{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"<b>主要是看一看剪枝前后的效果,不要太看重准确率.</b>","metadata":{}},{"cell_type":"code","source":"import pandas as pd#导入csv文件的库\nimport math\nimport numpy as np#进行矩阵运算的库\nimport matplotlib.pyplot as plt#作图的库\nimport torch#一个深度学习的库Pytorch\nimport torch.nn as nn#neural network,神经网络\nimport torch.optim as optim#一个实现了各种优化算法的库\nimport torch.nn.functional as F#神经网络函数库\nimport warnings#避免一些可以忽略的报错\nwarnings.filterwarnings('ignore')","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2023-07-06T13:02:32.749066Z","iopub.execute_input":"2023-07-06T13:02:32.749534Z","iopub.status.idle":"2023-07-06T13:02:37.049063Z","shell.execute_reply.started":"2023-07-06T13:02:32.749507Z","shell.execute_reply":"2023-07-06T13:02:37.047912Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#设置随机种子\nimport random\ntorch.backends.cudnn.deterministic = True#将cudnn框架中的随机数生成器设为确定性模式\ntorch.backends.cudnn.benchmark = False#关闭CuDNN框架的自动寻找最优卷积算法的功能，以避免不同的算法对结果产生影响\ntorch.manual_seed(42)\nnp.random.seed(42)\nrandom.seed(42)","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:02:37.052793Z","iopub.execute_input":"2023-07-06T13:02:37.053744Z","iopub.status.idle":"2023-07-06T13:02:37.065628Z","shell.execute_reply.started":"2023-07-06T13:02:37.053708Z","shell.execute_reply":"2023-07-06T13:02:37.064470Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train=pd.read_csv(\"/kaggle/input/ece597697-sp2023/train.csv/train.csv\")\ntrain.head()","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:02:37.070850Z","iopub.execute_input":"2023-07-06T13:02:37.073220Z","iopub.status.idle":"2023-07-06T13:02:43.431501Z","shell.execute_reply.started":"2023-07-06T13:02:37.073185Z","shell.execute_reply":"2023-07-06T13:02:43.430585Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"datay=train['label'].values\ntrain=train.drop(['label','id'],axis=1)\ndataX=train.values","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:02:43.434478Z","iopub.execute_input":"2023-07-06T13:02:43.434833Z","iopub.status.idle":"2023-07-06T13:02:43.551277Z","shell.execute_reply.started":"2023-07-06T13:02:43.434800Z","shell.execute_reply":"2023-07-06T13:02:43.550034Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#划分训练集和测试集的函数\ndef train_test_split(dataX,datay,shuffle=True,percentage=0.8):\n    \"\"\"\n    将训练数据X和标签y以numpy.array数组的形式传入\n    划分的比例定为训练集:测试集=8:2\n    \"\"\"\n    if shuffle:\n        random_num=[index for index in range(len(dataX))]\n        np.random.shuffle(random_num)\n        dataX=dataX[random_num]\n        datay=datay[random_num]\n    split_num=int(len(dataX)*percentage)\n    train_X=dataX[:split_num]\n    train_y=datay[:split_num]\n    test_X=dataX[split_num:]\n    test_y=datay[split_num:]\n    return train_X,train_y,test_X,test_y","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:02:43.552962Z","iopub.execute_input":"2023-07-06T13:02:43.553374Z","iopub.status.idle":"2023-07-06T13:02:43.562126Z","shell.execute_reply.started":"2023-07-06T13:02:43.553338Z","shell.execute_reply":"2023-07-06T13:02:43.561075Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_X,train_y,valid_X,valid_y=train_test_split(dataX,datay,percentage=0.8)\nlabel_count=np.max(train_y)+1\nfor index in range(label_count):\n    print(f\"np.sum(train_y==index)={np.sum(train_y==index)}\")\nprint(f\"train_X.shape:{train_X.shape},valid_X.shape:{valid_X.shape}\")","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:02:43.563922Z","iopub.execute_input":"2023-07-06T13:02:43.564311Z","iopub.status.idle":"2023-07-06T13:02:44.130486Z","shell.execute_reply.started":"2023-07-06T13:02:43.564253Z","shell.execute_reply":"2023-07-06T13:02:44.129576Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class CNN(nn.Module):\n\n    def __init__(self):\n        \n        super(CNN,self).__init__()\n        \n        self.conv=nn.Sequential(\n        nn.Conv2d(1,16,3,1,0,bias=False),#30*30->28*28\n        nn.BatchNorm2d(16),\n        nn.SiLU(),\n        nn.Conv2d(16,32,3,1,1,bias=False),#28*28->26*26\n        nn.BatchNorm2d(32),\n        nn.ReLU(),\n        nn.Conv2d(32,72,5,1,0,bias=False),#26*26->22*22\n        nn.BatchNorm2d(72),\n        nn.Tanh(),\n        nn.Conv2d(72,128,5,1,1,bias=False),#20*20\n        nn.MaxPool2d(2,2,0),#10*10\n        nn.BatchNorm2d(128),\n        nn.GELU(),\n        nn.Conv2d(128,256,5,1,0,bias=False),#6*6\n        nn.BatchNorm2d(256),\n        nn.SiLU(),\n        )\n        \n        self.fc=nn.Sequential( \n        nn.Linear(256*6*6,128),\n        nn.BatchNorm1d(128),\n        nn.Dropout(p=0.2),\n        nn.GELU(),\n        nn.Linear(128,10),\n        )\n  \n\n    def forward(self,input): \n        output=self.conv(input)\n        output=output.view(-1,256*6*6)#reshape\n        output=self.fc(output)\n        return F.log_softmax(output/5,dim=1)","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:02:44.132003Z","iopub.execute_input":"2023-07-06T13:02:44.132405Z","iopub.status.idle":"2023-07-06T13:02:44.142699Z","shell.execute_reply.started":"2023-07-06T13:02:44.132372Z","shell.execute_reply":"2023-07-06T13:02:44.141483Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"%%time\n\ndevice='cuda' if torch.cuda.is_available() else\"cpu\"\n\ntrain_X=np.array(train_X).reshape(-1,1,28,28)/255.0-0.5\ntrain_y=np.array(train_y)\nvalid_X=np.array(valid_X).reshape(-1,1,28,28)/255.0-0.5\n\nvalid_X1=torch.Tensor(valid_X).to(device)\nnetC=CNN()\ntrain_accs=[]#存储训练集的准确率\nvalid_accs=[]#存储测试集的准确率\n#训练周期为500次\nnum_epochs=500\nbatch_size=200#一次训练的数据量\nweight_decay=0.0001\n\n#优化器\noptimizer=optim.Adam(netC.parameters(),lr=0.00025,betas=(0.5,0.999))\n#交叉熵损失函数\ncriterion=nn.CrossEntropyLoss()\nnetC=netC.to(device)\nfor epoch in range(num_epochs):\n    \n    random_num=[index for index in range(len(train_X))]\n    np.random.shuffle(random_num)\n    \n    train_X=train_X[random_num]\n    \n    train_y=train_y[random_num]\n    \n    train_X1=torch.Tensor(train_X[:batch_size]).to(device)\n    train_y1=torch.Tensor(train_y[:batch_size]).long().to(device)\n    \n    #训练\n    netC.train()\n    #将数据放进去训练\n    output=netC(train_X1).to(device)\n    #计算每次的损失函数\n    loss=criterion(output,F.one_hot(train_y1,10).type(torch.FloatTensor).to(device)).to(device)\n    #L1正则化\n    l1_regularization = torch.tensor(0.).to(device)\n    for param in netC.parameters():\n        l1_regularization += torch.norm(param, 1)\n    loss += weight_decay * l1_regularization\n    #反向传播\n    loss.backward()\n    #优化器进行优化(梯度下降,降低误差)\n    optimizer.step()\n    #将梯度清空\n    optimizer.zero_grad()\n    if epoch%50==0:\n        netC.eval()\n        with torch.no_grad():\n            output=netC(train_X1).to(device)\n            output=output.detach().cpu().numpy()\n            train_pred=np.argmax(output,axis=1)\n\n            output=netC(valid_X1).to(device)\n            output=output.detach().cpu().numpy()\n            valid_pred=np.argmax(output,axis=1)\n        \n        train_acc=np.sum(train_pred==train_y[:batch_size])/len(train_y[:batch_size])\n        valid_acc=np.sum(valid_pred==valid_y)/len(valid_y)\n\n        train_accs.append(train_acc)\n        valid_accs.append(valid_acc)\n        print(f\"epoch:{epoch},loss:{loss}\\n train_acc:{train_acc},valid_acc:{valid_acc}\")","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:02:44.144334Z","iopub.execute_input":"2023-07-06T13:02:44.145093Z","iopub.status.idle":"2023-07-06T13:04:03.117589Z","shell.execute_reply.started":"2023-07-06T13:02:44.145057Z","shell.execute_reply":"2023-07-06T13:04:03.116590Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"percentage=0.25#剪枝的比例\nparameter=[]#每层的参数\nparameter_sorted=np.array([])\nfor name,module in netC.named_modules():\n    if isinstance(module,nn.BatchNorm2d): \n        parameter.append(module.weight.detach().cpu().numpy())\n        parameter_sorted=np.concatenate((parameter_sorted,module.weight.detach().cpu().numpy()),axis=0)\nparameter_sorted,index_sorted=torch.sort(torch.abs(torch.Tensor(parameter_sorted)))\nsplit_weight=parameter_sorted[int(percentage*len(parameter_sorted))]#小于这个数字的特征图不要\nfor i in range(len(parameter)):\n    for j in range(len(parameter[i])):\n        if abs(parameter[i][j])<split_weight:\n            parameter[i][j]=0\nfor index in range(len(parameter)):\n    print(f\"np.sum(parameter[index]!=0)={np.sum(parameter[index]!=0)}\")","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:04:03.118901Z","iopub.execute_input":"2023-07-06T13:04:03.119280Z","iopub.status.idle":"2023-07-06T13:04:03.151362Z","shell.execute_reply.started":"2023-07-06T13:04:03.119242Z","shell.execute_reply":"2023-07-06T13:04:03.150079Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#存储卷积层和batchNorm层剪枝剩下的参数\nlayer=0#记录一下这是第几层\ntotal_parameter=[]\nfor name,module in netC.named_modules():\n    if isinstance(module,nn.Conv2d):\n        conv_parameter=[]\n        conv_weight=module.weight.detach().cpu().numpy()\n        for index in range(len(conv_weight)):\n            if parameter[layer][index]:#如果这个特征图重要\n                #conv_weight[index]\n                weight=conv_weight[index]\n                remain_weight=[]\n                if layer==0:\n                    remain_weight=weight\n                for j in range(len(weight)):\n                    if layer and parameter[layer-1][j]:#上一层的这个位置重要吗?\n                        remain_weight.append(weight[j])\n                conv_parameter.append(np.array(remain_weight))\n        #print(f\"conv_weight[index].shape:{conv_weight[index].shape}\")\n        total_parameter.append(conv_parameter)\n    if isinstance(module,nn.BatchNorm2d): \n        bn_parameter=[]\n        bn_weight=module.weight.detach().cpu().numpy()\n        for index in range(len(bn_weight)):\n            if parameter[layer][index]:#如果这个特征图重要\n                bn_parameter.append(bn_weight[index])\n        total_parameter.append(bn_parameter)\n        layer+=1\nprint(f\"len(total_parameter):{len(total_parameter)}\")","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:04:03.154563Z","iopub.execute_input":"2023-07-06T13:04:03.154900Z","iopub.status.idle":"2023-07-06T13:04:03.208774Z","shell.execute_reply.started":"2023-07-06T13:04:03.154874Z","shell.execute_reply":"2023-07-06T13:04:03.207758Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#score:0.92976\nclass prune_CNN(nn.Module):\n\n    def __init__(self):\n        \n        super(prune_CNN,self).__init__()\n        \n        self.conv=nn.Sequential(\n        nn.Conv2d(1,16,3,1,0,bias=False),#30*30->28*28\n        nn.BatchNorm2d(16),\n        nn.SiLU(),\n        nn.Conv2d(16,32,3,1,1,bias=False),#28*28->26*26\n        nn.BatchNorm2d(32),\n        nn.ReLU(),\n        nn.Conv2d(32,72,5,1,0,bias=False),#26*26->22*22\n        nn.BatchNorm2d(72),\n        nn.Tanh(),\n        nn.Conv2d(72,107,5,1,1,bias=False),#20*20\n        nn.MaxPool2d(2,2,0),#10*10\n        nn.BatchNorm2d(107),\n        nn.GELU(),\n        nn.Conv2d(107,151,5,1,0,bias=False),#6*6\n        nn.BatchNorm2d(151),\n        nn.SiLU(),\n        )\n        \n        self.fc=nn.Sequential( \n        nn.Linear(151*6*6,128),\n        nn.BatchNorm1d(128),\n        nn.Dropout(p=0.2),\n        nn.GELU(),\n        nn.Linear(128,10),\n        )\n\n    def forward(self,input): \n        output=self.conv(input)\n        output=output.view(-1,151*6*6)#reshape\n        output=self.fc(output)\n        return F.log_softmax(output/5,dim=1)","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:04:03.210069Z","iopub.execute_input":"2023-07-06T13:04:03.210383Z","iopub.status.idle":"2023-07-06T13:04:03.220447Z","shell.execute_reply.started":"2023-07-06T13:04:03.210352Z","shell.execute_reply":"2023-07-06T13:04:03.219480Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"netC1=prune_CNN()\nlayer=0\nfor name,module in netC1.named_modules():\n    if isinstance(module,nn.Conv2d) or isinstance(module,nn.BatchNorm2d):  \n        module.weight.data=torch.Tensor(total_parameter[layer]).clone()\n        print(f\"module.weight.shape:{module.weight.shape}\\n total_parameter[layer].shape:{np.array(total_parameter[layer]).shape:}\")\n        print(f\"len(total_parameter[layer]):{len(total_parameter[layer])}\")\n        layer+=1","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:04:03.221812Z","iopub.execute_input":"2023-07-06T13:04:03.222435Z","iopub.status.idle":"2023-07-06T13:04:03.659529Z","shell.execute_reply.started":"2023-07-06T13:04:03.222386Z","shell.execute_reply":"2023-07-06T13:04:03.658547Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"%%time\n\ntrain_accs=[]#存储训练集的准确率\nvalid_accs=[]#存储测试集的准确率\n#训练周期为500次\nnum_epochs=500\nbatch_size=200#一次训练的数据量\n\n#优化器\noptimizer=optim.Adam(netC1.parameters(),lr=0.00025,betas=(0.5,0.999))\n#交叉熵损失函数\ncriterion=nn.CrossEntropyLoss()\n\nnetC1.to(device)\n\nfor epoch in range(num_epochs):\n    \n    random_num=[index for index in range(len(train_X))]\n    np.random.shuffle(random_num)\n    \n    train_X=train_X[random_num]\n    \n    train_y=train_y[random_num]\n    \n    train_X1=torch.Tensor(train_X[:batch_size]).to(device)\n    train_y1=torch.Tensor(train_y[:batch_size]).long().to(device)\n    \n    #训练\n    netC1.train()\n    #将数据放进去训练\n    output=netC1(train_X1).to(device)\n    #计算每次的损失函数\n    loss=criterion(output,F.one_hot(train_y1,10).type(torch.FloatTensor).to(device)).to(device)\n    #反向传播\n    loss.backward()\n    #优化器进行优化(梯度下降,降低误差)\n    optimizer.step()\n    #将梯度清空\n    optimizer.zero_grad()\n    if epoch%50==0:\n        netC.eval()\n        with torch.no_grad():\n            output=netC1(train_X1).to(device)\n            output=output.detach().cpu().numpy()\n            train_pred=np.argmax(output,axis=1)\n\n            output=netC1(valid_X1).to(device)\n            output=output.detach().cpu().numpy()\n            valid_pred=np.argmax(output,axis=1)\n        \n        train_acc=np.sum(train_pred==train_y[:batch_size])/len(train_y[:batch_size])\n        valid_acc=np.sum(valid_pred==valid_y)/len(valid_y)\n\n        train_accs.append(train_acc)\n        valid_accs.append(valid_acc)\n        print(f\"epoch:{epoch},loss:{loss}\\n train_acc:{train_acc},valid_acc:{valid_acc}\")","metadata":{"execution":{"iopub.status.busy":"2023-07-06T13:04:03.660924Z","iopub.execute_input":"2023-07-06T13:04:03.661489Z","iopub.status.idle":"2023-07-06T13:05:10.736802Z","shell.execute_reply.started":"2023-07-06T13:04:03.661451Z","shell.execute_reply":"2023-07-06T13:05:10.735796Z"},"trusted":true},"execution_count":null,"outputs":[]}]}