{"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 numpy as np\nimport pandas as pd\n\n\ndf = pd.read_csv(\"../input/h-and-m-personalized-fashion-recommendations/transactions_train.csv\", dtype={\"article_id\": str})\nprint(df.shape)\ndf.head()","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2022-08-18T02:02:48.145279Z","iopub.execute_input":"2022-08-18T02:02:48.145564Z","iopub.status.idle":"2022-08-18T02:03:20.775094Z","shell.execute_reply.started":"2022-08-18T02:02:48.145529Z","shell.execute_reply":"2022-08-18T02:03:20.774353Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df[\"t_dat\"] = pd.to_datetime(df[\"t_dat\"])\ndf[\"t_dat\"].max()","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:03:27.137762Z","iopub.execute_input":"2022-08-18T02:03:27.138512Z","iopub.status.idle":"2022-08-18T02:03:31.964212Z","shell.execute_reply.started":"2022-08-18T02:03:27.138467Z","shell.execute_reply":"2022-08-18T02:03:31.963338Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"active_articles = df.groupby(\"article_id\")[\"t_dat\"].max().reset_index()\nactive_articles = active_articles[active_articles[\"t_dat\"] >= \"2019-09-01\"].reset_index()\nactive_articles.shape","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:03:41.220806Z","iopub.execute_input":"2022-08-18T02:03:41.221078Z","iopub.status.idle":"2022-08-18T02:03:46.57329Z","shell.execute_reply.started":"2022-08-18T02:03:41.221047Z","shell.execute_reply":"2022-08-18T02:03:46.57262Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df = df[df[\"article_id\"].isin(active_articles[\"article_id\"])].reset_index(drop=True)\ndf.shape","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:03:48.305016Z","iopub.execute_input":"2022-08-18T02:03:48.305732Z","iopub.status.idle":"2022-08-18T02:03:56.037093Z","shell.execute_reply.started":"2022-08-18T02:03:48.305694Z","shell.execute_reply":"2022-08-18T02:03:56.036347Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df[\"week\"] = (df[\"t_dat\"].max() - df[\"t_dat\"]).dt.days // 7\ndf[\"week\"].value_counts()","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:03:58.734359Z","iopub.execute_input":"2022-08-18T02:03:58.735071Z","iopub.status.idle":"2022-08-18T02:04:00.214549Z","shell.execute_reply.started":"2022-08-18T02:03:58.735035Z","shell.execute_reply":"2022-08-18T02:04:00.213687Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from sklearn.preprocessing import LabelEncoder\n\n\narticle_ids = np.concatenate([[\"placeholder\"], np.unique(df[\"article_id\"].values)])\n\nle_article = LabelEncoder()\nle_article.fit(article_ids)\ndf[\"article_id\"] = le_article.transform(df[\"article_id\"])","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:04:13.745632Z","iopub.execute_input":"2022-08-18T02:04:13.746377Z","iopub.status.idle":"2022-08-18T02:05:01.182265Z","shell.execute_reply.started":"2022-08-18T02:04:13.74634Z","shell.execute_reply":"2022-08-18T02:05:01.18149Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"WEEK_HIST_MAX = 5\n\ndef create_dataset(df, week):\n    hist_df = df[(df[\"week\"] > week) & (df[\"week\"] <= week + WEEK_HIST_MAX)]\n    hist_df = hist_df.groupby(\"customer_id\").agg({\"article_id\": list, \"week\": list}).reset_index()\n    hist_df.rename(columns={\"week\": 'week_history'}, inplace=True)\n    \n    target_df = df[df[\"week\"] == week]\n    target_df = target_df.groupby(\"customer_id\").agg({\"article_id\": list}).reset_index()\n    target_df.rename(columns={\"article_id\": \"target\"}, inplace=True)\n    target_df[\"week\"] = week\n    \n    return target_df.merge(hist_df, on=\"customer_id\", how=\"left\")\n\nval_weeks = [0]\ntrain_weeks = [1, 2, 3, 4]\n\n\nval_df = pd.concat([create_dataset(df, w) for w in val_weeks]).reset_index(drop=True)\ntrain_df = pd.concat([create_dataset(df, w) for w in train_weeks]).reset_index(drop=True)\ntrain_df.shape, val_df.shape","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:05:11.772632Z","iopub.execute_input":"2022-08-18T02:05:11.773096Z","iopub.status.idle":"2022-08-18T02:05:43.412201Z","shell.execute_reply.started":"2022-08-18T02:05:11.773062Z","shell.execute_reply":"2022-08-18T02:05:43.411344Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from torch.utils.data import Dataset, DataLoader\nimport torch\nfrom tqdm import tqdm\n\nclass HMDataset(Dataset):\n    def __init__(self, df, seq_len, is_test=False):\n        self.df = df.reset_index(drop=True)\n        self.seq_len = seq_len\n        self.is_test = is_test\n    \n    def __len__(self):\n        return self.df.shape[0]\n    \n    def __getitem__(self, index):\n        row = self.df.iloc[index]\n        \n        if self.is_test:\n            target = torch.zeros(2).float()\n        else:\n            rand_target = np.random.choice(row.target,1)\n            target = torch.tensor(rand_target).squeeze().int()\n            #target = torch.zeros(len(article_ids)).float()\n            #for t in row.target:\n                #target[t] = 1.0\n            \n        article_hist = torch.zeros(self.seq_len).long()\n        week_hist = torch.ones(self.seq_len).float()\n        \n        \n        if isinstance(row.article_id, list):\n            if len(row.article_id) >= self.seq_len:\n                article_hist = torch.LongTensor(row.article_id[-self.seq_len:])\n                week_hist = (torch.LongTensor(row.week_history[-self.seq_len:]) - row.week)/WEEK_HIST_MAX/2\n            else:\n                article_hist[-len(row.article_id):] = torch.LongTensor(row.article_id)\n                week_hist[-len(row.article_id):] = (torch.LongTensor(row.week_history) - row.week)/WEEK_HIST_MAX/2\n                \n        return article_hist, week_hist, target\n    \n#HMDataset(val_df, 64)[1]\nHMDataset(val_df, 64)[2]","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:05:46.580699Z","iopub.execute_input":"2022-08-18T02:05:46.581402Z","iopub.status.idle":"2022-08-18T02:05:48.199346Z","shell.execute_reply.started":"2022-08-18T02:05:46.581364Z","shell.execute_reply":"2022-08-18T02:05:48.198296Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def adjust_lr(optimizer, epoch):\n    if epoch < 1:\n        lr = 5e-5\n    elif epoch < 6:\n        lr = 1e-3\n    elif epoch < 9:\n        lr = 1e-4\n    else:\n        lr = 1e-5\n\n    for p in optimizer.param_groups:\n        p['lr'] = lr\n    return lr\n    \ndef get_optimizer(net):\n    optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, net.parameters()), lr=3e-4, betas=(0.9, 0.999),\n                                 eps=1e-08)\n    return optimizer","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:05:50.808121Z","iopub.execute_input":"2022-08-18T02:05:50.808808Z","iopub.status.idle":"2022-08-18T02:05:50.814907Z","shell.execute_reply.started":"2022-08-18T02:05:50.808768Z","shell.execute_reply":"2022-08-18T02:05:50.813918Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import torch.nn as nn\nclass HierarchicalSoftmax(nn.Module):\n    def __init__(self, ntokens, nhid, ntokens_per_class = None):\n        super(HierarchicalSoftmax, self).__init__()\n\n        # Parameters\n        self.ntokens = ntokens\n        self.nhid = nhid\n\n#         if ntokens_per_class is None:\n#             ntokens_per_class = int(np.ceil(np.sqrt(ntokens)))\n\n        self.ntokens_per_class = ntokens_per_class\n\n        self.nclasses = int(np.ceil(self.ntokens * 1. / self.ntokens_per_class))\n        self.ntokens_actual = self.nclasses * self.ntokens_per_class\n\n        self.layer_top_W = nn.Parameter(torch.FloatTensor(self.nhid, self.nclasses), requires_grad=True)\n        self.layer_top_b = nn.Parameter(torch.FloatTensor(self.nclasses), requires_grad=True)\n\n        self.layer_bottom_W = nn.Parameter(torch.FloatTensor(self.ntokens_per_class, self.nhid), requires_grad=True)\n        self.layer_bottom_b = nn.Parameter(torch.FloatTensor(self.nclasses), requires_grad=True)\n\n       # self.softmax = nn.Softmax(dim=1)\n\n        self.init_weights()\n\n    def init_weights(self):\n\n        initrange = 0.1\n        self.layer_top_W.data.uniform_(-initrange, initrange)\n        self.layer_top_b.data.fill_(0)\n        self.layer_bottom_W.data.uniform_(-initrange, initrange)\n        self.layer_bottom_b.data.fill_(0)\n\n\n    def forward(self, inputs):\n        # add labels\n        labels = torch.arange(self.ntokens_actual)\n\n        batch_size, d = inputs.size()\n        \n        if labels is not None:\n\n            label_position_top = (labels / self.ntokens_per_class).long()\n            label_position_bottom = (labels % self.ntokens_per_class).long()\n\n            layer_top_logits = torch.matmul(inputs, self.layer_top_W) + self.layer_top_b\n           # layer_top_probs = self.softmax(layer_top_logits)\n        \n            multi_bias = self.layer_bottom_b[label_position_top].repeat(batch_size,1)\n            \n            layer_bottom_logits = torch.matmul(inputs,self.layer_bottom_W[label_position_top].T) + multi_bias\n\n            #layer_bottom_logits = torch.squeeze(torch.bmm(torch.unsqueeze(inputs, dim=1), self.layer_bottom_W[label_position_top]), dim=1) + self.layer_bottom_b[label_position_top]\n            #layer_bottom_probs = self.softmax(layer_bottom_logits)\n\n            #target_probs = layer_top_probs[torch.arange(batch_size).long(), label_position_top] * layer_bottom_probs[torch.arange(batch_size).long(), label_position_bottom]\n            \n            layer_top_logits = layer_top_logits.repeat_interleave(self.ntokens_per_class,dim=1)\n            target_logits = torch.mul(layer_top_logits,layer_bottom_logits)\n\n            return target_logits\n\n#         else:\n#             # Remain to be implemented\n#             layer_top_logits = torch.matmul(inputs, self.layer_top_W) + self.layer_top_b\n#             layer_top_probs = self.softmax(layer_top_logits)\n\n#             word_probs = layer_top_probs[:,0] * self.softmax(torch.matmul(inputs, self.layer_bottom_W[0]) + self.layer_bottom_b[0])\n\n#             for i in range(1, self.nclasses):\n#                 word_probs = torch.cat((word_probs, layer_top_probs[:,i] * self.softmax(torch.matmul(inputs, self.layer_bottom_W[i]) + self.layer_bottom_b[i])), dim=1)\n\n#             return word_probs","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:05:54.812414Z","iopub.execute_input":"2022-08-18T02:05:54.812874Z","iopub.status.idle":"2022-08-18T02:05:54.826883Z","shell.execute_reply.started":"2022-08-18T02:05:54.812839Z","shell.execute_reply":"2022-08-18T02:05:54.826062Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import torch.nn.functional as F\n\nclass HMModel(nn.Module):\n    def __init__(self, article_shape):\n        super(HMModel, self).__init__()\n        \n        self.article_emb = nn.Embedding(article_shape[0], embedding_dim=article_shape[1])\n        \n       # self.article_likelihood = nn.Parameter(torch.zeros(article_shape[0]), requires_grad=True)\n        \n        # No Conv1D\n#         self.top = nn.Sequential(nn.Conv1d(3, 32, kernel_size=1), nn.LeakyReLU(),\n#                                  nn.Conv1d(32, 8, kernel_size=1), nn.LeakyReLU(),\n#                                  nn.Conv1d(8, 1, kernel_size=1))\n#         self.softmax = nn.Softmax(dim=1)\n\n        self.hierSoftmax = HierarchicalSoftmax(72582, 512,ntokens_per_class = 20)\n        \n    def forward(self, inputs):\n        article_hist, week_hist = inputs[0], inputs[1]\n        x = self.article_emb(article_hist)\n        x = F.normalize(x, dim=2)\n        \n        #x = x@F.normalize(self.article_emb.weight).T\n        \n        x, indices = x.max(axis=1)\n        \n        # obtain logits\n        logits = self.hierSoftmax(x)\n        logits = logits[:,:72582]\n        \n        # simplifiy logic\n#         x = self.softmax(x)\n        \n#         x = x.clamp(1e-3, 0.999)\n#         x = -torch.log(1/x - 1)\n        \n#         max_week = week_hist.unsqueeze(2).repeat(1, 1, x.shape[-1]).gather(1, indices.unsqueeze(1).repeat(1, week_hist.shape[1], 1))\n#         max_week = max_week.mean(axis=1).unsqueeze(1)\n        \n#         x = torch.cat([x.unsqueeze(1), max_week,\n#                        self.article_likelihood[None, None, :].repeat(x.shape[0], 1, 1)], axis=1)\n        \n#         x = self.top(x).squeeze(1)\n\n        return logits\n    \n    \nmodel = HMModel((len(le_article.classes_), 512))\nmodel = model.cuda()","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:06:32.607761Z","iopub.execute_input":"2022-08-18T02:06:32.608622Z","iopub.status.idle":"2022-08-18T02:06:36.458863Z","shell.execute_reply.started":"2022-08-18T02:06:32.608583Z","shell.execute_reply":"2022-08-18T02:06:36.458061Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import sys\n\ndef calc_map(topk_preds, target_array, k=12):\n    metric = []\n    tp, fp = 0, 0\n    \n    for pred in topk_preds:\n        if target_array[pred]:\n            tp += 1\n            metric.append(tp/(tp + fp))\n        else:\n            fp += 1\n            \n    return np.sum(metric) / min(k, target_array.sum())\n\ndef read_data(data):\n    return tuple(d.cuda() for d in data[:-1]), data[-1].cuda()\n\n\ndef validate(model, val_loader, k=12):\n    model.eval()\n    \n    tbar = tqdm(val_loader, file=sys.stdout)\n    \n    maps = []\n    \n    with torch.no_grad():\n        for idx, data in enumerate(tbar):\n            inputs, target = read_data(data)\n\n            logits = model(inputs)\n\n            _, indices = torch.topk(logits, k, dim=1)\n\n            indices = indices.detach().cpu().numpy()\n            target = target.detach().cpu().numpy()\n\n            for i in range(indices.shape[0]):\n                maps.append(calc_map(indices[i], target[i]))\n        \n    \n    return np.mean(maps)\n\nSEQ_LEN = 16\n\nBS = 256\nNW = 8\n\nval_dataset = HMDataset(val_df, SEQ_LEN)\nval_loader = DataLoader(val_dataset, batch_size=BS, shuffle=False, num_workers=NW,\n                          pin_memory=False, drop_last=False)","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:06:41.694658Z","iopub.execute_input":"2022-08-18T02:06:41.695595Z","iopub.status.idle":"2022-08-18T02:06:41.712991Z","shell.execute_reply.started":"2022-08-18T02:06:41.695551Z","shell.execute_reply":"2022-08-18T02:06:41.712242Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Train and validate","metadata":{}},{"cell_type":"code","source":"def dice_loss(y_pred, y_true):\n    y_pred = y_pred.sigmoid()\n    intersect = (y_true*y_pred).sum(axis=1)\n    \n    return 1 - (intersect/(intersect + y_true.sum(axis=1) + y_pred.sum(axis=1))).mean()\n\n\ndef train(model, train_loader, val_loader, epochs):\n    np.random.seed(SEED)\n    \n    optimizer = get_optimizer(model)\n    scaler = torch.cuda.amp.GradScaler()\n\n    criterion = torch.nn.BCEWithLogitsLoss()\n    \n    for e in range(epochs):\n        model.train()\n        tbar = tqdm(train_loader, file=sys.stdout)\n        \n        lr = adjust_lr(optimizer, e)\n        \n        loss_list = []\n\n        for idx, data in enumerate(tbar):\n            inputs, target = read_data(data)\n\n            optimizer.zero_grad()\n            \n            with torch.cuda.amp.autocast():\n                logits = model(inputs)\n                loss = criterion(logits, target) + dice_loss(logits, target)\n            \n            \n            #loss.backward()\n            scaler.scale(loss).backward()\n            #optimizer.step()\n            scaler.step(optimizer)\n            scaler.update()\n            \n            loss_list.append(loss.detach().cpu().item())\n            \n            avg_loss = np.round(100*np.mean(loss_list), 4)\n\n            tbar.set_description(f\"Epoch {e+1} Loss: {avg_loss} lr: {lr}\")\n            \n      #  val_map = validate(model, val_loader)\n\n        #log_text = f\"Epoch {e+1}\\nTrain Loss: {avg_loss}\\nValidation MAP: {val_map}\\n\"\n            \n       # print(log_text)\n        \n        #logfile = open(f\"models/{MODEL_NAME}_{SEED}.txt\", 'a')\n        #logfile.write(log_text)\n        #logfile.close()\n    return model\n\n\nMODEL_NAME = \"exp001\"\nSEED = 0\n\ntrain_dataset = HMDataset(train_df, SEQ_LEN)\ntrain_loader = DataLoader(train_dataset, batch_size=BS, shuffle=True, num_workers=NW,\n                          pin_memory=False, drop_last=True)\n\nmodel = train(model, train_loader, val_loader, epochs=10)","metadata":{"execution":{"iopub.status.busy":"2022-08-18T02:06:47.613631Z","iopub.execute_input":"2022-08-18T02:06:47.614196Z","iopub.status.idle":"2022-08-18T02:06:52.522825Z","shell.execute_reply.started":"2022-08-18T02:06:47.614159Z","shell.execute_reply":"2022-08-18T02:06:52.521689Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### Finetune with more recent data for submission (include validation set)","metadata":{}},{"cell_type":"code","source":"train_dataset = HMDataset(train_df[train_df[\"week\"] < 4].append(val_df), SEQ_LEN)\ntrain_loader = DataLoader(train_dataset, batch_size=BS, shuffle=True, num_workers=NW,\n                          pin_memory=False, drop_last=True)\n\nmodel = train(model, train_loader, val_loader, epochs=10)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_df = pd.read_csv('../input/h-and-m-personalized-fashion-recommendations/sample_submission.csv').drop(\"prediction\", axis=1)\nprint(test_df.shape)\ntest_df.head()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def create_test_dataset(test_df):\n    week = -1\n    test_df[\"week\"] = week\n    \n    hist_df = df[(df[\"week\"] > week) & (df[\"week\"] <= week + WEEK_HIST_MAX)]\n    hist_df = hist_df.groupby(\"customer_id\").agg({\"article_id\": list, \"week\": list}).reset_index()\n    hist_df.rename(columns={\"week\": 'week_history'}, inplace=True)\n    \n    \n    return test_df.merge(hist_df, on=\"customer_id\", how=\"left\")\n\ntest_df = create_test_dataset(test_df)\ntest_df.head()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_df[\"article_id\"].isnull().mean()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_ds = HMDataset(test_df, SEQ_LEN, is_test=True)\ntest_loader = DataLoader(test_ds, batch_size=BS, shuffle=False, num_workers=NW,\n                          pin_memory=False, drop_last=False)\n\n\ndef inference(model, loader, k=12):\n    model.eval()\n    \n    tbar = tqdm(loader, file=sys.stdout)\n    \n    preds = []\n    \n    with torch.no_grad():\n        for idx, data in enumerate(tbar):\n            inputs, target = read_data(data)\n\n            logits = model(inputs)\n\n            _, indices = torch.topk(logits, k, dim=1)\n\n            indices = indices.detach().cpu().numpy()\n            target = target.detach().cpu().numpy()\n\n            for i in range(indices.shape[0]):\n                preds.append(\" \".join(list(le_article.inverse_transform(indices[i]))))\n        \n    \n    return preds\n\n\ntest_df[\"prediction\"] = inference(model, test_loader)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_df.to_csv(\"submission.csv\", index=False, columns=[\"customer_id\", \"prediction\"])","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}