{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceId":59093,"databundleVersionId":7469972,"sourceType":"competition"}],"dockerImageVersionId":30636,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"## Created by yunsuxiaozi","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19"}},{"cell_type":"markdown","source":"### Import necessary libraries","metadata":{}},{"cell_type":"code","source":"#https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-training\n#https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-inference\n#necessary\nimport pandas as pd#导入csv文件的库\nimport numpy as np#进行矩阵运算的库\nimport matplotlib.pyplot as plt#导入强大的绘图库\nimport torch #一个深度学习的库Pytorch\nimport timm#图像分类预训练模型库\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 torchvision.transforms as transforms#Pytorch下面的图像处理库,用于对图像进行数据增强\n#设置随机种子\nimport random\nimport warnings#避免一些可以忽略的报错\nwarnings.filterwarnings('ignore')#filterwarnings()方法是用于设置警告过滤器的方法，它可以控制警告信息的输出方式和级别。","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Config","metadata":{}},{"cell_type":"code","source":"class Config:\n seed=2024\n image_transform=transforms.Resize((512,512))\n batch_size=64\n num_epochs=10#由于是预训练模型,训练10个epoch足够了\n num_folds=5#5折交叉验证.","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Seed ","metadata":{}},{"cell_type":"code","source":"def seed_everything(seed):\n torch.backends.cudnn.deterministic = True#将cuda加速的随机数生成器设为确定性模式\n torch.backends.cudnn.benchmark = True#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\n torch.manual_seed(seed)#pytorch的随机种子\n np.random.seed(seed)#numpy的随机种子\n random.seed(seed)#python内置的随机种子\nseed_everything(Config.seed)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### labels","metadata":{}},{"cell_type":"code","source":"train_df=pd.read_csv(\"/kaggle/input/hms-harmful-brain-activity-classification/train.csv\")\n\nlabels=['seizure','lpd','gpd','lrda','grda','other']\nfor label in labels:\n group=train_df[f'{label}_vote'].groupby(train_df['spectrogram_id']).sum()\n label_vote_sum = pd.DataFrame({'spectrogram_id': group.index, f'{label}_vote_sum': group.values})\n if label=='seizure':\n train_feats=label_vote_sum\n else:\n train_feats=train_feats.merge(label_vote_sum,on='spectrogram_id',how='left')\ntrain_feats['total_vote']=0\nfor label in labels:\n train_feats['total_vote']+=train_feats[f'{label}_vote_sum']\nfor label in labels:\n train_feats[f'{label}_vote']=train_feats[f'{label}_vote_sum']/train_feats['total_vote']\nchoose_cols=['spectrogram_id']\nfor label in labels:\n choose_cols+=[f'{label}_vote']\ntrain_feats=train_feats[choose_cols]\ntrain_feats['path']=train_feats['spectrogram_id'].apply(lambda x: \"/kaggle/input/hms-harmful-brain-activity-classification/train_spectrograms/\"+str(x)+\".parquet\" )\ntrain_feats.head()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Loss function :Kullback Leibler Divergence","metadata":{}},{"cell_type":"code","source":"def KL_loss(p,q):\n epsilon=10**(-15)\n p=torch.clip(p,epsilon,1-epsilon)\n q = nn.functional.log_softmax(q,dim=1)\n #对第一个维度,就是num_classes维度的损失求和,得到每个样本的损失,然后对第0维求平均,得到每个样本平均KL散度.\n return torch.mean(torch.sum(p*(torch.log(p)-q),dim=1))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### get_batch_data","metadata":{}},{"cell_type":"code","source":"def get_batch(paths,batch_size=Config.batch_size):\n eps=1e-6\n batch_data=[]\n for path in paths:\n data=pd.read_parquet(path[0])\n #这里最小值是0,故用-1填充.第一列是时间列,故去掉 ,行是不同列,列是时间\n data = data.fillna(-1).values[:,1:].T\n #选取一段时间的数据进行训练\n data=data[:,0:300]#(400,300)\n data=np.clip(data,np.exp(-6),np.exp(10))#最大值为89209464.0\n data= np.log(data)#对数变换\n #对数据进行归一化\n data_mean=data.mean(axis=(0,1))\n data_std=data.std(axis=(0,1))\n data=(data-data_mean)/(data_std+eps)\n data_tensor = torch.unsqueeze(torch.Tensor(data), dim=0)\n data=Config.image_transform(data_tensor)\n batch_data.append(data)\n batch_data=torch.stack(batch_data)\n return batch_data","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Model_training","metadata":{}},{"cell_type":"code","source":"device ='cuda' if torch.cuda.is_available() else 'cpu'\nprint(f\"device:{device}\")\n\n\ntotal_idx=np.arange(len(train_feats))\nnp.random.shuffle(total_idx)\n\nfor fold in range(Config.num_folds):\n\n test_idx=total_idx[fold*len(total_idx)//Config.num_folds:(fold+1)*len(total_idx)//Config.num_folds]\n train_idx=np.array([idx for idx in total_idx if idx not in test_idx])\n\n #下载模型resnet34d,并且将训练好的参数加载进来.每折初始化一个模型.\n model = timm.create_model('resnet34d',pretrained=True, num_classes=6, in_chans=1)\n \n #优化器\n optimizer=optim.AdamW(model.parameters(),lr=0.001,betas=(0.5,0.999),weight_decay=0.01)\n\n best_test_loss=1.0#目前损失最好的成绩,每一折都要重新初始化\n train_losses=[]\n test_losses=[]\n\n print(f\"start\")\n\n for epoch in range(Config.num_epochs):\n print(f\"epoch {epoch}:\")\n model.to(device)#保存模型的时候可能会转成CPU,故设置这个来保证模型会在GPU上训练.\n #训练\n model.train()\n train_loss=[]\n random_num=np.arange(len(train_idx))\n np.random.shuffle(random_num)\n train_idx=train_idx[random_num]\n\n for idx in range(0,len(train_idx),Config.batch_size): \n #将梯度清空\n optimizer.zero_grad()\n train_idx1=train_idx[idx:idx+Config.batch_size]\n train_X1_path=train_feats[['path']].iloc[train_idx1].values\n train_X1=get_batch(train_X1_path,batch_size=Config.batch_size)\n train_y1=train_feats[['seizure_vote','lpd_vote','gpd_vote','lrda_vote','grda_vote','other_vote']].iloc[train_idx1].values\n train_y1=torch.Tensor(train_y1)\n #将数据放进去训练\n train_pred=model(train_X1.to(device)).to(device)\n #计算每次的损失函数\n loss=KL_loss(train_y1.to(device),train_pred.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 print(f\"idx:{idx},loss:{loss}\")\n train_loss.append(loss.detach().cpu().numpy())\n train_loss=np.mean(np.array(train_loss))\n print(f\"train_loss:{train_loss}\")\n test_loss=[]\n model.eval()\n with torch.no_grad():\n for idx in range(0,len(test_idx),Config.batch_size): \n test_idx1=test_idx[idx:idx+Config.batch_size]\n test_X1_path=train_feats[['path']].iloc[test_idx1].values\n test_X1=get_batch(test_X1_path,batch_size=Config.batch_size)\n\n test_y1=train_feats[['seizure_vote','lpd_vote','gpd_vote','lrda_vote','grda_vote','other_vote']].iloc[test_idx1].values\n test_y1=torch.Tensor(test_y1)\n #将数据放进去训练\n test_pred=model(test_X1.to(device)).to(device)\n #计算每次的损失函数\n loss=KL_loss(test_y1.to(device),test_pred.to(device)).to(device)\n test_loss.append(loss.detach().cpu().numpy())\n test_loss=np.mean(np.array(test_loss))\n print(f\"test_loss:{test_loss}\")\n if test_loss