{"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":"import os\nimport gc\nimport pickle\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\nimport torch\n\nimport torch.nn as nn\nimport matplotlib.pyplot as plt\n\nfrom sklearn.model_selection import StratifiedKFold","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2022-06-27T13:36:50.317501Z","iopub.execute_input":"2022-06-27T13:36:50.31841Z","iopub.status.idle":"2022-06-27T13:36:50.326536Z","shell.execute_reply.started":"2022-06-27T13:36:50.318359Z","shell.execute_reply":"2022-06-27T13:36:50.325343Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Gru Baseline Reference:\nhttps://www.kaggle.com/code/cdeotte/tensorflow-gru-starter-0-790/notebook","metadata":{}},{"cell_type":"markdown","source":"# Config","metadata":{}},{"cell_type":"code","source":"class CFG:\n    BATCH_SIZE=512\n    N_EPOCHS=8","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:36:50.348635Z","iopub.execute_input":"2022-06-27T13:36:50.349267Z","iopub.status.idle":"2022-06-27T13:36:50.355501Z","shell.execute_reply.started":"2022-06-27T13:36:50.349231Z","shell.execute_reply":"2022-06-27T13:36:50.354689Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"device=torch.device('cuda' if torch.cuda.is_available() else 'cpu')\nprint(device)","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:36:50.363627Z","iopub.execute_input":"2022-06-27T13:36:50.365182Z","iopub.status.idle":"2022-06-27T13:36:50.372721Z","shell.execute_reply.started":"2022-06-27T13:36:50.365128Z","shell.execute_reply":"2022-06-27T13:36:50.371936Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# loading the dataset","metadata":{}},{"cell_type":"code","source":"%%time\nallX=[]\nally=[]\n\nfor i in range(100):\n    xpath = \"../input/amex-gru-train-dataset/Xchunk_{}.npy\".format(i)\n    ypath = \"../input/amex-gru-train-dataset/ychunk_{}.npy\".format(i)\n    \n    if not os.path.exists(xpath):\n        break\n    allX.append(np.load(xpath))\n    ally.append(np.load(ypath))\n    gc.collect()\n    \nallX=np.concatenate(allX)\nally=np.concatenate(ally)\nprint(allX.shape, ally.shape)","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:36:50.417392Z","iopub.execute_input":"2022-06-27T13:36:50.418165Z","iopub.status.idle":"2022-06-27T13:36:59.569614Z","shell.execute_reply.started":"2022-06-27T13:36:50.41812Z","shell.execute_reply":"2022-06-27T13:36:59.567731Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# dataset","metadata":{}},{"cell_type":"code","source":"class AmexDataset(torch.utils.data.Dataset):\n    def __init__(self, X, y, idxs, phase='train'):\n        self.idxs=idxs\n        self.X = X\n        self.y = y\n        self.phase=phase\n    \n    def __getitem__(self, idx):\n        idx=self.idxs[idx]\n        X = torch.tensor(self.X[idx], dtype=torch.float32)\n        if self.phase !='train':\n            return X\n        y = torch.tensor(self.y[idx], dtype=torch.float32)\n        return (X, y)\n    \n    def __len__(self):\n        return len(self.idxs)","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:36:59.572952Z","iopub.execute_input":"2022-06-27T13:36:59.573583Z","iopub.status.idle":"2022-06-27T13:36:59.5862Z","shell.execute_reply.started":"2022-06-27T13:36:59.573528Z","shell.execute_reply":"2022-06-27T13:36:59.584339Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Model","metadata":{}},{"cell_type":"code","source":"class AmexGruModel(nn.Module):\n    def __init__(self):\n        super().__init__()\n        self.bn=nn.BatchNorm1d(169)\n        \n        self.mlp=nn.Sequential(\n            nn.Linear(169, 512),\n            nn.LeakyReLU(),\n            nn.Dropout(0.1),\n            nn.Linear(512, 256),\n            nn.Dropout(0.1)\n        )\n        \n        self.gru = nn.GRU(256, 128, batch_first = True)\n        self.out = nn.Sequential(\n            nn.Dropout(0.1),\n            \n            nn.Linear(128, 64),\n            nn.ReLU(),\n            \n            nn.BatchNorm1d(64),\n            nn.Linear(64, 32),\n            nn.ReLU(),\n            \n            nn.BatchNorm1d(32),\n            nn.Linear(32, 1),\n        )\n    def forward(self, x):\n        x = x.permute(0, 2, 1)\n        x = self.bn(x)\n        x = x.permute(0, 2, 1)\n        x = self.mlp(x)\n        \n        \n        (_, h) = self.gru(x)\n        h = h.squeeze(0)\n        y = self.out(h).view(-1)\n        return y","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:36:59.589484Z","iopub.execute_input":"2022-06-27T13:36:59.591815Z","iopub.status.idle":"2022-06-27T13:36:59.609927Z","shell.execute_reply.started":"2022-06-27T13:36:59.591735Z","shell.execute_reply":"2022-06-27T13:36:59.608014Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Metrics","metadata":{}},{"cell_type":"code","source":"def top_4percent(pred_df):\n    df = pred_df.copy()\n    df = df.sort_values('pred', ascending=False)\n    df['weight'] = df['target'].apply(lambda v: 20 if v==0 else 1)\n    four_percent_cutoff = 0.04 * sum(df['weight'])\n    df['weight_cumsum'] = df['weight'].cumsum()\n    df_cutoff = df[df.weight_cumsum <= four_percent_cutoff]\n    \n    return df_cutoff['target'].sum()/df['target'].sum()\n\ndef weighted_gini(pred_df):\n    df = pred_df.copy()\n    df = df.sort_values('pred', ascending=False)\n    df['weight'] = df['target'].apply(lambda v: 20 if v==0 else 1)\n    df['random'] = (df['weight'] / df['weight'].sum()).cumsum()\n    total_pos = (df['target'] * df['weight']).sum()\n    df['cum_pos_found'] = (df['target'] * df['weight']).cumsum()\n    df['lorentz'] = df['cum_pos_found'] / total_pos\n    df['gini'] = (df['lorentz'] - df['random']) * df['weight']\n    return df['gini'].sum()\n\n\ndef normalized_gini(df):\n    df_true=df[['target']].copy()\n    df_true['pred'] = df_true['target'].copy()\n    \n    G = weighted_gini(df)/weighted_gini(df_true)\n    return G","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:36:59.614015Z","iopub.execute_input":"2022-06-27T13:36:59.615296Z","iopub.status.idle":"2022-06-27T13:36:59.632131Z","shell.execute_reply.started":"2022-06-27T13:36:59.615225Z","shell.execute_reply":"2022-06-27T13:36:59.630783Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# losses","metadata":{}},{"cell_type":"code","source":"def label_smoothing(y, yhat):\n    y = torch.clamp(y, 0.01, 0.99)\n    loss = -y*torch.log(torch.sigmoid(yhat)) - (1-y) * torch.log(1-torch.sigmoid(yhat))\n    return loss.mean()\n\ndef get_rank_loss(y, yhat):\n    loss = torch.tensor(0.0, device=device)\n    ypos = yhat[y==1]\n    yneg = yhat[y==0]\n    \n    if len(ypos) == 0 or len(yneg) == 0:\n        return loss\n    \n    yneg = yneg.repeat((len(ypos), 1))\n    ypos = ypos.unsqueeze(dim=-1)\n    loss = -torch.log( torch.sigmoid( ypos-yneg ) ).mean()\n    return loss","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:49:11.036891Z","iopub.execute_input":"2022-06-27T13:49:11.037382Z","iopub.status.idle":"2022-06-27T13:49:11.049898Z","shell.execute_reply.started":"2022-06-27T13:49:11.037346Z","shell.execute_reply":"2022-06-27T13:49:11.04829Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# train model","metadata":{}},{"cell_type":"code","source":"def get_lr(epoch_num):\n    lrs = [1e-3, 1e-3, 1e-3, 1e-4, 1e-4, 1e-4, 1e-5, 1e-5]\n    if epoch_num < len(lrs):\n        return lrs[foldnum]\n    return 1e-5","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:49:37.248599Z","iopub.execute_input":"2022-06-27T13:49:37.25011Z","iopub.status.idle":"2022-06-27T13:49:37.256446Z","shell.execute_reply.started":"2022-06-27T13:49:37.250051Z","shell.execute_reply":"2022-06-27T13:49:37.254965Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def evaluate(model, val_dataloader):\n    model.eval()\n    ytrue=[]\n    ypred=[]\n    \n    for X,y in val_dataloader:\n        X = X.to(device)\n        y = y.to(device)\n        with torch.no_grad():\n            yhat=model(X)\n            yhat = yhat.sigmoid()\n            ytrue += y.cpu().tolist()\n            ypred += yhat.cpu().tolist()\n    \n    df = pd.DataFrame.from_dict({\n        'target': ytrue,\n        'pred': ypred\n    })\n    \n    G = normalized_gini(df[['target', 'pred']])\n    D = top_4percent(df[['target', 'pred']])\n    \n    M = (G+D)/2\n    return (G, D, M)","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:49:37.329722Z","iopub.execute_input":"2022-06-27T13:49:37.330824Z","iopub.status.idle":"2022-06-27T13:49:37.340958Z","shell.execute_reply.started":"2022-06-27T13:49:37.330781Z","shell.execute_reply":"2022-06-27T13:49:37.339626Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def train_model(foldnum, train_dataloader, val_dataloader):\n    best_eval=None\n    \n    model = AmexGruModel().to(device)\n    criterion = nn.BCEWithLogitsLoss()\n    optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3, weight_decay=1e-6)\n    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, \n                                                           T_max = CFG.N_EPOCHS * len(train_dataloader), \n                                                           eta_min=1e-5)\n    \n    for e in range(CFG.N_EPOCHS):\n        epoch_loss=[]\n        epoch_rank_loss=[]\n        \n        model.train()\n        \n        for it, (X, y) in enumerate(train_dataloader):\n            X = X.to(device)\n            y = y.to(device)\n            \n            yhat = model(X)\n            loss1 = criterion(yhat, y)\n            rank_loss = get_rank_loss(y, yhat)\n            #loss2 = criterion(yhat, 1-y)\n            #loss = 0.99 * loss1 + 0.01*loss2\n            loss = loss1 + 0.4*rank_loss\n            \n            optimizer.zero_grad(set_to_none=True)\n            loss.backward()\n            optimizer.step()\n            scheduler.step()\n\n            epoch_loss.append(loss.item())\n            epoch_rank_loss.append(rank_loss.item())\n        \n        #Evaluating\n        (G, D, M) = evaluate(model, val_dataloader)\n        if best_eval is None or best_eval<M:\n            best_eval = M\n            torch.save(model, \"model{}.pt\".format(foldnum))\n        \n        \n        print(\"epoch:{} | loss:{:.4f} | rank loss:{:.4f}\".format(e, np.mean(epoch_loss), np.mean(epoch_rank_loss)))\n        print(\"current Eval:{:.4f} | best Eval:{:.4f}\".format(M, best_eval))\n        print(\"Gini:{:.4f} | Default Rate:{:4f}\".format(G, D))\n        print()\n        print()\n        \n        plt.title(\"train epoch loss.\")\n        plt.plot(epoch_loss)\n        plt.show()\n        \n        plt.title(\"train epoch rank loss.\")\n        plt.plot(epoch_rank_loss)\n        plt.show()\n","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:50:28.331228Z","iopub.execute_input":"2022-06-27T13:50:28.331662Z","iopub.status.idle":"2022-06-27T13:50:28.351274Z","shell.execute_reply.started":"2022-06-27T13:50:28.331628Z","shell.execute_reply":"2022-06-27T13:50:28.349816Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"skf = StratifiedKFold(n_splits=5, random_state=33, shuffle=True)\nfor foldnum, (train_index, val_index) in enumerate(skf.split(ally, ally)):\n    train_dataset = AmexDataset(allX,ally, train_index)\n    val_dataset = AmexDataset(allX, ally, val_index)\n    \n    train_dataloader = torch.utils.data.DataLoader(train_dataset, \n                                                   batch_size=CFG.BATCH_SIZE, \n                                                   shuffle=True,\n                                                   drop_last=True)\n    \n    val_dataloader = torch.utils.data.DataLoader(val_dataset, batch_size=CFG.BATCH_SIZE, \n                                                   shuffle=False,\n                                                   drop_last=False)\n    \n    \n    print(\"Foldnumber:\", foldnum)\n    print(\"number of train iterations:\", len(train_dataloader))\n    print(\"number of val iterations:\", len(val_dataloader))\n    \n    train_model(foldnum, train_dataloader, val_dataloader)","metadata":{"execution":{"iopub.status.busy":"2022-06-27T13:50:30.491436Z","iopub.execute_input":"2022-06-27T13:50:30.492536Z","iopub.status.idle":"2022-06-27T13:51:03.302015Z","shell.execute_reply.started":"2022-06-27T13:50:30.492483Z","shell.execute_reply":"2022-06-27T13:51:03.300602Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Inference","metadata":{}},{"cell_type":"code","source":"def load_pickle_obj(filename):\n    with open(filename, 'rb') as file:\n        obj = pickle.load(file)\n    return obj\n\ntest_id2customer = load_pickle_obj(\"../input/amex-datasetcategorical-encoders/test_id2customer.pkl\")\nprint(len(test_id2customer))","metadata":{"execution":{"iopub.status.busy":"2022-06-27T10:16:13.565728Z","iopub.status.idle":"2022-06-27T10:16:13.566112Z","shell.execute_reply.started":"2022-06-27T10:16:13.565916Z","shell.execute_reply":"2022-06-27T10:16:13.565935Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"models=[]\nfor i in range(5):\n    model = torch.load(\"model{}.pt\".format(i))\n    models.append(model)","metadata":{"execution":{"iopub.status.busy":"2022-06-27T07:31:08.211064Z","iopub.status.idle":"2022-06-27T07:31:08.21211Z","shell.execute_reply.started":"2022-06-27T07:31:08.211615Z","shell.execute_reply":"2022-06-27T07:31:08.211645Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sub_df = []\nfor fileid in range(200):\n    xpath = \"../input/amex-rnn-chunk-dataset/Xchunk_{}.npy\".format(fileid)\n    customer_path = \"../input/amex-rnn-chunk-dataset/customerIds_chunk_{}.npy\".format(fileid)\n    \n    if os.path.exists(xpath):\n        Xtest = np.load(xpath)\n        customerids = np.load(customer_path)\n        test_ids = np.arange(len(Xtest))\n        test_dataset = AmexDataset(Xtest, None, test_ids, phase=\"infer\")\n        test_loader  = torch.utils.data.DataLoader(test_dataset, shuffle=False, drop_last=False, batch_size=512)\n        \n        all_preds=[]\n        for X in test_loader:\n            X = X.to(device)\n            preds=np.zeros(len(X))\n            \n            for model in models:\n                model.eval()\n                with torch.no_grad():\n                    yhat = model(X).sigmoid()\n                    preds += yhat.cpu().numpy()\n            preds = preds/len(models)\n            all_preds += list(preds)\n        \n        df = pd.DataFrame.from_dict({\n            'customer_ID': customerids,\n            'prediction': all_preds\n        })\n        df.fillna(0.0, inplace=True)\n        sub_df.append(df)","metadata":{"execution":{"iopub.status.busy":"2022-06-27T07:31:08.214229Z","iopub.status.idle":"2022-06-27T07:31:08.21527Z","shell.execute_reply.started":"2022-06-27T07:31:08.214966Z","shell.execute_reply":"2022-06-27T07:31:08.214996Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sub_df=pd.concat(sub_df)\nsub_df['customer_ID'] = sub_df['customer_ID'].apply(lambda k: test_id2customer[k])\nsub_df.head()\n","metadata":{"execution":{"iopub.status.busy":"2022-06-27T07:31:08.217495Z","iopub.status.idle":"2022-06-27T07:31:08.218443Z","shell.execute_reply.started":"2022-06-27T07:31:08.218118Z","shell.execute_reply":"2022-06-27T07:31:08.21816Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sub_df.shape","metadata":{"execution":{"iopub.status.busy":"2022-06-27T07:31:08.220516Z","iopub.status.idle":"2022-06-27T07:31:08.221591Z","shell.execute_reply.started":"2022-06-27T07:31:08.221268Z","shell.execute_reply":"2022-06-27T07:31:08.221299Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sub_df.to_csv(\"submission.csv\", index=False)\n","metadata":{"execution":{"iopub.status.busy":"2022-06-27T07:31:08.223593Z","iopub.status.idle":"2022-06-27T07:31:08.224865Z","shell.execute_reply.started":"2022-06-27T07:31:08.224342Z","shell.execute_reply":"2022-06-27T07:31:08.224373Z"},"trusted":true},"execution_count":null,"outputs":[]}]}