{"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":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n    for filename in filenames:\n        print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","execution":{"iopub.execute_input":"2022-09-16T21:08:05.102930Z","iopub.status.busy":"2022-09-16T21:08:05.102208Z","iopub.status.idle":"2022-09-16T21:08:05.127012Z","shell.execute_reply":"2022-09-16T21:08:05.125799Z"},"papermill":{"duration":0.035138,"end_time":"2022-09-16T21:08:05.131027","exception":false,"start_time":"2022-09-16T21:08:05.095889","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# 导入库\nimport pandas as pd\nimport numpy as np\n# from tqdm import tqdm,tqdm_notebook \nfrom tqdm.notebook import tqdm,trange\nimport gc\nimport os\nimport sys\nfrom scipy import sparse\nos.environ['CUDA_LAUNCH_BLOCKING'] = \"1\"\nimport torch\nimport torch.nn.functional as F\nfrom torch import Tensor\nimport pickle\n!pip install tables","metadata":{"execution":{"iopub.execute_input":"2022-09-16T21:08:05.141197Z","iopub.status.busy":"2022-09-16T21:08:05.140610Z","iopub.status.idle":"2022-09-16T21:08:19.618745Z","shell.execute_reply":"2022-09-16T21:08:19.617533Z"},"papermill":{"duration":14.48625,"end_time":"2022-09-16T21:08:19.622040","exception":false,"start_time":"2022-09-16T21:08:05.135790","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# from sqrt4kaido\nclass NegativeCorrLoss(torch.nn.Module):\n    \"\"\"Negative correlation loss function for Keras\n\n    Precondition:\n    y_true.mean(axis=1) == 0\n    y_true.std(axis=1) == 1\n\n    Returns:\n    -1 = perfect positive correlation\n    1 = totally negative correlation\n    \"\"\"\n\n    def __init__(self):\n        super().__init__()\n\n    def forward(self, preds, targets):\n\n        my = torch.mean(preds, dim=1)\n        my = torch.tile(torch.unsqueeze(my, dim=1), (1, targets.shape[1]))\n        ym = preds - my\n        r_num = torch.sum(torch.multiply(targets, ym), dim=1)\n        r_den = torch.sqrt(\n            torch.sum(torch.square(ym), dim=1) * float(targets.shape[-1])\n        )\n        r = torch.mean(r_num / r_den)\n        return -r\ndef correlation_score(y_true, y_pred):\n    \"\"\"Scores the predictions according to the competition rules. \n    \n    It is assumed that the predictions are not constant.\n    \n    Returns the average of each sample's Pearson correlation coefficient\"\"\"\n    if type(y_true) == pd.DataFrame: y_true = y_true.values\n    if type(y_pred) == pd.DataFrame: y_pred = y_pred.values\n    corrsum = 0\n    for i in range(len(y_true)):\n        corrsum += np.corrcoef(y_true[i], y_pred[i])[1, 0]\n    return corrsum / len(y_true)","metadata":{"execution":{"iopub.execute_input":"2022-09-16T21:08:19.635025Z","iopub.status.busy":"2022-09-16T21:08:19.633532Z","iopub.status.idle":"2022-09-16T21:08:19.644498Z","shell.execute_reply":"2022-09-16T21:08:19.643636Z"},"papermill":{"duration":0.019186,"end_time":"2022-09-16T21:08:19.646511","exception":false,"start_time":"2022-09-16T21:08:19.627325","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class DatasetSparse_cite(torch.utils.data.Dataset):\n    def __init__(self,data_x_path,data_y_path):\n        \"\"\"\n        \"\"\"\n        self.x = sparse.load_npz(data_x_path) \n        Y = pd.read_hdf(data_y_path).values\n        Y -= Y.mean(axis=1).reshape(-1, 1)\n        Y /= Y.std(axis=1).reshape(-1, 1)\n\n        self.y = Y\n        self.number = self.x.shape[0]\n        print('Finish loading data into memory')\n        \n    def __len__(self):\n        return self.number\n\n    def __getitem__(self,idx):\n        return self.x[idx],self.y[idx]\n    \ndef collate_fn_sparse_cite(batch):\n    \"\"\"\n    \"\"\"\n    index,feats,values,labels = [],[],[],[]\n    for idx, (feature, label) in enumerate(batch):\n        index+=[idx]*len(feature.indices)\n        feats+=feature.indices.tolist()\n        values+=feature.data.tolist()\n        labels+=[label.tolist()]\n    return torch.tensor(index), torch.tensor(feats), torch.tensor(values),torch.tensor(labels)\n\n    \n\nclass DatasetSparse_multi(torch.utils.data.Dataset):\n    def __init__(self,data_x_path,data_y_path):\n        \"\"\"\n        \"\"\"\n        self.x = sparse.load_npz(data_x_path) \n        self.y = sparse.load_npz(data_y_path) \n        self.y_mean = np.load('../input/data-process/multi_y_mean.npy')\n        self.y_std = np.load('../input/data-process/multi_y_std.npy')\n        self.number = self.x.shape[0]\n        \n        print('Finish loading data into memory')\n        \n    def __len__(self):\n        return self.number\n\n    def __getitem__(self,idx):\n        return self.x[idx],self.y[idx],self.y_mean[idx],self.y_std[idx] \n\ndef collate_fn_sparse_multi(batch):\n    \"\"\"\n    \"\"\"\n    index,feats,values,labels = [],[],[],[]\n    for idx, (feature, label, label_mean, label_std) in enumerate(batch):\n        index+=[idx]*len(feature.indices)\n        feats+=feature.indices.tolist()\n        values+=feature.data.tolist()\n        labels+=((label.toarray()-label_mean)/label_std).tolist()\n    return torch.tensor(index), torch.tensor(feats), torch.tensor(values),torch.tensor(labels)\n\n\nclass DatasetSparse_pred(torch.utils.data.Dataset):\n    def __init__(self,data_x_path):\n        \"\"\"\n        \"\"\"\n        self.pred_x = sparse.load_npz(data_x_path) \n        self.number = self.pred_x.shape[0]\n        print('Finish loading data into memory')\n    def __len__(self):\n        return self.number\n\n    def __getitem__(self,idx):\n        return self.pred_x[idx]\n    \ndef collate_fn_sparse_pred(batch):\n    \"\"\"\n    \"\"\"\n    index,feats,values,labels = [],[],[],[]\n    for idx, feature in enumerate(batch):\n        index+=[idx]*len(feature.indices)\n        feats+=feature.indices.tolist()\n        values+=feature.data.tolist()\n    return torch.tensor(index), torch.tensor(feats), torch.tensor(values)","metadata":{"execution":{"iopub.execute_input":"2022-09-16T21:08:19.657957Z","iopub.status.busy":"2022-09-16T21:08:19.657066Z","iopub.status.idle":"2022-09-16T21:08:19.674142Z","shell.execute_reply":"2022-09-16T21:08:19.673335Z"},"papermill":{"duration":0.024615,"end_time":"2022-09-16T21:08:19.676064","exception":false,"start_time":"2022-09-16T21:08:19.651449","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# model define\nclass LinearReg(torch.nn.Module):\n    \n    def __init__(self, model_config):\n        \"\"\"\n        model_config : {'input_dim':20000,'device':'cpu','output':140,'layer_one':160,'fc_dims':[64,32]}\n        \"\"\"\n        super(LinearReg, self).__init__()\n\n        self.input_dim = model_config['input_dim']\n        self.device = model_config['device']\n        self.output = model_config['output']\n        self.layer_one = model_config['layer_one']\n        self.fc_dims = model_config['fc_dims']\n        self.layer_one_lst = torch.nn.ParameterList()\n        \n        for index,i in enumerate(range(self.layer_one)):\n            self.layer_one_lst.append(torch.nn.Parameter(torch.zeros(1, 1,device=self.device)))\n            self.layer_one_lst.append(torch.nn.Parameter(torch.randn(self.input_dim, 1,device=self.device))) # 多1维，从 1 开始\n            torch.nn.init.xavier_uniform_(self.layer_one_lst[index * 2+1],gain=1)\n        \n        num_dim = self.layer_one\n        dim = self.layer_one\n        \n        self.mats = torch.nn.ParameterList()\n        for (index, fc_dim) in enumerate(self.fc_dims): # [10,5,1]\n            num_dim+=fc_dim\n            self.mats.append(torch.nn.Parameter(torch.randn(dim, fc_dim,device=self.device)))\n            self.mats.append(torch.nn.Parameter(torch.zeros(1, fc_dim,device=self.device))) # 可以多个 b\n            torch.nn.init.kaiming_uniform_(self.mats[index * 2], mode='fan_in', nonlinearity='relu')\n            dim = fc_dim\n            \n        self.output_w = torch.nn.Parameter(torch.randn(num_dim, self.output,device=self.device))\n        self.output_b = torch.nn.Parameter(torch.zeros(1, self.output,device=self.device)) # 可以多个 b\n        torch.nn.init.kaiming_uniform_(self.output_w, mode='fan_in', nonlinearity='relu')\n        \n        \n            \n            \n    def first_order(self, batch_size, index, values, bias, weights):\n        # type: (int, Tensor, Tensor, Tensor, Tensor) -> Tensor\n        size = batch_size # 64\n        srcs = weights.view(1, -1).mul(values.view(1, -1)).view(-1) # weight [1234,1] values[1234]  -> 1x1232 -> 1234\n        output = torch.zeros(size, dtype=torch.float32,device = self.device) # 64 \n        # output=output.to(device=torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\"))\n        output.scatter_add_(0, index, srcs) # dim,index,src  index 是1维   output[index[i]]+=srcs[i] 遍历 index 的 i index:[0,0,0,0,0,\\cdots,1,1,\\cdots,63,63] e.eg [1234] \n        first = output + bias # [64] # wx+b 一阶\n        return first\n    \n    \n    def higher_order(self, batch_size, index, mats):\n        for i in range(int(len(mats) / 2)):\n            output = torch.relu(output.matmul(mats[i * 2]) + mats[i * 2 + 1]) # 全连接层\n    \n\n    def forward(self, batch_size, index, feats, values):\n        # type: (int, Tensor, Tensor, Tensor, Tensor) -> Tensor\n        # batch_size : 64\n        ret_lst = []\n        for idx,i in enumerate(range(self.layer_one)):\n            batch_first = F.embedding(feats, self.layer_one_lst[idx * 2+1]) # weights [100000,1]  feats: [a_1,a_1,\\cdots,a_64,a_64] e.g [1234] batch_first: [1234,1]\n            ret_lst.append(self.first_order(batch_size, index, values, self.layer_one_lst[idx * 2],batch_first).view(-1,1)) # [64,1]\n        output = torch.relu(torch.concat(ret_lst,axis=1))    \n        \n        ret_lst_two = [output]\n        for i in range(int(len(self.mats) / 2)):\n            output = torch.relu(output.matmul(self.mats[i * 2]) + self.mats[i * 2 + 1])\n            ret_lst_two.append(output) # 全连接层\n        \n        second = torch.relu(torch.concat(ret_lst_two,axis=1))   \n            \n        return second.matmul(self.output_w)+self.output_b\n\n    @torch.jit.export\n    def get_name(self):\n        return \"LinearReg\"","metadata":{"execution":{"iopub.execute_input":"2022-09-16T21:08:19.686896Z","iopub.status.busy":"2022-09-16T21:08:19.686632Z","iopub.status.idle":"2022-09-16T21:08:19.704178Z","shell.execute_reply":"2022-09-16T21:08:19.703282Z"},"papermill":{"duration":0.025427,"end_time":"2022-09-16T21:08:19.706217","exception":false,"start_time":"2022-09-16T21:08:19.680790","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train  = 'cite'\n\nif train == 'cite':\n    # train cite\n    train_set = DatasetSparse_cite('../input/data-process/train_cite_inputs_sparse.npz','../input/open-problems-multimodal/train_cite_targets.h5')\n    train_set_loader = torch.utils.data.DataLoader(train_set, batch_size=64, shuffle=True, \n                                                   num_workers=2,drop_last=False,pin_memory=False,collate_fn=collate_fn_sparse_cite)\nif train == 'multi':\n    # train multi\n    train_set = DatasetSparse_multi('../input/open-problems-msci-multiome-sparse-matrices/test_multi_inputs_sparse.npz','../input/open-problems-msci-multiome-sparse-matrices/train_multi_targets_sparse.npz')\n    train_set_loader = torch.utils.data.DataLoader(train_set, batch_size=64, shuffle=True, \n                                                   num_workers=2,drop_last=False,pin_memory=False,collate_fn=collate_fn_sparse_multi)","metadata":{"execution":{"iopub.execute_input":"2022-09-16T21:08:19.717317Z","iopub.status.busy":"2022-09-16T21:08:19.716545Z","iopub.status.idle":"2022-09-16T21:08:39.813118Z","shell.execute_reply":"2022-09-16T21:08:39.811840Z"},"papermill":{"duration":20.104413,"end_time":"2022-09-16T21:08:39.815356","exception":false,"start_time":"2022-09-16T21:08:19.710943","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model_config = {\n'input_dim':train_set.x.shape[1]+1,\n'device':torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\") ,\n'output':140, # cite:140 multi:23418\n'layer_one':160, # cite:160  multi:256\n'fc_dims':[64,32],\n'epochs' : 20\n}\nreg_model = LinearReg(model_config)\noptimizer = torch.optim.SGD(reg_model.parameters(), lr=0.025, momentum=0.9, weight_decay=5e-4) # 5e-4\nloss_fn = NegativeCorrLoss()","metadata":{"execution":{"iopub.execute_input":"2022-09-16T21:08:39.827465Z","iopub.status.busy":"2022-09-16T21:08:39.826634Z","iopub.status.idle":"2022-09-16T21:08:43.078270Z","shell.execute_reply":"2022-09-16T21:08:43.077331Z"},"papermill":{"duration":3.260133,"end_time":"2022-09-16T21:08:43.080788","exception":false,"start_time":"2022-09-16T21:08:39.820655","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(model_config)","metadata":{"execution":{"iopub.execute_input":"2022-09-16T21:08:43.092567Z","iopub.status.busy":"2022-09-16T21:08:43.091687Z","iopub.status.idle":"2022-09-16T21:08:43.097190Z","shell.execute_reply":"2022-09-16T21:08:43.096288Z"},"papermill":{"duration":0.013592,"end_time":"2022-09-16T21:08:43.099480","exception":false,"start_time":"2022-09-16T21:08:43.085888","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"for epoch in trange(model_config['epochs'],desc = 'epochs'):\n    losses = []\n    train_batch_iterator = tqdm(train_set_loader, disable=False, file=sys.stdout)\n    for index,feats,values,labels in train_batch_iterator:\n        index = index.to(model_config['device'])\n        feats = feats.to(model_config['device'])\n        values = values.to(model_config['device'])\n        labels = labels.to(model_config['device'])\n\n        optimizer.zero_grad() \n        # batch_size 最后一个不一样\n        batch_size = len(labels)\n        outputs = reg_model(batch_size, index, feats, values)\n\n        loss  = loss_fn(outputs,labels)\n\n        loss.backward() \n        optimizer.step() \n\n        losses.append(loss.item())\n        train_batch_iterator.set_postfix(loss=loss.item()) # 实时 batch 的loss\n\n    print('epoch:{} \\n loss:{:.6}'.format(epoch,np.mean(losses)))","metadata":{"execution":{"iopub.execute_input":"2022-09-16T21:08:43.110215Z","iopub.status.busy":"2022-09-16T21:08:43.109950Z","iopub.status.idle":"2022-09-16T22:14:39.016548Z","shell.execute_reply":"2022-09-16T22:14:39.015291Z"},"papermill":{"duration":3955.915157,"end_time":"2022-09-16T22:14:39.019354","exception":false,"start_time":"2022-09-16T21:08:43.104197","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"torch.save(reg_model.state_dict(),'./reg_model_{}'.format(train))","metadata":{"execution":{"iopub.execute_input":"2022-09-16T22:14:39.035418Z","iopub.status.busy":"2022-09-16T22:14:39.034497Z","iopub.status.idle":"2022-09-16T22:14:39.093519Z","shell.execute_reply":"2022-09-16T22:14:39.092580Z"},"papermill":{"duration":0.069708,"end_time":"2022-09-16T22:14:39.096186","exception":false,"start_time":"2022-09-16T22:14:39.026478","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# data leak\nif train == 'cite':\n    data_leak_cite = train_set.y[:7476]\ndel train_batch_iterator\ngc.collect()\ndel train_set_loader\ngc.collect()\ndel train_set\ngc.collect()                                       ","metadata":{"execution":{"iopub.execute_input":"2022-09-16T22:14:39.111563Z","iopub.status.busy":"2022-09-16T22:14:39.111250Z","iopub.status.idle":"2022-09-16T22:14:39.465303Z","shell.execute_reply":"2022-09-16T22:14:39.464396Z"},"papermill":{"duration":0.364138,"end_time":"2022-09-16T22:14:39.467430","exception":false,"start_time":"2022-09-16T22:14:39.103292","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# train cite\nif train == 'cite':\n    test_set = DatasetSparse_pred('../input/data-process/test_cite_inputs_sparse.npz')\n    test_set_loader = torch.utils.data.DataLoader(test_set, batch_size=64, shuffle=False, \n                                                   num_workers=2,drop_last=False,pin_memory=False,collate_fn=collate_fn_sparse_pred)\n    \nif train == 'multi':\n    test_set = DatasetSparse_pred('../input/open-problems-msci-multiome-sparse-matrices/test_multi_inputs_sparse.npz')\n    test_set_loader = torch.utils.data.DataLoader(test_set, batch_size=64, shuffle=False, \n                                                   num_workers=2,drop_last=False,pin_memory=False,collate_fn=collate_fn_sparse_pred)\n    multi_test_row = np.load('../input/data-process/multi_test_row.npy')\n    multi_test_row_name = np.load('../input/data-process/multi_test_row_name.npy')\n    multi_test_col = np.load('../input/data-process/multi_test_col.npy')\n    multi_test_col_name =pd.read_hdf('../input/open-problems-multimodal/train_multi_targets.h5',start=0,stop=1).columns.to_numpy()\n    \n    \n\n\nwith torch.no_grad():\n    test_batch_iterator = tqdm(test_set_loader, disable=False, file=sys.stdout)\n    test_pred = []\n    start_r = 0\n    start_c = 0\n    for index,feats,values in test_batch_iterator:\n        index = index.to(model_config['device'])\n        feats = feats.to(model_config['device'])\n        values = values.to(model_config['device'])\n\n        # batch_size 最后一个不一样, cite 和 multi 在这里的处理不一样\n        batch_size = index[-1].tolist()+1\n        outputs = reg_model(batch_size, index, feats, values)\n        if train == 'cite':\n            test_pred+=outputs.tolist()\n        if train == 'multi':\n            outputs = outputs[multi_test_row[start_r:start_r+batch_size]]\n            if len(outputs)>0:\n                c_cnt = np.sum(multi_test_row[start_r:start_r+batch_size])\n                index = torch.tensor(multi_test_col[start_c:start_c+c_cnt])\n                outputs = outputs.gather(1, index)\n                start_r+=batch_size\n                start_c+=c_cnt\n                test_pred+=outputs.tolist()\n        \ntest_pred = np.array(test_pred)\n\nif train == 'cite':\n    test_pred[:7476] = data_leak_cite\n    np.save('./cite_test_pred',test_pred)\nif train == 'multi':\n    np.save('./multi_test_pred',test_pred)","metadata":{"execution":{"iopub.execute_input":"2022-09-16T22:14:39.482659Z","iopub.status.busy":"2022-09-16T22:14:39.482357Z","iopub.status.idle":"2022-09-16T22:16:35.325614Z","shell.execute_reply":"2022-09-16T22:16:35.324388Z"},"papermill":{"duration":115.853955,"end_time":"2022-09-16T22:16:35.328203","exception":false,"start_time":"2022-09-16T22:14:39.474248","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"del test_batch_iterator\ngc.collect()\ndel test_set_loader\ngc.collect()\ndel test_set\ngc.collect()     ","metadata":{"execution":{"iopub.execute_input":"2022-09-16T22:16:35.344555Z","iopub.status.busy":"2022-09-16T22:16:35.344195Z","iopub.status.idle":"2022-09-16T22:16:35.672088Z","shell.execute_reply":"2022-09-16T22:16:35.671020Z"},"papermill":{"duration":0.338907,"end_time":"2022-09-16T22:16:35.674691","exception":false,"start_time":"2022-09-16T22:16:35.335784","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# submit\n# Convert the string columns to more efficient categorical types\neval_ids = pd.read_csv('../input/open-problems-multimodal/evaluation_ids.csv', index_col='row_id')\neval_ids.cell_id = eval_ids.cell_id.astype(pd.CategoricalDtype())\neval_ids.gene_id = eval_ids.gene_id.astype(pd.CategoricalDtype())\nsubmission = pd.Series(name='target',\n                       index=pd.MultiIndex.from_frame(eval_ids), \n                       dtype=np.float32)\nsubmission.reset_index(drop=True, inplace=True)\nsubmission.index.name = 'row_id'\nsubmission.iloc[:len(test_pred.ravel())] = test_pred.ravel()\nwith open(\"partial_submission_cite.pickle\", 'wb') as f: pickle.dump(submission, f)","metadata":{"execution":{"iopub.execute_input":"2022-09-16T22:16:35.690373Z","iopub.status.busy":"2022-09-16T22:16:35.689747Z","iopub.status.idle":"2022-09-16T22:18:28.539374Z","shell.execute_reply":"2022-09-16T22:18:28.538378Z"},"papermill":{"duration":112.860027,"end_time":"2022-09-16T22:18:28.541851","exception":false,"start_time":"2022-09-16T22:16:35.681824","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{"papermill":{"duration":0.006671,"end_time":"2022-09-16T22:18:28.556379","exception":false,"start_time":"2022-09-16T22:18:28.549708","status":"completed"},"tags":[]},"execution_count":null,"outputs":[]}]}