推荐系统作业

gmf.py

+import numpy as np
+import pandas as pd
+import os
+import torch
+import argparse
+import heapq
+import pdb
+
+from tqdm import tqdm,trange
+from time import time
+from scipy.sparse import load_npz
+from torch import nn
+from torch.optim.lr_scheduler import CyclicLR
+from torch.utils.data import DataLoader, Dataset
+from utils import get_train_instances, get_scores
+
+os.environ["CUDA_VISIBLE_DEVICES"] = '0'  # assign GPU
+
+def parse_args():
+
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--datadir", type=str, default=".",
+        help="data directory.")
+    parser.add_argument("--modeldir", type=str, default="./models",
+        help="models directory")
+    parser.add_argument("--dataname", type=str, default="neuralcf_split.npz",
+        help="npz file with dataset")
+    parser.add_argument("--train_matrix", type=str, default="neuralcf_train_sparse.npz",
+        help="train matrix for faster iteration")
+    parser.add_argument("--epochs", type=int, default=20,
+        help="number of epochs.")
+    parser.add_argument("--batch_size", type=int, default=1024,
+        help="batch size.")
+    parser.add_argument("--n_emb", type=int, default=8,
+        help="embedding size.")
+    parser.add_argument("--lr", type=float, default=0.01,
+        help="if lr_scheduler this will be max_lr")
+    parser.add_argument("--learner", type=str, default="adam",
+        help="Specify an optimizer: adagrad, adam, rmsprop, sgd")
+    parser.add_argument("--lr_scheduler", action="store_true",
+        help="boolean to set the use of CyclicLR during training")
+    parser.add_argument("--validate_every", type=int, default=1,
+        help="validate every n epochs")
+    parser.add_argument("--save_model", type=int, default=1)
+    parser.add_argument("--n_neg", type=int, default=4,
+        help="number of negative instances to consider per positive instance")
+    parser.add_argument("--topk", type=int, default=10,
+        help="number of items to retrieve for recommendation")
+
+    return parser.parse_args()
+
+
+class GMF(nn.Module):
+    def __init__(self, n_user, n_item, n_emb=8):
+        super(GMF, self).__init__()
+
+        self.n_emb = n_emb
+        self.n_user = n_user
+        self.n_item = n_item
+
+        self.embeddings_user = nn.Embedding(n_user, n_emb)
+        self.embeddings_item = nn.Embedding(n_item, n_emb)
+        self.out = nn.Linear(in_features=n_emb, out_features=1)
+
+        for m in self.modules():
+            if isinstance(m, nn.Embedding):
+                nn.init.normal_(m.weight)
+            elif isinstance(m, nn.Linear):
+                nn.init.uniform_(m.weight)
+
+    def forward(self, users, items):
+
+        user_emb = self.embeddings_user(users)
+        item_emb = self.embeddings_item(items)
+
+        # Task-1: complete the proces to compute preds
+        # preds = '...'
+        preds = torch.sigmoid(self.out(torch.mul(user_emb, item_emb)))
+        return preds
+
+
+def train(model, criterion, optimizer, scheduler, epoch, batch_size,
+    use_cuda, train_ratings, negatives, n_items, n_neg):
+    model.train()
+    train_dataset = get_train_instances(train_ratings,
+        negatives,
+        n_items,
+        n_neg)
+    train_loader = DataLoader(dataset=train_dataset,
+        batch_size=batch_size,
+        num_workers=4,
+        shuffle=True)
+    train_steps = (len(train_loader.dataset) // train_loader.batch_size) + 1
+    running_loss=0
+    for data in train_loader:
+        users = data[:,0]
+        items = data[:,1]
+        labels = data[:,2].float()
+        if use_cuda:
+            users, items, labels = users.cuda(), items.cuda(), labels.cuda()
+        optimizer.zero_grad()
+        preds =  model(users, items)
+        if scheduler:
+            scheduler.step()
+        loss = criterion(preds.squeeze(1), labels)
+        loss.backward()
+        optimizer.step()
+        running_loss += loss.item()
+    return running_loss/train_steps
+
+
+def evaluate(model, test_loader, use_cuda, topk):
+    model.eval()
+    scores=[]
+    with torch.no_grad():
+        for data in test_loader:
+            users = data[:,0]
+            items = data[:,1]
+            labels = data[:,2].float()
+            if use_cuda:
+                users, items, labels = users.cuda(), items.cuda(), labels.cuda()
+            preds = model(users, items)
+            items_cpu = items.cpu().numpy()
+            preds_cpu = preds.squeeze(1).detach().cpu().numpy()
+            split_chuncks = preds_cpu.shape[0]//100
+            litems=np.split(items_cpu, split_chuncks)
+            lpreds=np.split(preds_cpu, split_chuncks)
+            scores += [get_scores(it,pr,topk) for it,pr in zip(litems,lpreds)]
+    hits = [s[0] for s in scores]
+    ndcgs = [s[1] for s in scores]
+    return (np.array(hits).mean(),np.array(ndcgs).mean())
+
+
+def checkpoint(model, modelpath):
+    torch.save(model.state_dict(), modelpath)
+
+
+if __name__ == '__main__':
+
+    args = parse_args()
+
+    datadir = args.datadir
+    dataname = args.dataname
+    train_matrix = args.train_matrix
+    modeldir = args.modeldir
+    n_emb = args.n_emb
+    batch_size = args.batch_size
+    epochs = args.epochs
+    learner = args.learner
+    lr = args.lr
+    lr_scheduler = args.lr_scheduler
+    lrs = "wlrs" if lr_scheduler else "wolrs"
+    validate_every = args.validate_every
+    save_model = args.save_model
+    topk = args.topk
+    n_neg = args.n_neg
+
+    modelfname = "GMF" + \
+        "_".join(["_bs", str(batch_size)]) + \
+        "_".join(["_lr", str(lr).replace(".", "")]) + \
+        "_".join(["_n_emb", str(n_emb)]) + \
+        "_".join(["_lrnr", learner]) + \
+        "_".join(["_lrs", lrs]) + \
+        ".pt"
+    if not os.path.exists(modeldir): os.makedirs(modeldir)
+    modelpath = os.path.join(modeldir, modelfname)
+    resultsdfpath = os.path.join(modeldir, 'results_df.p')
+
+    dataset = np.load(os.path.join(datadir, dataname))
+    train_ratings = load_npz(os.path.join(datadir, train_matrix)).todok()
+    test_ratings, negatives = dataset['test_negative'], dataset['negatives']
+    n_users, n_items = dataset['n_users'].item(), dataset['n_items'].item()
+
+    test_loader = DataLoader(dataset=test_ratings,
+        batch_size=1000,
+        shuffle=False
+        )
+
+    model = GMF(n_users, n_items, n_emb=n_emb)
+
+    if learner.lower() == "adagrad":
+        optimizer = torch.optim.Adagrad(model.parameters(), lr=lr)
+    elif learner.lower() == "rmsprop":
+        optimizer = torch.optim.RMSprop(model.parameters(), lr=lr, momentum=0.9)
+    elif learner.lower() == "adam":
+        optimizer = torch.optim.Adam(model.parameters(), lr=lr)
+    else:
+        optimizer = torch.optim.SGD(model.parameters(), lr=lr, momentum=0.9, nesterov=True)
+
+    criterion = nn.BCELoss()
+
+    training_steps = ((len(train_ratings)+len(train_ratings)*n_neg)//batch_size)+1
+    step_size = training_steps*3 # one cycle every 6 epochs
+    cycle_momentum=True
+    if learner.lower() == "adagrad" or learner.lower()=="adam":
+        cycle_momentum=False
+    if lr_scheduler:
+        scheduler = CyclicLR(optimizer, step_size_up=step_size, base_lr=lr/10., max_lr=lr,
+            cycle_momentum=cycle_momentum)
+    else:
+        scheduler = None
+
+    use_cuda = torch.cuda.is_available()
+    if use_cuda:
+        model = model.cuda()
+
+    best_hr, best_ndcgm, best_iter=0,0,0
+    for epoch in range(1,epochs+1):
+        t1 = time()
+        loss = train(model, criterion, optimizer, scheduler, epoch, batch_size,
+            use_cuda, train_ratings, negatives, n_items, n_neg)
+        t2 = time()
+        if epoch % validate_every == 0:
+            (hr, ndcg) = evaluate(model, test_loader, use_cuda, topk)
+            print("Epoch: {} {:.2f}s, LOSS = {:.4f}, HR = {:.4f}, NDCG = {:.4f}, validated in {:.2f}s".
+                format(epoch, t2-t1, loss, hr, ndcg, time()-t2))
+            if hr > best_hr:
+                iter_loss, best_hr, best_ndcg, best_iter, train_time = \
+                    loss, hr, ndcg, epoch, t2-t1
+                if save_model:
+                    checkpoint(model, modelpath)
+
+    print("End. Best Iteration {}: HR = {:.4f}, NDCG = {:.4f}. ".format(best_iter, best_hr, best_ndcg))
+    if save_model:
+        print("The best GMF model is saved to {}".format(modelpath))
+
+    if save_model:
+        cols = ["modelname", "iter_loss","best_hr", "best_ndcg", "best_iter","train_time"]
+        vals = [modelfname, iter_loss, best_hr, best_ndcg, best_iter, train_time]
+        if not os.path.isfile(resultsdfpath):
+            results_df = pd.DataFrame(columns=cols)
+            experiment_df = pd.DataFrame(data=[vals], columns=cols)
+            results_df = results_df.append(experiment_df, ignore_index=True)
+            results_df.to_pickle(resultsdfpath)
+        else:
+            results_df = pd.read_pickle(resultsdfpath)
+            experiment_df = pd.DataFrame(data=[vals], columns=cols)
+            results_df = results_df.append(experiment_df, ignore_index=True)
+            results_df.to_pickle(resultsdfpath)

mlp.py

+import numpy as np
+import pandas as pd
+import os
+import torch
+import argparse
+import heapq
+import pdb
+
+from tqdm import tqdm,trange
+from time import time
+from scipy.sparse import load_npz
+from torch import nn
+from torch.optim.lr_scheduler import CyclicLR
+from torch.utils.data import DataLoader, Dataset
+from utils import get_train_instances, get_scores
+from gmf import train, evaluate, checkpoint
+
+os.environ["CUDA_VISIBLE_DEVICES"] = '0'  # assign GPU
+
+def parse_args():
+
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--datadir", type=str, default=".",
+        help="data directory.")
+    parser.add_argument("--modeldir", type=str, default="./models",
+        help="models directory")
+    parser.add_argument("--dataname", type=str, default="neuralcf_split.npz",
+        help="npz file with dataset")
+    parser.add_argument("--train_matrix", type=str, default="neuralcf_train_sparse.npz",
+        help="train matrix for faster iteration")
+    parser.add_argument("--epochs", type=int, default=20,
+        help="number of epochs.")
+    parser.add_argument("--batch_size", type=int, default=1024,
+        help="batch size.")
+    parser.add_argument("--layers", type=str, default="[64,32,16,8]",
+        help="layer architecture. The first elements is used for the embedding \
+        layers and equals n_emb*2")
+    parser.add_argument("--dropouts", type=str, default="[0,0,0]",
+        help="dropout per dense layer. len(dropouts) = len(layers)-1")
+    parser.add_argument("--l2reg", type=float, default=0.,
+        help="l2 regularization")
+    parser.add_argument("--lr", type=float, default=0.01,
+        help="if lr_scheduler this will be max_lr")
+    parser.add_argument("--learner", type=str, default="adam",
+        help="Specify an optimizer: adagrad, adam, rmsprop, sgd")
+    parser.add_argument("--lr_scheduler", action="store_true",
+        help="use CyclicLR during training")
+    parser.add_argument("--validate_every", type=int, default=1,
+        help="validate every n epochs")
+    parser.add_argument("--save_model", type=int, default=1)
+    parser.add_argument("--n_neg", type=int, default=4,
+        help="number of negative instances to consider per positive instance.")
+    parser.add_argument("--topk", type=int, default=10,
+        help="number of items to retrieve for recommendation.")
+
+    return parser.parse_args()
+
+
+class MLP(nn.Module):
+    """
+    Concatenate Embeddings that are then passed through a series of Dense  layers
+    """
+    def __init__(self, n_user, n_item, layers, dropouts):
+        super(MLP, self).__init__()
+
+        self.layers = layers
+        self.n_layers = len(layers)
+        self.dropouts = dropouts
+        self.n_user = n_user
+        self.n_item = n_item
+
+        self.embeddings_user = nn.Embedding(n_user, int(layers[0]/2))
+        self.embeddings_item = nn.Embedding(n_item, int(layers[0]/2))
+
+        self.mlp = nn.Sequential()
+        for i in range(1,self.n_layers):
+            self.mlp.add_module("linear%d" %i, nn.Linear(layers[i-1],layers[i]))
+            self.mlp.add_module("relu%d" %i, torch.nn.ReLU())
+            self.mlp.add_module("dropout%d" %i , torch.nn.Dropout(p=dropouts[i-1]))
+
+        # Task-2: replace '?' with two proper numbers in below code
+        # self.out = nn.Linear(in_features='?', out_features='?')
+        self.out = nn.Linear(in_features= layers[-1], out_features= 1)
+
+        for m in self.modules():
+            if isinstance(m, nn.Embedding):
+                nn.init.normal_(m.weight)
+
+    def forward(self, users, items):
+
+        user_emb = self.embeddings_user(users)
+        item_emb = self.embeddings_item(items)
+        emb_vector = torch.cat([user_emb,item_emb], dim=1)
+        emb_vector = self.mlp(emb_vector)
+        preds = torch.sigmoid(self.out(emb_vector))
+
+        return preds
+
+
+if __name__ == '__main__':
+
+    args = parse_args()
+
+    datadir = args.datadir
+    dataname = args.dataname
+    train_matrix = args.train_matrix
+    modeldir = args.modeldir
+    layers = eval(args.layers)
+    ll = str(layers[-1]) #last layer
+    dropouts = eval(args.dropouts)
+    dp = "wdp" if dropouts[0]!=0 else "wodp"
+    l2reg = args.l2reg
+    n_emb = int(layers[0]/2)
+    batch_size = args.batch_size
+    epochs = args.epochs
+    learner = args.learner
+    lr = args.lr
+    lr_scheduler = args.lr_scheduler
+    lrs = "wlrs" if lr_scheduler else "wolrs"
+    validate_every = args.validate_every
+    save_model = args.save_model
+    topk = args.topk
+    n_neg = args.n_neg
+
+    modelfname = "MLP" + \
+        "_".join(["_bs", str(batch_size)]) + \
+        "_".join(["_reg", str(l2reg).replace(".", "")]) + \
+        "_".join(["_lr", str(lr).replace(".", "")]) + \
+        "_".join(["_n_emb", str(n_emb)]) + \
+        "_".join(["_ll", ll]) + \
+        "_".join(["_dp", dp]) + \
+        "_".join(["_lrnr", learner]) + \
+        "_".join(["_lrs", lrs]) + \
+        ".pt"
+    modelpath = os.path.join(modeldir, modelfname)
+    resultsdfpath = os.path.join(modeldir, 'results_df.p')
+
+    dataset = np.load(os.path.join(datadir, dataname))
+    train_ratings = load_npz(os.path.join(datadir, train_matrix)).todok()
+    test_ratings, negatives = dataset['test_negative'], dataset['negatives']
+    n_users, n_items = dataset['n_users'].item(), dataset['n_items'].item()
+
+    test_loader = DataLoader(dataset=test_ratings,
+        batch_size=1000,
+        shuffle=False
+        )
+
+    model = MLP(n_users, n_items, layers, dropouts)
+
+    if learner.lower() == "adagrad":
+        optimizer = torch.optim.Adagrad(model.parameters(), lr=lr, weight_decay=l2reg)
+    elif learner.lower() == "rmsprop":
+        optimizer = torch.optim.RMSprop(model.parameters(), lr=lr, weight_decay=l2reg,
+            momentum=0.9)
+    elif learner.lower() == "adam":
+        optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=l2reg)
+    else:
+        optimizer = torch.optim.SGD(model.parameters(), lr=lr, weight_decay=l2reg,
+            momentum=0.9, nesterov=True)
+
+    criterion = nn.BCELoss()
+
+    training_steps = ((len(train_ratings)+len(train_ratings)*n_neg)//batch_size)+1
+    step_size = training_steps*3 #one cycle every 6 epochs
+    cycle_momentum=True
+    if learner.lower() == "adagrad" or learner.lower()=="adam":
+        cycle_momentum=False
+    if lr_scheduler:
+        scheduler = CyclicLR(optimizer, step_size_up=step_size, base_lr=lr/10., max_lr=lr,
+            cycle_momentum=cycle_momentum)
+    else:
+        scheduler = None
+
+    use_cuda = torch.cuda.is_available()
+    if use_cuda:
+        model = model.cuda()
+
+    best_hr, best_ndcgm, best_iter=0,0,0
+    for epoch in range(1,epochs+1):
+        t1 = time()
+        loss = train(model, criterion, optimizer, scheduler, epoch, batch_size,
+            use_cuda, train_ratings, negatives, n_items, n_neg)
+        t2 = time()
+        if epoch % validate_every == 0:
+            (hr, ndcg) = evaluate(model, test_loader, use_cuda, topk)
+            print("Epoch: {} {:.2f}s, LOSS = {:.4f}, HR = {:.4f}, NDCG = {:.4f}, validated in {:.2f}s".
+                format(epoch, t2-t1, loss, hr, ndcg, time()-t2))
+            if hr > best_hr:
+                iter_loss, best_hr, best_ndcg, best_iter, train_time = \
+                    loss, hr, ndcg, epoch, t2-t1
+                if save_model:
+                    checkpoint(model, modelpath)
+
+    print("End. Best Iteration {}: HR = {:.4f}, NDCG = {:.4f}. ".format(best_iter, best_hr, best_ndcg))
+    if save_model:
+        print("The best MLP model is saved to {}".format(modelpath))
+
+    if save_model:
+        cols = ["modelname", "iter_loss","best_hr", "best_ndcg", "best_iter","train_time"]
+        vals = [modelfname, iter_loss, best_hr, best_ndcg, best_iter, train_time]
+        if not os.path.isfile(resultsdfpath):
+            results_df = pd.DataFrame(columns=cols)
+            experiment_df = pd.DataFrame(data=[vals], columns=cols)
+            results_df = results_df.append(experiment_df, ignore_index=True)
+            results_df.to_pickle(resultsdfpath)
+        else:
+            results_df = pd.read_pickle(resultsdfpath)
+            experiment_df = pd.DataFrame(data=[vals], columns=cols)
+            results_df = results_df.append(experiment_df, ignore_index=True)
+            results_df.to_pickle(resultsdfpath)

prepare_data.py

+import numpy as np
+import pandas as pd
+import gzip
+import pickle
+import argparse
+import scipy.sparse as sp
+
+from time import time
+from pathlib import Path
+from scipy.sparse import save_npz
+from joblib import Parallel, delayed
+from sklearn.model_selection import train_test_split
+
+
+def array2mtx(interactions):
+    num_users = interactions[:,0].max()
+    num_items = interactions[:,1].max()
+    mat = sp.dok_matrix((num_users+1, num_items+1), dtype=np.float32)
+    for user, item, rating in interactions:
+            mat[user, item] = rating
+    return mat.tocsr()
+
+
+def standard_split(df, data_path):
+
+	# Cardinality
+	n_users = df.user.nunique()
+	n_items = df.item.nunique()
+	n_ranks = df['rank'].nunique()
+	train, test, = train_test_split(df.values.astype(np.int64), test_size=0.2, random_state=1)
+
+	# Save
+	np.savez(data_path/"standard_split.npz", train=train, test=test, n_users=n_users,
+		n_items=n_items, n_ranks=n_ranks, columns=df.columns.tolist())
+
+
+def neuralcf_split(df, data_path):
+	# Xiangnan He, et al, 2017 train/test split with implicit negative feedback
+
+	# sort by rank
+	dfc = df.copy()
+
+	# Cardinality
+	n_users = df.user.nunique()
+	n_items = df.item.nunique()
+
+	dfc.sort_values(['user','rank'], ascending=[True,True], inplace=True)
+	dfc.reset_index(inplace=True, drop=True)
+
+	# use last ratings for testing and all the previous for training
+	test = dfc.groupby('user').tail(1)
+	train = pd.merge(dfc, test, on=['user','item'],
+		how='outer', suffixes=('', '_y'))
+	train = train[train.rating_y.isnull()]
+	test = test[['user','item','rating']]
+	train = train[['user','item','rating']]
+
+	# select 99 random movies per user that were never rated by that user
+	all_items = dfc.item.unique()
+	rated_items = (dfc.groupby("user")['item']
+	    .apply(list)
+	    .reset_index()
+	    ).item.tolist()
+
+	def sample_not_rated(item_list, rseed=1, n=99):
+		np.random.seed=rseed
+		return np.random.choice(np.setdiff1d(all_items, item_list), n)
+
+	print("sampling not rated items...")
+	start = time()
+	non_rated_items = Parallel(n_jobs=4)(delayed(sample_not_rated)(ri) for ri in rated_items)
+	end = time() - start
+	print("sampling took {} min".format(round(end/60,2)))
+
+	negative = pd.DataFrame({'negative':non_rated_items})
+	negative[['item_n'+str(i) for i in range(99)]] =\
+		pd.DataFrame(negative.negative.values.tolist(), index= negative.index)
+	negative.drop('negative', axis=1, inplace=True)
+	negative = negative.stack().reset_index()
+	negative = negative.iloc[:, [0,2]]
+	negative.columns = ['user','item']
+	negative['rating'] = 0
+	assert negative.shape[0] == len(non_rated_items)*99
+	test_negative = (pd.concat([test,negative])
+		.sort_values('user', ascending=True)
+		.reset_index(drop=True)
+		)
+	# Ensuring that the 1st element every 100 is the rated item. This is
+	# fundamental for testing
+	test_negative.sort_values(['user', 'rating'], ascending=[True,False], inplace=True)
+	assert np.all(test_negative.values[0::100][:,2] != 0)
+
+	# Save
+	np.savez(data_path/"neuralcf_split.npz", train=train.values, test=test.values,
+		test_negative=test_negative.values, negatives=np.array(non_rated_items),
+		n_users=n_users, n_items=n_items)
+
+	# Save training as sparse matrix
+	print("saving training set as sparse matrix...")
+	train_mtx = array2mtx(train.values)
+	save_npz(data_path/"neuralcf_train_sparse.npz", train_mtx)
+
+
+def prepare_amazon(data_path, input_fname):
+
+	df = pd.read_json(data_path/input_fname, lines=True)
+
+	keep_cols = ['reviewerID', 'asin', 'unixReviewTime', 'overall']
+	new_colnames = ['user', 'item', 'timestamp', 'rating']
+	df = df[keep_cols]
+	df.columns = new_colnames
+
+	# rank of items bought
+	df['rank'] = df.groupby("user")["timestamp"].rank(ascending=True, method='dense')
+	df.drop("timestamp", axis=1, inplace=True)
+
+	# mapping user and item ids to integers
+	user_mappings = {k:v for v,k in enumerate(df.user.unique())}
+	item_mappings = {k:v for v,k in enumerate(df.item.unique())}
+	df['user'] = df['user'].map(user_mappings)
+	df['item'] = df['item'].map(item_mappings)
+	df = df[['user','item','rank','rating']].astype(np.int64)
+
+	pickle.dump(user_mappings, open(data_path/'user_mappings.p', 'wb'))
+	pickle.dump(item_mappings, open(data_path/'item_mappings.p', 'wb'))
+
+	standard_split(df, data_path)
+	neuralcf_split(df, data_path)
+
+
+if __name__ == '__main__':
+
+	parser = argparse.ArgumentParser(description="prepare Amazon dataset")
+
+	parser.add_argument("--input_dir",type=str, default=".")
+	parser.add_argument("--input_fname",type=str, default="reviews_Movies_and_TV_5.json.gz")
+	args = parser.parse_args()
+
+	DATA_PATH = Path(args.input_dir)
+	reviews = args.input_fname
+
+	prepare_amazon(DATA_PATH, reviews)
+