_name
def complete_operate(operate_type, folder_path, trees, neurons,
neuronLayersCount, maxpoolings, fullLayers):
#TODO 添加增强健壮性的代码
if os.path.isdir(folder_path):
if (operate_type != '5'):
print "Processing " + folder_path
#首先根据目录名字拿到数据集的文件名
file_name_split = folder_path.split('/')
file_or_folder_name = file_name_split[len(file_name_split) - 1]
dataset_name_sub = file_or_folder_name.split('_')
#TODO：添加判断目录是否为有效实验结果的代码，有时候非实验结果的文件夹会不小心混在目录中，应该予以剔除
# print dataset_name_sub
dataset_name = dataset_name_sub[0] + '_' + dataset_name_sub[
1] + '_' + dataset_name_sub[2]
#搜寻该文件夹中的网络配置文件
#TODO:统一的网络配置文件格式
#查询已有框架自动恢复网络结构的方式？？？
#TODO:与下面的else配套
#neurons = 0
#neuronLayersCount = 0
neighbors = int(dataset_name_sub[1]) + 1
#maxpoolings = 0
#fullLayers = 0
raws = 0
lines = 0
batch_size = 100
ratio = 0.001
decay = 0.00001
LabelsMat = sio.loadmat(data_prefix + dataset_name_sub[0] + '/' +
dataset_name_sub[0] + 'Gt.mat')
key_label_name = LabelsMat.keys()
label_key = ''
for temp_key in key_label_name:
if temp_key != '__version__' and temp_key != '__header__' and temp_key != '__globals__':
label_key = temp_key
break
Labels = LabelsMat[label_key]
raws = int(len(Labels))
lines = int(len(Labels[0]))
if operate_type == '1':
print 'CNN+RF base on trained CNN model.'
if len(trees.split('-')) > 0:
small_tree_number = trees.split('-')[0]
big_tree_number = trees.split('-')[1]
exp_trees = range(int(small_tree_number),
int(big_tree_number) + 1)
cnnrf_acc_list = []
rf_acc_list = []
for trees_number in exp_trees:
print "Tree Number:" + str(trees_number)
cnnrf_acc, rf_acc = cnnrf.run(folder_path + "/" +
dataset_name,
trees_number,
int(neurons),
int(neuronLayersCount),
int(neighbors),
int(maxpoolings),
int(fullLayers),
raws,
lines,
test_cnn=-1)
# cnnrf.run(file_name,trees, neurons, neuronLayersCount, /neighbors, maxpoolings, fullLayers, raws_size, lines_size)
cnnrf_acc_list.append(cnnrf_acc)
rf_acc_list.append(rf_acc)
sio.savemat(folder_path + "/ALL_CNN_RF_EXP_RESULT.mat", {
'cnnrf': cnnrf_acc_list,
'rf': rf_acc_list
})
else:
cnnrf.run(dataset_name, trees_number, neurons,
neuronLayersCount, neighbors, maxpoolings,
fullLayers, 0, 0, 0, raws, lines)
elif operate_type == '2':
print 'Rebuild CNN_+ SVM model experiments...'
cnnsvm.run(folder_path + "/" + dataset_name, neurons,
neuronLayersCount, neighbors, maxpoolings, fullLayers,
1, 1, 1)
elif operate_type == '3':
print 'TODO'
elif operate_type == '4':
print 'TODO'
# logic to compute the OA, AA and kappa index.
elif operate_type == '5':
#print 'Computing OA, AA and Kappa for ' + folder_path
train_ratio = 0
neighbor_strategy = 0
folder_name_current = folder_path.split('_')
dataset_name = folder_name_current[0].split('/')[3]
strategy_current = folder_name_current[1]
training_ratio_current = folder_name_current[2]
current_dataset = dataset_name + "_" + strategy_current + "_" + training_ratio_current
cnn_result = current_dataset + "CNNOnlyResult.mat"
ccr_result = current_dataset + "CNNRFResult.mat"
ccr_result_backup = current_dataset + "_trees_200_CNNRFResult.mat"
ccs_result = current_dataset + "CNNSVMResult.mat"
rf_result = current_dataset + "RFonlyResult.mat"
svm_result = current_dataset + "SVMonlyResult.mat"
results_index = [
cnn_result, ccr_result, ccr_result_backup, ccs_result,
rf_result, svm_result
]
#检查并创建sum_up_results目录用于存放所有文件
sum_up_result_folder = "../experiments/" + dataset_name + "/SumResults"
if not os.path.exists(sum_up_result_folder):
os.makedirs(sum_up_result_folder)
for temp_index in results_index:
if os.path.exists(folder_path + "/" + temp_index):
#开始取出结果并汇总
data = sio.loadmat(folder_path + "/" + temp_index)
actual = data['actual'][0]
predict = data['predict'][0]
file_name = sum_up_result_folder + "/" + temp_index
if os.path.exists(file_name):
data_old = sio.loadmat(file_name)
actual_old = data_old['actual'][0]
predict_old = data_old['predict'][0]
actual_new = np.append(actual_old, actual)
predict_new = np.append(predict_old, predict)
sio.savemat(file_name, {
'actual': actual_new,
'predict': predict_new
},
appendmat=False)
else:
sio.savemat(file_name, {
'actual': actual,
'predict': predict
})
else:
print 'Not under selection list, skip it.'
else:
print "not a folder, skipt it."
def run_single(learning_ratio, network_type):
prompt = '>'
#mix_model_svm_ratio是为了以后采用组合混合模型的时候，保存一个svm在所有模型中的占比。以后根据需求进行扩充。。。TODO
mix_model_svm_ratio = 0
file_name, neighbors, raws_size, lines_size = data_util.prepare(
learning_ratio, "NONE", 0, 2)
print "now constructing the network..."
#print "enter the layers each convolutional kernel covers: "
#neighbors = int(raw_input(prompt))
neighbors = neighbors + 1
print "the neighbors strategy is: " + str(neighbors)
print "enter the number of convolutional neurons:"
neurons = int(raw_input(prompt))
print "enter the number of layers you want the CNN to operate convolutional operation:"
neuronLayersCount = int(raw_input(prompt))
print "enter the kernel size of the maxpooling layer:"
maxpoolings = int(raw_input(prompt))
print "enter the number of full layers\' neurons, default is 100:"
fullLayers = int(raw_input(prompt))
#if tempfullLayers > 1:
# fullLayers = tempfullLayers
print "enter the batch size for bsgd:"
batch_size = int(raw_input(prompt))
print "enter the learning ratio:"
learning = float(raw_input(prompt))
print "enter the train decay:"
train_decay = float(raw_input(prompt))
print "enter the epoches you want the network to be trained:"
epoches = int(raw_input(prompt))
print "now choose the following strategy after the cnn network been trained:"
print "#1:train a cnn-svm joint framework;"
print "#2:train a cnn-rf joint framework;"
print "#3:train both cnn-svm and cnn-rf joint frameworks;"
print "#4:TODO: train a mix assemble cnn-classifier model."
print "#5: train a CNN model only."
following_strategy = int(raw_input(prompt))
if following_strategy == 4:
print "enter the ratio of svm classifier:"
mix_model_svm_ratio = int(row_input(prompt))
tress = 0
if following_strategy == 2 or following_strategy == 3:
print "enter the count of trees you want to set in Random Forest:"
trees = int(raw_input(prompt))
print "starting ..."
if network_type == '1':
HSICNN.run_network(file_name, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, batch_size, learning,
train_decay, epoches)
elif network_type == '2':
hic.run_network(file_name, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, batch_size, learning,
train_decay, epoches)
print "the training of the network have done."
if following_strategy == 1 and following_strategy != 5:
#CNN + SVM
print "now processing the cnn + svm joint framework..."
cnnsvm.run(file_name, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, batch_size, learning, train_decay)
elif following_strategy == 2 and following_strategy != 5:
#CNN + rfind
print "now processing the cnn + rf joint framework..."
# print "enter the count of trees you want to set in Random Forest:"
# trees = int(raw_input(prompt))
cnnrf.run(file_name, trees, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, raws_size, lines_size)
elif following_strategy == 3 != following_strategy != 5:
#CNN+svm and CNN+RF
print "now processing the cnn + svm joint framework..."
cnnsvm.run(file_name, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, batch_size, learning, train_decay)
print "now processing the cnn + rf joint framework..."
cnnrf.run(file_name, trees, neurons, neuronLayersCount, neighbors,
maxpoolings, fullLayers, raws_size, lines_size)
file = open(file_name + "_experiment_description.txt", 'w')
file.write("-------------Experiment Description-------------\n")
file.write("Data set:" + file_name + "#\n")
file.write("neighbor strategy:" + str(neighbors) + "#\n")
file.write("Convolutional Neurons:" + str(neurons) + "#\n")
file.write("Each convolutional neuron operates " + str(neuronLayersCount))
file.write("Max Polling Kernel Size:" + str(maxpoolings) + "#\n")
file.write("Full Layer Neuron number:" + str(fullLayers) + "#\n")
file.write("Batch size of SGD training:" + str(batch_size) + "#\n")
file.write("Training epoches of deep CNN:" + str(epoches) + "#\n")
file.write("Learning ratio:" + str(learning) + "#\n")
file.write("Train decay:" + str(train_decay) + "#\n")
if following_strategy == 2 or following_strategy == 3:
file.write("Number of trees in random forest: " + str(trees) + "#\n")
file.write("===============================================\n")
file.close()
return file_name
def complete_operate(operate_type, folder_path, trees, neurons, neuronLayersCount,maxpoolings, fullLayers):
#TODO 添加增强健壮性的代码
if os.path.isdir(folder_path):
if(operate_type != '5'):
print "Processing " + folder_path
#首先根据目录名字拿到数据集的文件名
file_name_split = folder_path.split('/')
file_or_folder_name = file_name_split[len(file_name_split) - 1]
dataset_name_sub = file_or_folder_name.split('_')
#TODO：添加判断目录是否为有效实验结果的代码，有时候非实验结果的文件夹会不小心混在目录中，应该予以剔除
# print dataset_name_sub
dataset_name = dataset_name_sub[0] + '_' + dataset_name_sub[1] + '_' + dataset_name_sub[2]
#搜寻该文件夹中的网络配置文件
#TODO:统一的网络配置文件格式
#查询已有框架自动恢复网络结构的方式？？？
#TODO:与下面的else配套
#neurons = 0
#neuronLayersCount = 0
neighbors = int(dataset_name_sub[1]) + 1
#maxpoolings = 0
#fullLayers = 0
raws = 0
lines = 0
batch_size = 100
ratio = 0.001
decay = 0.00001
LabelsMat = sio.loadmat(data_prefix + dataset_name_sub[0] + '/' + dataset_name_sub[0] + 'Gt.mat')
key_label_name = LabelsMat.keys()
label_key = ''
for temp_key in key_label_name:
if temp_key != '__version__' and temp_key != '__header__' and temp_key != '__globals__':
label_key = temp_key
break
Labels = LabelsMat[label_key]
raws = int(len(Labels))
lines = int(len(Labels[0]))
if operate_type == '1':
print 'CNN+RF base on trained CNN model.'
if len(trees.split('-')) > 0:
small_tree_number = trees.split('-')[0]
big_tree_number = trees.split('-')[1]
exp_trees = range(int(small_tree_number),int(big_tree_number) + 1)
cnnrf_acc_list = []
rf_acc_list = []
for trees_number in exp_trees:
print "Tree Number:" + str(trees_number)
cnnrf_acc, rf_acc = cnnrf.run(folder_path + "/" + dataset_name,trees_number, int(neurons), int(neuronLayersCount), int(neighbors), int(maxpoolings), int(fullLayers), raws, lines,test_cnn = -1)
# cnnrf.run(file_name,trees, neurons, neuronLayersCount, /neighbors, maxpoolings, fullLayers, raws_size, lines_size)
cnnrf_acc_list.append(cnnrf_acc)
rf_acc_list.append(rf_acc)
sio.savemat(folder_path + "/ALL_CNN_RF_EXP_RESULT.mat",{'cnnrf':cnnrf_acc_list, 'rf':rf_acc_list})
else:
cnnrf.run(dataset_name,trees_number, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, 0, 0, 0, raws, lines)
elif operate_type == '2':
print 'Rebuild CNN_+ SVM model experiments...'
cnnsvm.run(folder_path + "/" + dataset_name, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, 1, 1, 1)
elif operate_type == '3':
print 'TODO'
elif operate_type == '4':
print 'TODO'
# logic to compute the OA, AA and kappa index.
elif operate_type == '5':
#print 'Computing OA, AA and Kappa for ' + folder_path
train_ratio = 0
neighbor_strategy = 0
folder_name_current = folder_path.split('_')
dataset_name = folder_name_current[0].split('/')[3]
strategy_current = folder_name_current[1]
training_ratio_current = folder_name_current[2]
current_dataset = dataset_name + "_" + strategy_current + "_" + training_ratio_current
cnn_result = current_dataset + "CNNOnlyResult.mat"
ccr_result = current_dataset + "CNNRFResult.mat"
ccr_result_backup = current_dataset + "_trees_200_CNNRFResult.mat"
ccs_result = current_dataset + "CNNSVMResult.mat"
rf_result = current_dataset + "RFonlyResult.mat"
svm_result = current_dataset + "SVMonlyResult.mat"
results_index = [cnn_result, ccr_result, ccr_result_backup, ccs_result, rf_result, svm_result]
#检查并创建sum_up_results目录用于存放所有文件
sum_up_result_folder = "../experiments/" + dataset_name + "/SumResults"
if not os.path.exists(sum_up_result_folder):
os.makedirs(sum_up_result_folder)
for temp_index in results_index:
if os.path.exists(folder_path + "/" + temp_index):
#开始取出结果并汇总
data = sio.loadmat(folder_path + "/" + temp_index)
actual = data['actual'][0]
predict = data['predict'][0]
file_name = sum_up_result_folder + "/" + temp_index
if os.path.exists(file_name):
data_old = sio.loadmat(file_name)
actual_old = data_old['actual'][0]
predict_old = data_old['predict'][0]
actual_new = np.append(actual_old, actual)
predict_new = np.append(predict_old, predict)
sio.savemat(file_name,{'actual':actual_new,'predict':predict_new},appendmat=False)
else:
sio.savemat(file_name,{'actual':actual, 'predict':predict})
else:
print 'Not under selection list, skip it.'
else:
print "not a folder, skipt it."
def run_single(learning_ratio, network_type):
prompt = '>'
#mix_model_svm_ratio是为了以后采用组合混合模型的时候，保存一个svm在所有模型中的占比。以后根据需求进行扩充。。。TODO
mix_model_svm_ratio = 0
file_name, neighbors, raws_size, lines_size = data_util.prepare(learning_ratio,"NONE", 0, 2)
print "now constructing the network..."
#print "enter the layers each convolutional kernel covers: "
#neighbors = int(raw_input(prompt))
neighbors = neighbors + 1
print "the neighbors strategy is: " + str(neighbors)
print "enter the number of convolutional neurons:"
neurons = int(raw_input(prompt))
print "enter the number of layers you want the CNN to operate convolutional operation:"
neuronLayersCount = int(raw_input(prompt))
print "enter the kernel size of the maxpooling layer:"
maxpoolings = int(raw_input(prompt))
print "enter the number of full layers\' neurons, default is 100:"
fullLayers = int(raw_input(prompt))
#if tempfullLayers > 1:
# fullLayers = tempfullLayers
print "enter the batch size for bsgd:"
batch_size = int(raw_input(prompt))
print "enter the learning ratio:"
learning = float(raw_input(prompt))
print "enter the train decay:"
train_decay = float(raw_input(prompt))
print "enter the epoches you want the network to be trained:"
epoches = int(raw_input(prompt))
print "now choose the following strategy after the cnn network been trained:"
print "#1:train a cnn-svm joint framework;"
print "#2:train a cnn-rf joint framework;"
print "#3:train both cnn-svm and cnn-rf joint frameworks;"
print "#4:TODO: train a mix assemble cnn-classifier model."
print "#5: train a CNN model only."
following_strategy = int(raw_input(prompt))
if following_strategy == 4:
print "enter the ratio of svm classifier:"
mix_model_svm_ratio = int(row_input(prompt))
tress = 0
if following_strategy == 2 or following_strategy == 3:
print "enter the count of trees you want to set in Random Forest:"
trees = int(raw_input(prompt))
print "starting ..."
if network_type == '1':
HSICNN.run_network(file_name, neurons,neuronLayersCount, neighbors, maxpoolings, fullLayers,batch_size, learning, train_decay, epoches)
elif network_type == '2':
hic.run_network(file_name, neurons,neuronLayersCount, neighbors, maxpoolings, fullLayers,batch_size, learning, train_decay, epoches)
print "the training of the network have done."
if following_strategy == 1 and following_strategy != 5:
#CNN + SVM
print "now processing the cnn + svm joint framework..."
cnnsvm.run(file_name, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, batch_size, learning, train_decay)
elif following_strategy == 2 and following_strategy != 5:
#CNN + rfind
print "now processing the cnn + rf joint framework..."
# print "enter the count of trees you want to set in Random Forest:"
# trees = int(raw_input(prompt))
cnnrf.run(file_name,trees, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, raws_size, lines_size)
elif following_strategy == 3 != following_strategy != 5:
#CNN+svm and CNN+RF
print "now processing the cnn + svm joint framework..."
cnnsvm.run(file_name, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, batch_size, learning, train_decay)
print "now processing the cnn + rf joint framework..."
cnnrf.run(file_name, trees, neurons, neuronLayersCount, neighbors, maxpoolings, fullLayers, raws_size, lines_size)
file = open(file_name + "_experiment_description.txt", 'w')
file.write("-------------Experiment Description-------------\n")
file.write("Data set:" + file_name + "#\n")
file.write("neighbor strategy:" + str(neighbors) + "#\n")
file.write("Convolutional Neurons:" + str(neurons) + "#\n")
file.write("Each convolutional neuron operates " + str(neuronLayersCount))
file.write("Max Polling Kernel Size:" + str(maxpoolings) + "#\n")
file.write("Full Layer Neuron number:" + str(fullLayers) + "#\n")
file.write("Batch size of SGD training:" + str(batch_size) + "#\n")
file.write("Training epoches of deep CNN:" + str(epoches) + "#\n")
file.write("Learning ratio:" + str(learning) + "#\n")
file.write("Train decay:" + str(train_decay) +"#\n")
if following_strategy == 2 or following_strategy == 3:
file.write("Number of trees in random forest: " + str(trees) + "#\n")
file.write("===============================================\n")
file.close()
return file_name