Merge pull request #18 from recohut/stage

commit

_notebooks/2022-01-24-cdl.ipynb

@@ -0,0 +1 @@
+{"nbformat":4,"nbformat_minor":0,"metadata":{"colab":{"name":"2022-01-24-cdl.ipynb","provenance":[{"file_id":"https://github.com/recohut/nbs/blob/main/raw/T249728%20%7C%20CDL%20on%20ML-1m.ipynb","timestamp":1644665590116},{"file_id":"16z8K9gL8SxhneDP4Mi3L4Opbck-aYxO8","timestamp":1638115875999}],"collapsed_sections":[],"toc_visible":true,"mount_file_id":"16z8K9gL8SxhneDP4Mi3L4Opbck-aYxO8","authorship_tag":"ABX9TyPNtUryDQj38/+BfqMfah1L"},"kernelspec":{"name":"python3","display_name":"Python 3"},"language_info":{"name":"python"}},"cells":[{"cell_type":"markdown","source":["# CDL on ML-1m"],"metadata":{"id":"kFnMXHQetSFs"}},{"cell_type":"code","metadata":{"id":"RuiURj-F17mq"},"source":["import numpy as np\n","import pandas as pd\n","import os, sys\n","from collections import defaultdict\n","\n","import torch\n","import torch.nn as nn\n","from torch.utils.data import Dataset, DataLoader\n","import torch.optim as optim\n","import torch.nn.functional as F"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"colab":{"base_uri":"https://localhost:8080/"},"id":"aB3rWtCGOm0U","executionInfo":{"status":"ok","timestamp":1638115805771,"user_tz":-330,"elapsed":1255,"user":{"displayName":"Sparsh Agarwal","photoUrl":"https://lh3.googleusercontent.com/a/default-user=s64","userId":"13037694610922482904"}},"outputId":"c00a910d-f4e2-4f29-d461-e3e7bed154a7"},"source":["!wget -q --show-progress https://github.com/sparsh-ai/stanza/raw/S629908/rec/CDL/data/ml_100k_train.npy\n","!wget -q --show-progress https://github.com/sparsh-ai/stanza/raw/S629908/rec/CDL/data/ml_100k_test.npy\n","!wget -q --show-progress https://github.com/sparsh-ai/stanza/raw/S629908/rec/CDL/data/movies.csv"],"execution_count":null,"outputs":[{"output_type":"stream","name":"stdout","text":["ml_100k_train.npy   100%[===================>]  12.10M  --.-KB/s    in 0.1s    \n","ml_100k_test.npy    100%[===================>]  12.10M  --.-KB/s    in 0.1s    \n","movies.csv          100%[===================>] 232.34K  --.-KB/s    in 0.03s   \n"]}]},{"cell_type":"code","metadata":{"id":"1nkyosaO7P7h"},"source":["def add_noise(x, corrupt_ratio):\n","    noise = np.random.binomial(1, corrupt_ratio, size=x.shape)\n","    return x + noise\n","\n","class CDLDataset(Dataset):\n","    def __init__(self, xc, x0):\n","        super(CDLDataset, self).__init__()\n","        self.xc = xc\n","        self.x0 = x0\n","    \n","    def __len__(self):\n","        return self.xc.shape[0]\n","\n","    def __getitem__(self, idx):\n","        return {'clean':torch.FloatTensor(self.xc[idx, :]),\n","                'corrupt':torch.FloatTensor(self.x0[idx, :]),\n","                'idx':idx}"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"P3OBm1mg7sZM"},"source":["class SDAE(nn.Module):\n","    def __init__(self, input_dim, hidden_dim, embed_dim):\n","        super(SDAE, self).__init__()\n","        self.input_dim = input_dim\n","        self.hidden_dim = hidden_dim\n","        self.embed_dim = embed_dim\n","        self.enc1 = nn.Linear(input_dim, hidden_dim)\n","        self.enc2 = nn.Linear(hidden_dim, hidden_dim)\n","        self.enc3 = nn.Linear(hidden_dim, embed_dim)\n","        \n","        self.dec1 = nn.Linear(embed_dim, hidden_dim)\n","        self.dec2 = nn.Linear(hidden_dim, hidden_dim)\n","        self.dec3 = nn.Linear(hidden_dim, input_dim)\n","\n","    def forward(self, x):\n","        x = F.relu(self.enc1(x))\n","        x = F.relu(self.enc2(x))\n","        latent = F.relu(self.enc3(x))\n","        x = F.relu(self.dec1(latent))\n","        x = F.relu(self.dec2(x))\n","        x = self.dec3(x)\n","        return latent, x"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"4PCM4TvX7dgY"},"source":["class CDL:\n","    def __init__(self, train_imp, test_imp, input_dim, hidden_dim, dim_f, dataloader, seed, device, config):\n","        self.dim_f = dim_f\n","        self.user_num = train_imp.shape[0]\n","        self.item_num = train_imp.shape[1]\n","        self.input_dim = input_dim\n","\n","        self.R_tr = train_imp\n","        self.R_tst = test_imp\n","        self.C = np.where(self.R_tr > 0, 1, 0)\n","        self.C_u = np.zeros((self.item_num, self.item_num))\n","        self.C_i = np.zeros((self.user_num, self.user_num))\n","\n","        np.random.seed(seed)\n","        self.X = np.random.standard_normal((self.user_num, dim_f))\n","        self.Y = np.random.standard_normal((self.item_num, dim_f))\n","        \n","        self.loss_tr = defaultdict(float)\n","        self.loss_ae = defaultdict(float)\n","        self.loss_tst = defaultdict(float)\n","\n","        self.ae = SDAE(input_dim=input_dim, hidden_dim=hidden_dim, embed_dim=dim_f).to(device)\n","        self.optimizer = optim.Adam(self.ae.parameters(), lr=config.learning_rate, weight_decay=config.lambda_w)\n","        self.dataloader = dataloader\n","\n","        self.lambda_u = config.lambda_u\n","        self.lambda_w = config.lambda_w\n","        self.lambda_v = config.lambda_v\n","        self.lambda_n = config.lambda_n\n","\n","        self.device = device\n","        self.config = config\n","    \n","    def ae_train(self):\n","        latent_np = np.zeros((self.item_num, self.dim_f))\n","        loss_ae = []\n","        for batch in self.dataloader:\n","            y = batch['clean'].to(self.device)\n","            x = batch['corrupt'].to(self.device)\n","            idx = batch['idx']\n","            latent, pred = self.ae(x)\n","            latent_ = latent.detach().cpu().numpy()\n","            latent_np[idx.numpy()] = latent_\n","\n","            loss = self.loss_fn(pred, y, idx.to(self.device), latent_)\n","            loss.backward()\n","            self.optimizer.step()\n","            loss_ae.append(loss.item())\n","\n","        return latent_np, np.mean(loss_ae)\n","\n","    def fit(self):\n","        start = datetime.now()\n","        for epoch in range(self.config.epochs):\n","            start_epoch = datetime.now()\n","            self.ae.train()\n","            self.latent_feat, self.loss_ae[epoch] = self.ae_train()\n","            n = 0\n","            for u in range(self.user_num):\n","                yty = np.dot(self.Y.T, self.Y)\n","                self.X[u, :] = self.update_user_vector(u, yty)\n","\n","            for i in range(self.item_num):\n","                xtx = np.dot(self.X.T, self.X)\n","                self.Y[i, :] = self.update_item_vector(i, xtx)\n","            \n","            phat = self.scoring()\n","            \n","            train_loss = self.evaluate(train_eval=True)\n","            test_loss = self.evaluate(train_eval=False)\n","\n","            self.loss_tr[epoch] = train_loss\n","            self.loss_tst[epoch] = test_loss\n","\n","\n","            print(f'EPOCH {epoch+1} : TRAINING RANK {self.loss_tr[epoch]:.5f}, VALID RANK {self.loss_tst[epoch]:.5f}')\n","\n","            print(f'Time per one epoch {datetime.now() - start_epoch}')\n","        end = datetime.now()\n","        print(f'Training takes time {end-start}')\n","        \n","    def scoring(self):\n","        \n","        return np.dot(self.X, self.Y.T)\n","    \n","    def update_user_vector(self, u, yty):\n","        np.fill_diagonal(self.C_u, (self.C[u, :] - 1))\n","        comp1 = yty\n","        comp2 = np.dot(self.Y.T, self.C_u).dot(self.Y)\n","        comp3 = np.identity(self.config.dim_f) * self.config.lambda_u\n","        comp = np.linalg.inv(comp1 + comp2 + comp3)\n","        self.C_u = self.C_u + np.identity(self.C_u.shape[0])\n","        comp = np.dot(comp, self.Y.T).dot(self.C_u)\n","        \n","        return np.dot(comp, self.R_tr[u, :])\n","\n","    def update_item_vector(self, i, xtx): \n","        np.fill_diagonal(self.C_i, (self.C[:, i] - 1))\n","        comp1 = xtx\n","        comp2 = np.dot(self.X.T, self.C_i).dot(self.X)\n","        comp3 = np.identity(self.config.dim_f) * self.config.lambda_v\n","        comp = np.linalg.inv(comp1 + comp2 + comp3)\n","        self.C_i = self.C_i + np.identity(self.C_i.shape[0])\n","        comp4 = self.X.T.dot(self.C_i).dot(self.R_tr[:, i])\n","        comp5 = self.lambda_v * self.latent_feat[i, :]\n","        \n","        return np.dot(comp, comp4+comp5)\n","    \n","    def loss_fn(self, pred, xc, idx, latent_feat):\n","        X = torch.tensor(self.X).to(self.device)\n","        Y = torch.tensor(self.Y).to(self.device)[idx, :]\n","        R = torch.tensor(self.R_tr).float().to(self.device)[:, idx]\n","        C = torch.tensor(self.C).float().to(self.device)[:, idx]\n","        latent = torch.tensor(latent_feat).to(self.device)\n","\n","        comp1 = (X**2).sum(axis=1).sum() * self.lambda_u/2\n","        comp2 = ((Y - latent)**2).sum(axis=1).sum() * self.lambda_v/2\n","        comp3 = ((pred - xc)**2).sum(axis=1).sum() * self.lambda_n/2\n","        comp4 = torch.sum((torch.mm(X, Y.T) - R)**2 * C/2)\n","\n","        return comp1+comp2+comp3+comp4\n","\n","    def evaluate(self, train_eval):\n","        if train_eval:\n","            R = self.R_tr\n","        else:\n","            R = self.R_tst\n","\n","        phat = self.scoring()\n","        rank_mat = np.zeros(phat.shape)\n","        for u in range(self.user_num):\n","            pred_u = phat[u, :] * -1\n","            rank = pred_u.argsort().argsort()\n","            rank = rank / self.item_num\n","            rank_mat[u, :] = rank\n","\n","        return np.sum(R * rank_mat) / np.sum(R)"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"xHtxulpK8yts"},"source":["class Config:\n","    learning_rate = 0.001\n","    early_stopping_round = 0\n","    epochs = 15\n","    seed = 1995\n","    dim_f = 10\n","    batch_size = 16\n","    lambda_u = 1\n","    lambda_w = 5e-4\n","    lambda_v = 1\n","    lambda_n = 1\n","    device = 'cuda' if torch.cuda.is_available() else 'cpu'\n","    rankM = 50\n","\n","config = Config()"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"oGgWfwh7T8oe"},"source":["train = np.load('ml_100k_train.npy')\n","test = np.load('ml_100k_test.npy')\n","xc = pd.read_csv('movies.csv').iloc[:,5:].values\n","\n","train = np.where(train > 0, 1, 0)\n","test = np.where(test > 0, 1, 0)\n","x0 = add_noise(xc, corrupt_ratio=0.1)"],"execution_count":null,"outputs":[]},{"cell_type":"code","metadata":{"id":"MWpUNtyJ5eiV","colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"status":"ok","timestamp":1630644729721,"user_tz":-330,"elapsed":724462,"user":{"displayName":"Sparsh Agarwal","photoUrl":"","userId":"13037694610922482904"}},"outputId":"c04371e6-3de2-4704-e8b8-776f6a1bcea2"},"source":["config.n_item = train.shape[1]\n","config.n_user = train.shape[0]\n","idx = np.arange(config.n_item)\n","\n","config.neg_item_tr = {i :idx[~train[i, :].astype(bool)] for i in range(config.n_user)}\n","config.neg_item_tst = {i :idx[~test[i, :].astype(bool)] for i in range(config.n_user)}\n","\n","config.pos_item_tr_bool = {i :train[i, :].astype(bool) for i in range(config.n_user)}\n","config.pos_item_tst_bool = {i :test[i, :].astype(bool) for i in range(config.n_user)}\n","\n","dataset = CDLDataset(xc, x0)\n","trainloader = DataLoader(dataset, config.batch_size, drop_last=False, shuffle=False)\n","\n","model = CDL(\n","    train_imp=train, \n","    test_imp=test, \n","    input_dim=xc.shape[1], \n","    hidden_dim=config.dim_f, \n","    dim_f=config.dim_f,\n","    dataloader=trainloader,\n","    seed=1995,\n","    device=config.device,\n","    config=config\n",")\n","\n","model.fit()"],"execution_count":null,"outputs":[{"output_type":"stream","name":"stdout","text":["EPOCH 1 : TRAINING RANK 0.20875, VALID RANK 0.20875\n","Time per one epoch 0:00:50.467119\n","EPOCH 2 : TRAINING RANK 0.14438, VALID RANK 0.14438\n","Time per one epoch 0:00:48.940005\n","EPOCH 3 : TRAINING RANK 0.16976, VALID RANK 0.16976\n","Time per one epoch 0:00:47.577796\n","EPOCH 4 : TRAINING RANK 0.18687, VALID RANK 0.18687\n","Time per one epoch 0:00:48.211371\n","EPOCH 5 : TRAINING RANK 0.20023, VALID RANK 0.20023\n","Time per one epoch 0:00:49.610613\n","EPOCH 6 : TRAINING RANK 0.20773, VALID RANK 0.20773\n","Time per one epoch 0:00:48.596721\n","EPOCH 7 : TRAINING RANK 0.21310, VALID RANK 0.21310\n","Time per one epoch 0:00:49.022005\n","EPOCH 8 : TRAINING RANK 0.21953, VALID RANK 0.21953\n","Time per one epoch 0:00:48.815897\n","EPOCH 9 : TRAINING RANK 0.23348, VALID RANK 0.23348\n","Time per one epoch 0:00:48.488231\n","EPOCH 10 : TRAINING RANK 0.36056, VALID RANK 0.36056\n","Time per one epoch 0:00:48.563701\n","EPOCH 11 : TRAINING RANK 0.54516, VALID RANK 0.54516\n","Time per one epoch 0:00:48.318132\n","EPOCH 12 : TRAINING RANK 0.39444, VALID RANK 0.39444\n","Time per one epoch 0:00:46.785170\n","EPOCH 13 : TRAINING RANK 0.74885, VALID RANK 0.74885\n","Time per one epoch 0:00:46.347805\n","EPOCH 14 : TRAINING RANK 0.79147, VALID RANK 0.79147\n","Time per one epoch 0:00:46.180976\n","EPOCH 15 : TRAINING RANK 0.80281, VALID RANK 0.80281\n","Time per one epoch 0:00:47.471120\n","Training takes time 0:12:03.398521\n"]}]},{"cell_type":"markdown","metadata":{"id":"vhB5APuTQNQI"},"source":["---"]},{"cell_type":"code","metadata":{"colab":{"base_uri":"https://localhost:8080/"},"id":"jpKEP2FGQSKE","executionInfo":{"status":"ok","timestamp":1638115654441,"user_tz":-330,"elapsed":3216,"user":{"displayName":"Sparsh Agarwal","photoUrl":"https://lh3.googleusercontent.com/a/default-user=s64","userId":"13037694610922482904"}},"outputId":"c36674e7-df1a-4143-f355-9786cdfdf417"},"source":["!pip install -q watermark\n","%reload_ext watermark\n","%watermark -a \"Sparsh A.\" -m -iv -u -t -d"],"execution_count":null,"outputs":[{"output_type":"stream","name":"stdout","text":["Author: Sparsh A.\n","\n","Last updated: 2021-11-28 16:07:35\n","\n","Compiler    : GCC 7.5.0\n","OS          : Linux\n","Release     : 5.4.104+\n","Machine     : x86_64\n","Processor   : x86_64\n","CPU cores   : 2\n","Architecture: 64bit\n","\n","torch     : 1.10.0+cu111\n","IPython   : 5.5.0\n","sys       : 3.7.12 (default, Sep 10 2021, 00:21:48) \n","[GCC 7.5.0]\n","numpy     : 1.19.5\n","pandas    : 1.1.5\n","matplotlib: 3.2.2\n","\n"]}]},{"cell_type":"markdown","metadata":{"id":"-k-HLCIjQNQR"},"source":["---"]},{"cell_type":"markdown","metadata":{"id":"Mxkw17lSQNQS"},"source":["**END**"]}]}