{"cells":[{"metadata":{"_uuid":"1c7b7a7b-d138-4c7a-b7e2-36bcd78b03d6","_cell_guid":"91afaa34-a976-4fdf-bef0-8399ee88738d","trusted":true},"cell_type":"code","source":"import numpy as np \nimport pandas as pd \nimport os\nimport torch\nimport torch.nn as nn\nfrom sklearn.model_selection import train_test_split \nfrom matplotlib import pyplot as plt\nfrom torchvision import transforms\nfrom torch.optim.lr_scheduler import ReduceLROnPlateau,StepLR\nimport csv\nimport time\nimport cv2\n\ntraindata=pd.read_csv('/kaggle/input/Kannada-MNIST/train.csv')\ntestdata=pd.read_csv('/kaggle/input/Kannada-MNIST/test.csv')\n\ntrainset=traindata.drop('label',axis=1)\ntrainlabel=traindata['label']\n\ntestset=testdata.drop('id',axis=1)\n\ntrain_data,validation_data,train_label,validation_label=train_test_split(trainset,trainlabel,test_size=0.01,stratify=trainlabel,random_state=42)\n\ntrain_data=train_data.to_numpy().astype(np.float32).reshape(len(train_data),28,28,1)\ntrain_label=torch.from_numpy(train_label.to_numpy())\nvalidation_data=validation_data.to_numpy().astype(np.float32).reshape(len(validation_data),28,28,1)\nvalidation_label=torch.from_numpy(validation_label.to_numpy())\ntest_data=testset.to_numpy().reshape(len(testset),28,28,1)\n\nnew_train_data=train_data[0][np.newaxis,:]\nnew_train_label=np.array([train_label[0]])\n\ndef imgtrans(im,datagroup,labelgroup,imglabel):\n global j\n img1=cv2.warpAffine(im,rot_mat1,(28,28),flags=cv2.INTER_LINEAR)[np.newaxis,:,:,np.newaxis]\n img3=cv2.warpAffine(im,scale_mat1,(28,28),flags=cv2.INTER_LINEAR)[np.newaxis,:,:,np.newaxis]\n img5=cv2.warpAffine(im,shift_mat1,(28,28),flags=cv2.INTER_LINEAR)[np.newaxis,:,:,np.newaxis]\n img2=cv2.warpAffine(im,rot_mat2,(28,28),flags=cv2.INTER_LINEAR)[np.newaxis,:,:,np.newaxis]\n img4=cv2.warpAffine(im,scale_mat2,(28,28),flags=cv2.INTER_LINEAR)[np.newaxis,:,:,np.newaxis]\n img6=cv2.warpAffine(im,shift_mat2,(28,28),flags=cv2.INTER_LINEAR)[np.newaxis,:,:,np.newaxis]\n im=im[np.newaxis,:,:,np.newaxis]\n #data=np.concatenate((datagroup[j],im,img1),axis=0)\n #label=np.concatenate((labelgroup[j],np.array([imglabel]),np.array([imglabel])),axis=0)\n data=np.concatenate((datagroup[j],im,img1,img2,img3,img4,img5,img6),axis=0)\n label=np.concatenate((labelgroup[j],np.array([imglabel]),np.array([imglabel]),np.array([imglabel]),np.array([imglabel]),np.array([imglabel]),np.array([imglabel]),np.array([imglabel])),axis=0)\n datagroup[j]=data\n labelgroup[j]=label\n \n#start=time.time()\nrot_mat1=cv2.getRotationMatrix2D((14,14),10,1) #旋转\nrot_mat2=cv2.getRotationMatrix2D((14,14),-10,1)\nscale_mat1=cv2.getRotationMatrix2D((14,14),0,1.1) #缩放\nscale_mat2=cv2.getRotationMatrix2D((14,14),0,0.9)\nshift_mat1= np.float32([[1,0,3],[0,1,3]]) #平移\nshift_mat2 = np.float32([[1,0,-3],[0,1,-3]])\nj=0\ndatadict={}\nfor i in range(13):\n datadict[i]=new_train_data\nlabeldict={}\nfor i in range(13):\n labeldict[i]=new_train_label\nstart=time.time()\nfor i in range(len(train_data)):\n img=train_data[i,:,:,0]\n imgtrans(img,datadict,labeldict,train_label[i])\n if i%5000==0:\n print(datadict[j].shape,labeldict[j].shape,time.time()-start)\n start=time.time()\n j+=1\n \nfor i in range(13):\n new_train_data=np.concatenate((new_train_data,datadict[i]),axis=0)\n new_train_label=np.concatenate((new_train_label,labeldict[i]),axis=0)\n \nnew_train_data=new_train_data[1:,:,:,:]\nnew_train_label=new_train_label[1:]\n\ntrain_data=new_train_data\ntrain_label=torch.from_numpy(new_train_label)\n\nnew_train_data=np.concatenate((train_data,test_data),axis=0)\nprint(new_train_data.shape)\n\n#fig,axes=plt.subplots(10,10,figsize=(10,10))\n#for i in range(10):\n# for j in range(10):\n# axes[i][j].axis('off')\n# axes[i][j].imshow(train_data[i*10+j,:,:,0])\n#fig.suptitle('train_data')\n\n#fig2,axes2=plt.subplots(10,10,figsize=(10,10))\n#for i in range(10):\n# for j in range(10):\n# axes2[i][j].axis('off')\n# axes2[i][j].imshow(test_data[i*10+j,:,:,0])\n#fig2.suptitle('test_data')","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"89df2c66-405d-4d8b-b892-f6c9718206a6","_cell_guid":"d41dad96-72f7-49ea-8353-b6edf4d55fee","trusted":true},"cell_type":"code","source":"\nclass mydataset(torch.utils.data.Dataset):\n def __init__(self,data,label=None):\n self.data=data\n self.label=label\n \n def __getitem__(self,index):\n img=self.data[index]\n img=transforms.ToTensor()(img)\n if self.label is not None:\n label=self.label[index]\n return img,label #train.val\n else:\n return img #test\n \n def __len__(self):\n return len(self.data)\ntrainset=mydataset(train_data,train_label)\ntrain_loader=torch.utils.data.DataLoader(trainset,batch_size=1024,shuffle=True)\nvalset=mydataset(validation_data,validation_label)\nval_loader=torch.utils.data.DataLoader(valset,batch_size=64,shuffle=False)\ntestset=mydataset(test_data)\ntest_loader=torch.utils.data.DataLoader(testset,batch_size=64,shuffle=False)\nprint(len(train_loader))\nprint(len(val_loader))\nprint(len(test_loader))","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"b82edf0a-405a-461e-9948-636814c29fdf","_cell_guid":"a14cc1f8-abce-4203-b7f4-7e70c1bd3b95","trusted":true},"cell_type":"code","source":"class maxout(nn.Module):\n def __init__(self):\n super(maxout,self).__init__()\n \n def forward(self,input):\n x1,x2=torch.chunk(input,2,1)\n output=torch.nn.functional.relu(x2-x1)+x1\n return output\n \nclass lenet(nn.Module):\n def __init__(self):\n super(lenet,self).__init__()\n self.features=nn.Sequential(\n nn.Conv2d(1,64,3), #26\n nn.BatchNorm2d(64),\n nn.LeakyReLU(0.1),\n nn.Conv2d(64,64,3), #24\n nn.BatchNorm2d(64),\n nn.LeakyReLU(0.1),\n nn.Conv2d(64,128,3), #22\n nn.BatchNorm2d(128),\n nn.LeakyReLU(0.1),\n nn.Conv2d(128,128,3), #20\n nn.BatchNorm2d(128),\n nn.LeakyReLU(0.1),\n nn.MaxPool2d(2,2), #10\n nn.Dropout(p=0.2),\n nn.Conv2d(128,256,3), #8\n nn.BatchNorm2d(256),\n nn.LeakyReLU(0.1),\n nn.Conv2d(256,256,3), #6\n nn.BatchNorm2d(256),\n nn.LeakyReLU(0.1),\n nn.MaxPool2d(2,2), #3\n nn.Dropout(p=0.2),\n )\n \n self.classify=nn.Sequential(\n nn.Linear(2304,512),\n nn.BatchNorm1d(512),\n nn.LeakyReLU(0.1),\n nn.Linear(512,10),\n )\n def forward(self,input):\n x=self.features(input)\n x=x.view(x.size(0),-1)\n x=self.classify(x)\n return x\nLenet=lenet()\nLenet.to('cuda')\nprint('building model......')","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"4fdf1b0e-bfe8-4642-8432-caf06258fcba","_cell_guid":"ac5d0917-f6e1-45d0-9d5f-d53e6760e605","trusted":true},"cell_type":"code","source":"print('train......')\nepoch=25\ncriterion=nn.CrossEntropyLoss()\noptimizer=torch.optim.SGD(Lenet.parameters(),lr=0.001,momentum=0.99,weight_decay=0.001)\n#lr_scheduler=ReduceLROnPlateau(optimizer,factor=0.5,patience=4)\nlr_scheduler=StepLR(optimizer,step_size=5,gamma=0.5)\n\n#train\nstart=time.time()\nfor i in range(epoch):\n Lenet.train()\n loss_total=0\n \n for idx,(img,label) in enumerate(train_loader):\n img=img.to('cuda')\n label=label.to('cuda')\n output=Lenet(img)\n loss=criterion(output,label)\n loss_total+=loss.item()\n optimizer.zero_grad()\n loss.backward()\n optimizer.step()\n \n if idx%100==0:\n print('epoch:{} , {}/{} , loss:{:.6f} {:.6f} , lr={}'.format(i,idx,len(train_loader),loss_total/(idx+1),loss.item(),optimizer.param_groups[0]['lr']))\n lr_scheduler.step()\nprint('time: {} s'.format(time.time()-start))","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"815f05c6-ea7f-4d32-8f53-8b751d5bf977","_cell_guid":"412d5043-4a76-4c4e-a5ac-688a27875f29","trusted":true},"cell_type":"code","source":"#validation\nprint('validate......')\nwith torch.no_grad():\n Lenet.eval()\n correct=0\n total=600\n for idx,(img,label) in enumerate(val_loader):\n img=img.to('cuda')\n label=label.to('cuda')\n output=Lenet(img)\n prediction=torch.max(output,dim=1)[1]\n correct+=sum(prediction==label).item()\n print('val accuracy:{:.6f}'.format(correct/total))","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"1fde06b5-1703-4709-bf47-d4043e57ae89","_cell_guid":"0602a0bb-c40a-40c8-b5a1-d31b6e52de2c","trusted":true},"cell_type":"code","source":"#pseudo label\nprint('pseudo label data......')\nnew_train_label=train_label.to('cuda')\n\nwith torch.no_grad():\n Lenet.eval()\n for idx,img in enumerate(test_loader):\n img=img.to('cuda')\n img=img.float()\n output=Lenet(img)\n _,predictions=torch.max(output,1)\n new_train_label=torch.cat((new_train_label,predictions),dim=0)\nprint(new_train_data.shape)\nprint(new_train_label.shape)","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"1c3b4db1-e687-4650-9afe-f6132c93c754","_cell_guid":"c00a2fab-8752-4d26-8590-853c84ffd782","trusted":true},"cell_type":"code","source":"#retrain\nprint('retrain with pseudo label......')\nnew_trainset=mydataset(new_train_data,new_train_label)\nnew_train_loader=torch.utils.data.DataLoader(new_trainset,batch_size=1024,shuffle=True)\n\nepoch=30\noptimizer=torch.optim.Adam(Lenet.parameters(),lr=0.001,weight_decay=0.001)\n#optimizer=torch.optim.SGD(Lenet.parameters(),lr=0.001,momentum=0.9,weight_decay=0.001)\nlr_scheduler=ReduceLROnPlateau(optimizer,mode='min',factor=0.5,patience=2)\n#lr_scheduler=StepLR(optimizer,step_size=5,gamma=0.5)\n\n#train\nbestnet=None\nbestloss=1\nstart=time.time()\nfor i in range(epoch):\n Lenet.train()\n loss_total=0\n \n for idx,(img,label) in enumerate(new_train_loader):\n img=img.to('cuda')\n label=label.to('cuda')\n img=img.float()\n output=Lenet(img)\n loss=criterion(output,label)\n loss_total+=loss.item()\n optimizer.zero_grad()\n loss.backward()\n optimizer.step()\n \n if idx%100==0:\n print('epoch:{} , {}/{} , loss:{:.6f} {:.6f} , lr={}'.format(i,idx,len(train_loader),loss_total/(idx+1),loss.item(),optimizer.param_groups[0]['lr']))\n if loss.item()