add k-fold

dataloader.py

@@ -105,7 +105,8 @@ def loaddata_sleep_edf(opt,filedir,filenum,signal_name,BID = 'median',filter = T
         signals.append(eeg[events[i][0]:events[i][0]+3000])
     stages=np.array(stages)
     signals=np.array(signals)
-    signals = signals*13/np.median(np.abs(signals))
+    if BID == 'median':
+        signals = signals*13/np.median(np.abs(signals))
 
     # #select sleep time 
     if opt.select_sleep_time:
@@ -122,35 +123,6 @@ def loaddata_sleep_edf(opt,filedir,filenum,signal_name,BID = 'median',filter = T
             cnt += 1
     print('shape:',signals.shape,stages.shape)
 
-    '''
-    f_stage = pyedflib.EdfReader(os.path.join(filedir,f_stage_name))
-    annotations = f_stage.readAnnotations()
-    number_of_annotations = f_stage.annotations_in_file
-    end_duration = int(annotations[0][number_of_annotations-1])+int(annotations[1][number_of_annotations-1])
-    stages = np.zeros(end_duration//30, dtype=int)
-    # print(number_of_annotations)
-    for i in range(number_of_annotations):
-        stages[int(annotations[0][i])//30:(int(annotations[0][i])+int(annotations[1][i]))//30] = stage_str2int(annotations[2][i])
-
-    f_signal = pyedflib.EdfReader(os.path.join(filedir,f_signal_name))
-    signals = f_signal.readSignal(0)
-    signals=trimdata(signals,3000)
-    signals = signals.reshape(-1,3000)
-    stages = stages[0:signals.shape[0]]
-
-    # #select sleep time 
-    # signals = signals[np.clip(int(annotations[1][0])//30-60,0,9999999):int(annotations[0][number_of_annotations-2])//30+60]
-    # stages = stages[np.clip(int(annotations[1][0])//30-60,0,9999999):int(annotations[0][number_of_annotations-2])//30+60]
-    
-    #del UND
-    stages_copy = stages.copy()
-    cnt = 0
-    for i in range(len(stages_copy)):
-        if stages_copy[i] == 5 :
-            signals = np.delete(signals,i-cnt,axis =0)
-            stages = np.delete(stages,i-cnt,axis =0)
-            cnt += 1
-    '''
     return signals.astype(np.int16),stages.astype(np.int16)
 
 
@@ -168,7 +140,7 @@ def loaddataset(opt,filedir,dataset_name = 'CinC_Challenge_2018',signal_name = '
         for i,filename in enumerate(filenames[:num],0):
 
             try:
-                signal,stage = loaddata(os.path.join(filedir,filename),signal_name,BID)
+                signal,stage = loaddata(os.path.join(filedir,filename),signal_name,BID = None)
                 if i == 0:
                     signals =signal.copy()
                     stages =stage.copy()
@@ -187,7 +159,7 @@ def loaddataset(opt,filedir,dataset_name = 'CinC_Challenge_2018',signal_name = '
         cnt = 0
         for filename in filenames:
             if 'PSG' in filename:
-                signal,stage = loaddata_sleep_edf(opt,filedir,filename[2:6],signal_name = 'EEG Fpz-Cz')
+                signal,stage = loaddata_sleep_edf(opt,filedir,filename[2:6],signal_name = signal_name)
                 if cnt == 0:
                     signals =signal.copy()
                     stages =stage.copy()

options.py

@@ -2,6 +2,9 @@ import argparse
 import os
 import numpy as np
 import torch
+#python3 train.py --dataset_dir '/media/hypo/Hypo/physionet_org_train' --dataset_name CinC_Challenge_2018 --signal_name C4-M1 --sample_num 200 --model_name resnet18 --batchsize 32 --epochs 10 --fold_num 5 --pretrained
+#python3 train_new.py --dataset_dir '/media/hypo/Hypo/physionet_org_train' --dataset_name CinC_Challenge_2018 --signal_name C4-M1 --sample_num 10 --model_name LSTM --batchsize 32 --network_save_freq 100 --epochs 10
+#python3 train.py --dataset_dir '/media/hypo/Hypo/physionet_org_train' --dataset_name CinC_Challenge_2018 --signal_name C4-M1 --sample_num 10 --model_name resnet18 --batchsize 32
 #filedir = '/media/hypo/Hypo/physionet_org_train'
 # filedir ='E:\physionet_org_train'
 #python3 train.py --dataset_name sleep-edf --model_name resnet50 --batchsize 4 --epochs 50 --pretrained
@@ -13,17 +16,19 @@ class Options():
 
     def initialize(self):
         self.parser.add_argument('--no_cuda', action='store_true', help='if input, do not use gpu')
+        self.parser.add_argument('--no_cudnn', action='store_true', help='if input, do not use cudnn')
         self.parser.add_argument('--pretrained', action='store_true', help='if input, use pretrained models')
         self.parser.add_argument('--lr', type=float, default=0.001,help='learning rate')
+        self.parser.add_argument('--fold_num', type=int, default=5,help='k-fold')
         self.parser.add_argument('--batchsize', type=int, default=16,help='batchsize')
         self.parser.add_argument('--dataset_dir', type=str, default='./datasets/sleep-edfx/',
                                 help='your dataset path')
-        self.parser.add_argument('--dataset_name', type=str, default='sleep-edf',help='Choose dataset')
+        self.parser.add_argument('--dataset_name', type=str, default='sleep-edf',help='Choose dataset sleep-edf|sleep-edf|CinC_Challenge_2018|')
         self.parser.add_argument('--select_sleep_time', action='store_true', help='if input, for sleep-cassette only use sleep time to train')
         self.parser.add_argument('--signal_name', type=str, default='EEG Fpz-Cz',help='Choose the EEG channel C4-M1|EEG Fpz-Cz')
         self.parser.add_argument('--sample_num', type=int, default=20,help='the amount you want to load')
         self.parser.add_argument('--model_name', type=str, default='resnet18',help='Choose model')
-        self.parser.add_argument('--epochs', type=int, default=20,help='end epoch')
+        self.parser.add_argument('--epochs', type=int, default=50,help='end epoch')
         self.parser.add_argument('--weight_mod', type=str, default='avg_best',help='Choose weight mode: avg_best|normal')
         self.parser.add_argument('--network_save_freq', type=int, default=5,help='the freq to save network')
 

statistics.py

 import numpy as np
 import matplotlib.pyplot as plt
-
-def writelog(log):
-    f = open('./log','a+')
-    f.write(log+'\n')
-    # print(log)
+import util
 
 def stage(stages):	
     #N3->0  N2->1  N1->2  REM->3  W->4
@@ -12,6 +8,8 @@ def stage(stages):
     for i in range(len(stages)):
         stage_cnt[stages[i]] += 1
     stage_cnt_per = stage_cnt/len(stages) 
+    util.writelog('statistics of dataset [S3 S2 S1 R W]: '+str(stage_cnt))
+    print('statistics of dataset [S3 S2 S1 R W]:\n',stage_cnt,'\n',stage_cnt_per)
     return stage_cnt,stage_cnt_per
 
 def result(mat):
@@ -20,7 +18,10 @@ def result(mat):
     sub_recall = np.zeros(wide)
     err = 0
     for i in range(wide):
-        sub_recall[i]=mat[i,i]/np.sum(mat[i])
+        if np.sum(mat[i]) == 0 :
+            sub_recall[i] = 0
+        else:
+            sub_recall[i]=mat[i,i]/np.sum(mat[i])
         err += mat[i,i]
         sub_acc[i] = (np.sum(mat)-((np.sum(mat[i])+np.sum(mat[:,i]))-2*mat[i,i]))/np.sum(mat)
     avg_recall = np.mean(sub_recall)
@@ -28,6 +29,14 @@ def result(mat):
     err = 1-err/np.sum(mat)
     return avg_recall,avg_acc,err
 
+def stagefrommat(mat):
+    wide=mat.shape[0]
+    stage_num = np.zeros(wide,dtype='int')
+    for i in range(wide):
+        stage_num[i]=np.sum(mat[i])
+    util.writelog('statistics of dataset [S3 S2 S1 R W]:\n'+str(stage_num),True)
+
+
 
 def show(plot_result,epoch):
     train_recall = np.array(plot_result['train'])

train.py

@@ -17,59 +17,56 @@ warnings.filterwarnings("ignore")
 
 opt = Options().getparse()
 localtime = time.asctime(time.localtime(time.time()))
-statistics.writelog('\n\n'+str(localtime)+'\n'+str(opt))
+util.writelog('\n\n'+str(localtime)+'\n'+str(opt))
 
 t1 = time.time()
-signals,stages = dataloader.loaddataset(opt,opt.dataset_dir,opt.dataset_name,opt.signal_name,opt.sample_num,shuffle=True,BID='median')
-stage_cnt_per = statistics.stage(stages)[1]
-print('stage_cnt_per:',stage_cnt_per,'\nlength of dataset:',len(stages))
-signals_train,stages_train,signals_eval,stages_eval, = transformer.batch_generator(signals,stages,opt.batchsize,shuffle = True)
-
-batch_length = len(signals_train)+len(signals_eval)
+signals,stages = dataloader.loaddataset(opt,opt.dataset_dir,opt.dataset_name,opt.signal_name,opt.sample_num,shuffle=True,BID=None)
+stage_cnt,stage_cnt_per = statistics.stage(stages)
+signals,stages = transformer.batch_generator(signals,stages,opt.batchsize,shuffle = True)
+batch_length = len(stages)
 print('length of batch:',batch_length)
-show_freq = int(len(signals_train)/5)
+train_sequences,test_sequences = transformer.k_fold_generator(batch_length,opt.fold_num)
+
+show_freq = int(len(train_sequences[0])/5)
 util.show_menory()
 t2 = time.time()
-print('load data cost time:',t2-t1)
+print('load data cost time: %.2f'% (t2-t1),'s')
 
 net=models.CreatNet(opt.model_name)
-# print(net)
-if opt.pretrained:
-    net.load_state_dict(torch.load('./checkpoints/pretrained/'+opt.model_name+'.pth'))
+torch.save(net.cpu().state_dict(),'./checkpoints/'+opt.model_name+'.pth')
 
 weight = np.array([1,1,1,1,1])
 if opt.weight_mod == 'avg_best':
     weight = np.log(1/stage_cnt_per)
     weight[2] = weight[2]+1
-print(weight)
+    weight = np.clip(weight,1,5)
+print('Loss_weight:',weight)
 weight = torch.from_numpy(weight).float()
 # print(net)
 if not opt.no_cuda:
     net.cuda()
     weight = weight.cuda()
-    # cudnn.benchmark = True
-# print(weight)
-# time.sleep(2000)
+if not opt.no_cudnn:
+    cudnn.benchmark = True
+
 optimizer = torch.optim.Adam(net.parameters(), lr=opt.lr)
-# scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=2, gamma=0.95)
+# scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.8)
 criterion = nn.CrossEntropyLoss(weight)
 
-
-
-def evalnet(net,signals,stages,epoch,plot_result={},mode = 'part'):
-    net.eval()
+def evalnet(net,signals,stages,sequences,epoch,plot_result={},mode = 'part'):
+    # net.eval()
     if mode =='part':
         transformer.shuffledata(signals,stages)
         signals=signals[0:int(len(stages)/2)]
         stages=stages[0:int(len(stages)/2)]
 
     confusion_mat = np.zeros((5,5), dtype=int)
-    for i, (signal, stage) in enumerate(zip(signals,stages), 1):
+    for i, sequence in enumerate(sequences, 1):
 
-        signal=transformer.ToInputShape(signal,opt.model_name,test_flag =True)
-        signal,stage = transformer.ToTensor(signal,stage,no_cuda =opt.no_cuda)
-        out = net(signal)
-        loss = criterion(out, stage)
+        signal=transformer.ToInputShape(signals[sequence],opt.model_name,test_flag =True)
+        signal,stage = transformer.ToTensor(signal,stages[sequence],no_cuda =opt.no_cuda)
+        with torch.no_grad():
+            out = net(signal)
         pred = torch.max(out, 1)[1]
 
         pred=pred.data.cpu().numpy()
@@ -83,56 +80,71 @@ def evalnet(net,signals,stages,epoch,plot_result={},mode = 'part'):
         plot_result['test'].append(recall)   
         heatmap.draw(confusion_mat,name = 'test')
         print('test avg_recall:','%.4f' % recall,'avg_acc:','%.4f' % acc,'error:','%.4f' % error)
-        statistics.writelog('epoch:'+str(epoch)+'  test avg_recall:'+str(round(recall,4))+'  avg_acc:'+str(round(acc,4))+'  error:'+str(round(error,4)))
-    if epoch%1==0:
-        statistics.writelog('confusion_mat:\n'+str(confusion_mat))
-    # torch.cuda.empty_cache()
-    return plot_result
+        #util.writelog('epoch:'+str(epoch)+'  test avg_recall:'+str(round(recall,4))+'  avg_acc:'+str(round(acc,4))+'  error:'+str(round(error,4)))
+    return plot_result,confusion_mat
 
-
-plot_result={}
-plot_result['train']=[0]
-plot_result['test']=[0]
 print('begin to train ...')
-for epoch in range(opt.epochs):
-    t1 = time.time()
-    confusion_mat = np.zeros((5,5), dtype=int)
-    print('epoch:',epoch+1)
-    net.train()
-    for i, (signal, stage) in enumerate(zip(signals_train,stages_train), 1):
-
-        signal=transformer.ToInputShape(signal,opt.model_name,test_flag =False)
-        signal,stage = transformer.ToTensor(signal,stage,no_cuda =opt.no_cuda)
-
-        out = net(signal)
-        loss = criterion(out, stage)
-        pred = torch.max(out, 1)[1]
-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
-
-        pred=pred.data.cpu().numpy()
-        stage=stage.data.cpu().numpy()
-        for x in range(len(pred)):
-            confusion_mat[stage[x]][pred[x]] += 1
-        if i%show_freq==0:
-            # torch.cuda.empty_cache()          
-            plot_result['train'].append(statistics.result(confusion_mat)[0])
-            heatmap.draw(confusion_mat,name = 'train')
-            # plot_result=evalnet(net,signals_eval,stages_eval,plot_result,show_freq,mode = 'part')
-            statistics.show(plot_result,epoch+i/(batch_length*0.8))
-            confusion_mat[:]=0
-            # net.train()
-
-    # torch.cuda.empty_cache() 
-    evalnet(net,signals_eval,stages_eval,epoch+1,plot_result,mode = 'all')
-    # scheduler.step()
-
-    if (epoch+1)%opt.network_save_freq == 0:
-        torch.save(net.cpu().state_dict(),'./checkpoints/'+opt.model_name+'_epoch'+str(epoch+1)+'.pth')
-        print('network saved.')
-        if not opt.no_cuda:
-            net.cuda()
-
-    t2=time.time()
-    print('cost time: %.2f' % (t2-t1))
\ No newline at end of file
+final_confusion_mat = np.zeros((5,5), dtype=int)
+for fold in range(opt.fold_num):
+    net.load_state_dict(torch.load('./checkpoints/'+opt.model_name+'.pth'))
+    if opt.pretrained:
+        net.load_state_dict(torch.load('./checkpoints/pretrained/'+opt.model_name+'.pth'))
+    if not opt.no_cuda:
+        net.cuda()
+    plot_result={'train':[0],'test':[0]}
+    confusion_mats = []
+
+    for epoch in range(opt.epochs):
+        t1 = time.time()
+        confusion_mat = np.zeros((5,5), dtype=int)
+        print('fold:',fold+1,'epoch:',epoch+1)
+        net.train()
+        for i, sequence in enumerate(train_sequences[fold], 1):
+
+            signal=transformer.ToInputShape(signals[sequence],opt.model_name,test_flag =False)
+            signal,stage = transformer.ToTensor(signal,stages[sequence],no_cuda =opt.no_cuda)
+
+            out = net(signal)
+            loss = criterion(out, stage)
+            pred = torch.max(out, 1)[1]
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+            pred=pred.data.cpu().numpy()
+            stage=stage.data.cpu().numpy()
+            for x in range(len(pred)):
+                confusion_mat[stage[x]][pred[x]] += 1
+            if i%show_freq==0:       
+                plot_result['train'].append(statistics.result(confusion_mat)[0])
+                heatmap.draw(confusion_mat,name = 'train')
+                # plot_result=evalnet(net,signals_eval,stages_eval,plot_result,show_freq,mode = 'part')
+                statistics.show(plot_result,epoch+i/(batch_length*0.8))
+                confusion_mat[:]=0
+
+        plot_result,confusion_mat = evalnet(net,signals,stages,test_sequences[fold],epoch+1,plot_result,mode = 'all')
+        confusion_mats.append(confusion_mat)
+        # scheduler.step()
+
+        if (epoch+1)%opt.network_save_freq == 0:
+            torch.save(net.cpu().state_dict(),'./checkpoints/'+opt.model_name+'_epoch'+str(epoch+1)+'.pth')
+            print('network saved.')
+            if not opt.no_cuda:
+                net.cuda()
+
+        t2=time.time()
+        print('cost time: %.2f' % (t2-t1),'s')
+    pos = plot_result['test'].index(max(plot_result['test']))-1
+    final_confusion_mat = final_confusion_mat+confusion_mats[pos]
+    recall,acc,error  = statistics.result(confusion_mats[pos])
+    print('\nfold:',fold+1,'finished',' avg_recall:','%.4f' % recall,'avg_acc:','%.4f' % acc,'error:','%.4f' % error,'\n')
+    util.writelog('fold:'+str(fold+1)+'  test avg_recall:'+str(round(recall,4))+'  avg_acc:'+str(round(acc,4))+'  error:'+str(round(error,4)))
+    util.writelog('confusion_mat:\n'+str(confusion_mat))
+
+recall,acc,error  = statistics.result(final_confusion_mat)
+#print('all finished!\n',final_confusion_mat)
+#print('avg_recall:','%.4f' % recall,'avg_acc:','%.4f' % acc,'error:','%.4f' % error)
+util.writelog('final:'+'  test avg_recall:'+str(round(recall,4))+'  avg_acc:'+str(round(acc,4))+'  error:'+str(round(error,4)),True)
+util.writelog('confusion_mat:\n'+str(confusion_mat),True)
+statistics.stagefrommat(confusion_mat)
+heatmap.draw(final_confusion_mat,name = 'final_test')
\ No newline at end of file

transformer.py

@@ -15,17 +15,39 @@ def shuffledata(data,target):
     np.random.shuffle(target)
     # return data,target
 
-
 def batch_generator(data,target,batchsize,shuffle = True):
     if shuffle:
         shuffledata(data,target)
+    data = trimdata(data,batchsize)
+    target = trimdata(target,batchsize)
+    data = data.reshape(-1,batchsize,3000)
+    target = target.reshape(-1,batchsize)
+    return data,target
 
+def k_fold_generator(length,fold_num):
+    sequence = np.linspace(0,length-1,length,dtype='int')
+    train_length = int(length/fold_num*(fold_num-1))
+    test_length = int(length/fold_num)
+    train_sequence = np.zeros((fold_num,train_length), dtype = 'int')
+    test_sequence = np.zeros((fold_num,test_length), dtype = 'int')
+    for i in range(fold_num):
+        test_sequence[i] = (sequence[test_length*i:test_length*(i+1)])[:test_length]
+        train_sequence[i] = np.concatenate((sequence[0:test_length*i],sequence[test_length*(i+1):]),axis=0)[:train_length]
+    return train_sequence,test_sequence
+
+
+'''
+def batch_generator(data,target,batchsize,shuffle = True):
     data = trimdata(data,batchsize)
     target = trimdata(target,batchsize)
     data = data.reshape(-1,batchsize,3000)
     target = target.reshape(-1,batchsize)
-    return data[0:int(0.8*len(target))],target[0:int(0.8*len(target))],data[int(0.8*len(target)):],target[int(0.8*len(target)):]
-    
+    signals_train,stages_train,signals_eval,stages_eval = data[0:int(0.8*len(target))],target[0:int(0.8*len(target))],data[int(0.8*len(target)):],target[int(0.8*len(target)):]
+    if shuffle:
+        shuffledata(signals_train,stages_train)
+        shuffledata(signals_eval,stages_eval)
+    return signals_train,stages_train,signals_eval,stages_eval
+'''   
 def Normalize(data,maxmin,avg,sigma):
     data = np.clip(data, -maxmin, maxmin)
     return (data-avg)/sigma
@@ -67,14 +89,14 @@ def random_transform_2d(img,finesize = (224,122),test_flag = True):
         result = img[h_move:h_move+finesize[0],w_move:w_move+finesize[1]]
     else:
         #random crop
-        h_move = int(5*random.random()) #do not loss low freq signal infos
+        h_move = int(10*random.random()) #do not loss low freq signal infos
         w_move = int((w-finesize[1])*random.random())
         result = img[h_move:h_move+finesize[0],w_move:w_move+finesize[1]]
         #random flip
         if random.random()<0.5:
             result = result[:,::-1]
         #random amp
-        result = result*random.uniform(0.95,1.05)+random.uniform(-0.02,0.02)
+        result = result*random.uniform(0.9,1.1)+random.uniform(-0.05,0.05)
     return result
 
 
@@ -127,7 +149,7 @@ def ToInputShape(data,net_name,norm=True,test_flag = False):
         if norm:
             #sleep_def : std,mean,median = 0.4157 0.3688 0.2473
             #challge 2018 : std,mean,median,max= 0.2972 0.3008 0.2006 2.0830
-            result=Normalize(result,3,0.3,1)
+            result=Normalize(result,2,0.3,1)
         result = result.reshape(batchsize,1,224,122)
         # print(result.shape)
 

util.py

@@ -3,4 +3,10 @@ import memory_profiler
 def show_menory():
 	usage=int(memory_profiler.memory_usage()[0])
 	print('menory usage:',usage,'MB')
-	return usage
\ No newline at end of file
+	return usage
+
+def writelog(log,printflag = False):
+    f = open('./log','a+')
+    f.write(log+'\n')
+    if printflag:
+        print(log)
\ No newline at end of file