Full Code of yingtaoluo/Spatial-Temporal-Attention-Network-for-POI-Recommendation for AI

Repository: yingtaoluo/Spatial-Temporal-Attention-Network-for-POI-Recommendation
Branch: master
Commit: 8b82f2840e7f
Files: 8
Total size: 27.3 KB

Directory structure:
gitextract_gnhvk2ih/

├── README.md
├── STAN_ORAL.pptx
├── data/
│   ├── NYC.npy
│   └── NYC_POI.npy
├── layers.py
├── load.py
├── models.py
└── train.py

================================================
FILE CONTENTS
================================================

================================================
FILE: README.md
================================================
# STAN: Spatial-Temporal-Attention-Network-for-Next-Location-Recommendation
[[Paper](https://arxiv.org/abs/2102.04095)]. [[Oral Youtube](https://www.youtube.com/watch?v=ajNzESvOvzs)] or [[Oral Bilibili](https://www.bilibili.com/video/BV1WL411E7Qm?from=search&seid=7472683569881802215)]. [[Implementation through LibCity](https://github.com/LibCity/Bigscity-LibCity)].

Thank you for your interest in our work! Thank you for reporting possible bugs and please make sure you are forking the latest repo to avoid eariler bugs. Before asking questions regarding the codes or the paper, I strongly recommend you to read the FAQ first. You can also use the [LibCity](https://github.com/LibCity/Bigscity-LibCity) version.

## Description
Because of the huge memory of the location matrix, the running speed of STAN is extremely low. You can refer to the implementation of masked attention [[here](https://github.com/yingtaoluo/PyHealth/blob/master/pyhealth/models/sequence/dipole.py)] if you wish to rewrite your own codes. 

Divide the dataset into different proportions of users to test the performance and then average. 

Run "load.py" first and then "train.py". You should see on the screen the result of the first proportion:   
100%|██████████| 100/100 [14:32<00:00,  8.72s/it]  
epoch:27, time:23587.941201210022, valid_acc:[0.18 0.49 0.56 0.67]  
epoch:27, time:23587.941201210022, test_acc:[0.15 0.46 0.59 0.67]

## FAQs
Q1: Can you provide a dataset?  
A1: Our datasets are collected from the following links. Please feel free to do your own data processing on your model while comparing STAN as baseline.
http://snap.stanford.edu/data/loc-gowalla.html;  
https://personal.ntu.edu.sg/gaocong/data/poidata.zip;
http://www-public.imtbs-tsp.eu/~zhang_da/pub/dataset_tsmc2014.zip  
  
Q2.1: What does it mean "The number of the training set is 𝑚 − 3, with the first 𝑚′ ∈ [1,𝑚 − 3] check-ins as input sequence and the [2,𝑚 − 2]-nd visited location as the label"?  
A2.1: We use [1] as input to predict [2], use [1,2] as input to predict [3], and ..., until we use [1,...,m-3] to predict [m-2]. Basically we do not use the last few steps and reserve them as a simulation of "future visits" to test the model since these last steps are not fed into the model during training.  
  
Q2.2: Can you please explain your trajectory encoding process? Do you create the location embeddings using skip-gram-like approaches?  
A2.2: Pre-training of embedding is an effective approach and can further improve the performance for sure. Unfortunately, the focus and contribution of this paper are not on embedding pre-training but on spatio-temporal linear embedding, and pretraining is not used in baselines, so we do not use it in our paper.

Q2.3: Would it be better to construct edges based on spatial distances instead of using distances?  
A2.3: If the edges can truly reflect the relations between each loaction and each user, then yes. Ideal 0-1 edge relation is a stronger representation. However, constructing edges merely based on spatial distances can raise problems. Consider that a 30-kilometer metro takes less time than a 5-kilometer walk. From the data, we only know distances.  

Q2.4: What do you mean by setting a unit spatiotemporal embedding?  
A2.4: ![image](https://github.com/yingtaoluo/Spatial-Temporal-Attention-Network-for-POI-Recommendation/blob/master/unit_embedding.png)

Q2.5: What does each column/row in NYC.npy mean?  
A2.5: Each row: [user id, check-in location id, time in minutes].  

Q2.6: Can we try a different division of train/dev/test datasets?  
A2.6: Our goal here is to generalize for the future visits of each user we have known (we do not want to test the model performance on biased past behavior), instead of generalizing to other users whose user-id embeddings are not known to the model. 

Q2.7: How is the value of the recall rate calculated in your paper? For example, the top5 probability of the NYC data set is 0.xx but in the paper it is 0.xxxx.  
A2.7: It is common practice to run under different seeds and get the average value. We averaged the ten times results and all of them are accepted by the statistical test of p=0.01. 

Q3: What is the environment to run the code? And version?  
A3: We use python 3.7.2, CUDA 10.1 and PyTorch 1.7.1. Make sure to install all libs that we import.  


================================================
FILE: layers.py
================================================
from load import *
import torch
from torch import nn
from torch.nn import functional as F

seed = 0
global_seed = 0
hours = 24*7
torch.manual_seed(seed)
device = 'cuda'


def to_npy(x):
    return x.cpu().data.numpy() if device == 'cuda' else x.detach().numpy()


class Attn(nn.Module):
    def __init__(self, emb_loc, loc_max, dropout=0.1):
        super(Attn, self).__init__()
        self.value = nn.Linear(max_len, 1, bias=False)
        self.emb_loc = emb_loc
        self.loc_max = loc_max

    def forward(self, self_attn, self_delta, traj_len):
        # self_attn (N, M, emb), candidate (N, L, emb), self_delta (N, M, L, emb), len [N]
        self_delta = torch.sum(self_delta, -1).transpose(-1, -2)  # squeeze the embed dimension
        [N, L, M] = self_delta.shape
        candidates = torch.linspace(1, int(self.loc_max), int(self.loc_max)).long()  # (L)
        candidates = candidates.unsqueeze(0).expand(N, -1).to(device)  # (N, L)
        emb_candidates = self.emb_loc(candidates)  # (N, L, emb)
        attn = torch.mul(torch.bmm(emb_candidates, self_attn.transpose(-1, -2)), self_delta)  # (N, L, M)
        # pdb.set_trace()
        attn_out = self.value(attn).view(N, L)  # (N, L)
        # attn_out = F.log_softmax(attn_out, dim=-1)  # ignore if cross_entropy_loss

        return attn_out  # (N, L)


class SelfAttn(nn.Module):
    def __init__(self, emb_size, output_size, dropout=0.1):
        super(SelfAttn, self).__init__()
        self.query = nn.Linear(emb_size, output_size, bias=False)
        self.key = nn.Linear(emb_size, output_size, bias=False)
        self.value = nn.Linear(emb_size, output_size, bias=False)

    def forward(self, joint, delta, traj_len):
        delta = torch.sum(delta, -1)  # squeeze the embed dimension
        # joint (N, M, emb), delta (N, M, M, emb), len [N]
        # construct attention mask
        mask = torch.zeros_like(delta, dtype=torch.float32)
        for i in range(mask.shape[0]):
            mask[i, 0:traj_len[i], 0:traj_len[i]] = 1

        attn = torch.add(torch.bmm(self.query(joint), self.key(joint).transpose(-1, -2)), delta)  # (N, M, M)
        attn = F.softmax(attn, dim=-1) * mask  # (N, M, M)

        attn_out = torch.bmm(attn, self.value(joint))  # (N, M, emb)

        return attn_out  # (N, M, emb)


class Embed(nn.Module):
    def __init__(self, ex, emb_size, loc_max, embed_layers):
        super(Embed, self).__init__()
        _, _, _, self.emb_su, self.emb_sl, self.emb_tu, self.emb_tl = embed_layers
        self.su, self.sl, self.tu, self.tl = ex
        self.emb_size = emb_size
        self.loc_max = loc_max

    def forward(self, traj_loc, mat2, vec, traj_len):
        # traj_loc (N, M), mat2 (L, L), vec (N, M), delta_t (N, M, L)
        delta_t = vec.unsqueeze(-1).expand(-1, -1, self.loc_max)
        delta_s = torch.zeros_like(delta_t, dtype=torch.float32)
        mask = torch.zeros_like(delta_t, dtype=torch.long)
        for i in range(mask.shape[0]):  # N
            mask[i, 0:traj_len[i]] = 1
            delta_s[i, :traj_len[i]] = torch.index_select(mat2, 0, (traj_loc[i]-1)[:traj_len[i]])

        # pdb.set_trace()

        esl, esu, etl, etu = self.emb_sl(mask), self.emb_su(mask), self.emb_tl(mask), self.emb_tu(mask)
        vsl, vsu, vtl, vtu = (delta_s - self.sl).unsqueeze(-1).expand(-1, -1, -1, self.emb_size), \
                             (self.su - delta_s).unsqueeze(-1).expand(-1, -1, -1, self.emb_size), \
                             (delta_t - self.tl).unsqueeze(-1).expand(-1, -1, -1, self.emb_size), \
                             (self.tu - delta_t).unsqueeze(-1).expand(-1, -1, -1, self.emb_size)

        space_interval = (esl * vsu + esu * vsl) / (self.su - self.sl)
        time_interval = (etl * vtu + etu * vtl) / (self.tu - self.tl)
        delta = space_interval + time_interval  # (N, M, L, emb)

        return delta


class MultiEmbed(nn.Module):
    def __init__(self, ex, emb_size, embed_layers):
        super(MultiEmbed, self).__init__()
        self.emb_t, self.emb_l, self.emb_u, \
        self.emb_su, self.emb_sl, self.emb_tu, self.emb_tl = embed_layers
        self.su, self.sl, self.tu, self.tl = ex
        self.emb_size = emb_size

    def forward(self, traj, mat, traj_len):
        # traj (N, M, 3), mat (N, M, M, 2), len [N]
        traj[:, :, 2] = (traj[:, :, 2]-1) % hours + 1  # segment time by 24 hours * 7 days
        time = self.emb_t(traj[:, :, 2])  # (N, M) --> (N, M, embed)
        loc = self.emb_l(traj[:, :, 1])  # (N, M) --> (N, M, embed)
        user = self.emb_u(traj[:, :, 0])  # (N, M) --> (N, M, embed)
        joint = time + loc + user  # (N, M, embed)

        delta_s, delta_t = mat[:, :, :, 0], mat[:, :, :, 1]  # (N, M, M)
        mask = torch.zeros_like(delta_s, dtype=torch.long)
        for i in range(mask.shape[0]):
            mask[i, 0:traj_len[i], 0:traj_len[i]] = 1

        esl, esu, etl, etu = self.emb_sl(mask), self.emb_su(mask), self.emb_tl(mask), self.emb_tu(mask)
        vsl, vsu, vtl, vtu = (delta_s - self.sl).unsqueeze(-1).expand(-1, -1, -1, self.emb_size), \
                             (self.su - delta_s).unsqueeze(-1).expand(-1, -1, -1, self.emb_size), \
                             (delta_t - self.tl).unsqueeze(-1).expand(-1, -1, -1, self.emb_size), \
                             (self.tu - delta_t).unsqueeze(-1).expand(-1, -1, -1, self.emb_size)

        space_interval = (esl*vsu+esu*vsl) / (self.su-self.sl)
        time_interval = (etl*vtu+etu*vtl) / (self.tu-self.tl)
        delta = space_interval + time_interval  # (N, M, M, emb)

        return joint, delta


================================================
FILE: load.py
================================================
import numpy as np
import torch
from math import radians, cos, sin, asin, sqrt
import joblib
from torch.nn.utils.rnn import pad_sequence

max_len = 100  # max traj len; i.e., M


def haversine(lon1, lat1, lon2, lat2):
    """
    Calculate the great circle distance between two points
    on the earth (specified in decimal degrees)
    """
    lon1, lat1, lon2, lat2 = map(radians, [lon1, lat1, lon2, lat2])

    dlon = lon2 - lon1
    dlat = lat2 - lat1
    a = sin(dlat / 2) ** 2 + cos(lat1) * cos(lat2) * sin(dlon / 2) ** 2
    c = 2 * asin(sqrt(a))
    r = 6371
    return c * r


def euclidean(point, each):
    lon1, lat1, lon2, lat2 = point[2], point[1], each[2], each[1]
    return np.sqrt((lon1 - lon2)**2 + (lat1 - lat2)**2)


def rst_mat1(traj, poi):
    # traj (*M, [u, l, t]), poi(L, [l, lat, lon])
    mat = np.zeros((len(traj), len(traj), 2))
    for i, item in enumerate(traj):
        for j, term in enumerate(traj):
            poi_item, poi_term = poi[item[1] - 1], poi[term[1] - 1]  # retrieve poi by loc_id
            mat[i, j, 0] = haversine(lon1=poi_item[2], lat1=poi_item[1], lon2=poi_term[2], lat2=poi_term[1])
            mat[i, j, 1] = abs(item[2] - term[2])
    return mat  # (*M, *M, [dis, tim])


def rs_mat2s(poi, l_max):
    # poi(L, [l, lat, lon])
    candidate_loc = np.linspace(1, l_max, l_max)  # (L)
    mat = np.zeros((l_max, l_max))  # mat (L, L)
    for i, loc1 in enumerate(candidate_loc):
        print(i) if i % 100 == 0 else None
        for j, loc2 in enumerate(candidate_loc):
            poi1, poi2 = poi[int(loc1) - 1], poi[int(loc2) - 1]  # retrieve poi by loc_id
            mat[i, j] = haversine(lon1=poi1[2], lat1=poi1[1], lon2=poi2[2], lat2=poi2[1])
    return mat  # (L, L)


def rt_mat2t(traj_time):  # traj_time (*M+1) triangle matrix
    # construct a list of relative times w.r.t. causality
    mat = np.zeros((len(traj_time)-1, len(traj_time)-1))
    for i, item in enumerate(traj_time):  # label
        if i == 0:
            continue
        for j, term in enumerate(traj_time[:i]):  # data
            mat[i-1, j] = np.abs(item - term)
    return mat  # (*M, *M)


def process_traj(dname):  # start from 1
    # data (?, [u, l, t]), poi (L, [l, lat, lon])
    data = np.load('./data/' + dname + '.npy')
    # add the code below if you are using dividing time into minutes instead of hours
    data[:, -1] = np.array(data[:, -1]/60, dtype=np.int)
    poi = np.load('./data/' + dname + '_POI.npy')
    num_user = data[-1, 0]  # max id of users, i.e. NUM
    data_user = data[:, 0]  # user_id sequence in data
    trajs, labels, mat1, mat2t, lens = [], [], [], [], []
    u_max, l_max = np.max(data[:, 0]), np.max(data[:, 1])

    for u_id in range(num_user+1):
        if u_id == 0:  # skip u_id == 0
            continue
        init_mat1 = np.zeros((max_len, max_len, 2))  # first mat (M, M, 2)
        init_mat2t = np.zeros((max_len, max_len))  # second mat of time (M, M)
        user_traj = data[np.where(data_user == u_id)]  # find all check-ins of u_id
        user_traj = user_traj[np.argsort(user_traj[:, 2])].copy()  # sort traj by time

        print(u_id, len(user_traj)) if u_id % 100 == 0 else None

        if len(user_traj) > max_len + 1:  # consider only the M+1 recent check-ins
            # 0:-3 are training data, 1:-2 is training label;
            # 1:-2 are validation data, 2:-1 is validation label;
            # 2:-1 are test data, 3: is the label for test.
            # *M would be the real length if <= max_len + 1
            user_traj = user_traj[-max_len-1:]  # (*M+1, [u, l, t])

        # spatial and temporal intervals
        user_len = len(user_traj[:-1])  # the len of data, i.e. *M
        user_mat1 = rst_mat1(user_traj[:-1], poi)  # (*M, *M, [dis, tim])
        user_mat2t = rt_mat2t(user_traj[:, 2])  # (*M, *M)
        init_mat1[0:user_len, 0:user_len] = user_mat1
        init_mat2t[0:user_len, 0:user_len] = user_mat2t

        trajs.append(torch.LongTensor(user_traj)[:-1])  # (NUM, *M, [u, l, t])
        mat1.append(init_mat1)  # (NUM, M, M, 2)
        mat2t.append(init_mat2t)  # (NUM, M, M)
        labels.append(torch.LongTensor(user_traj[1:, 1]))  # (NUM, *M)
        lens.append(user_len-2)  # (NUM), the real *M for every user

    # padding zero to the vacancies in the right
    mat2s = rs_mat2s(poi, l_max)  # contains dis of all locations, (L, L)
    zipped = zip(*sorted(zip(trajs, mat1, mat2t, labels, lens), key=lambda x: len(x[0]), reverse=True))
    trajs, mat1, mat2t, labels, lens = zipped
    trajs, mat1, mat2t, labels, lens = list(trajs), list(mat1), list(mat2t), list(labels), list(lens)
    trajs = pad_sequence(trajs, batch_first=True, padding_value=0)  # (NUM, M, 3)
    labels = pad_sequence(labels, batch_first=True, padding_value=0)  # (NUM, M)

    data = [trajs, np.array(mat1), mat2s, np.array(mat2t), labels, np.array(lens), u_max, l_max]
    data_pkl = './data/' + dname + '_data.pkl'
    open(data_pkl, 'a')
    with open(data_pkl, 'wb') as pkl:
        joblib.dump(data, pkl)


if __name__ == '__main__':
    name = 'NYC'
    process_traj(name)


================================================
FILE: models.py
================================================
from layers import *


class Model(nn.Module):
    def __init__(self, t_dim, l_dim, u_dim, embed_dim, ex, dropout=0.1):
        super(Model, self).__init__()
        emb_t = nn.Embedding(t_dim, embed_dim, padding_idx=0)
        emb_l = nn.Embedding(l_dim, embed_dim, padding_idx=0)
        emb_u = nn.Embedding(u_dim, embed_dim, padding_idx=0)
        emb_su = nn.Embedding(2, embed_dim, padding_idx=0)
        emb_sl = nn.Embedding(2, embed_dim, padding_idx=0)
        emb_tu = nn.Embedding(2, embed_dim, padding_idx=0)
        emb_tl = nn.Embedding(2, embed_dim, padding_idx=0)
        embed_layers = emb_t, emb_l, emb_u, emb_su, emb_sl, emb_tu, emb_tl

        self.MultiEmbed = MultiEmbed(ex, embed_dim, embed_layers)
        self.SelfAttn = SelfAttn(embed_dim, embed_dim)
        self.Embed = Embed(ex, embed_dim, l_dim-1, embed_layers)
        self.Attn = Attn(emb_l, l_dim-1)

    def forward(self, traj, mat1, mat2, vec, traj_len):
        # long(N, M, [u, l, t]), float(N, M, M, 2), float(L, L), float(N, M), long(N)
        joint, delta = self.MultiEmbed(traj, mat1, traj_len)  # (N, M, emb), (N, M, M, emb)
        self_attn = self.SelfAttn(joint, delta, traj_len)  # (N, M, emb)
        self_delta = self.Embed(traj[:, :, 1], mat2, vec, traj_len)  # (N, M, L, emb)
        output = self.Attn(self_attn, self_delta, traj_len)  # (N, L)
        return output


================================================
FILE: train.py
================================================
from load import *
import time
import random
from torch import optim
import torch.utils.data as data
from tqdm import tqdm
from models import *


def calculate_acc(prob, label):
    # log_prob (N, L), label (N), batch_size [*M]
    acc_train = [0, 0, 0, 0]
    for i, k in enumerate([1, 5, 10, 20]):
        # topk_batch (N, k)
        _, topk_predict_batch = torch.topk(prob, k=k)
        for j, topk_predict in enumerate(to_npy(topk_predict_batch)):
            # topk_predict (k)
            if to_npy(label)[j] in topk_predict:
                acc_train[i] += 1

    return np.array(acc_train)


def sampling_prob(prob, label, num_neg):
    num_label, l_m = prob.shape[0], prob.shape[1]-1  # prob (N, L)
    label = label.view(-1)  # label (N)
    init_label = np.linspace(0, num_label-1, num_label)  # (N), [0 -- num_label-1]
    init_prob = torch.zeros(size=(num_label, num_neg+len(label)))  # (N, num_neg+num_label)

    random_ig = random.sample(range(1, l_m+1), num_neg)  # (num_neg) from (1 -- l_max)
    while len([lab for lab in label if lab in random_ig]) != 0:  # no intersection
        random_ig = random.sample(range(1, l_m+1), num_neg)

    global global_seed
    random.seed(global_seed)
    global_seed += 1

    # place the pos labels ahead and neg samples in the end
    for k in range(num_label):
        for i in range(num_neg + len(label)):
            if i < len(label):
                init_prob[k, i] = prob[k, label[i]]
            else:
                init_prob[k, i] = prob[k, random_ig[i-len(label)]]

    return torch.FloatTensor(init_prob), torch.LongTensor(init_label)  # (N, num_neg+num_label), (N)


class DataSet(data.Dataset):
    def __init__(self, traj, m1, v, label, length):
        # (NUM, M, 3), (NUM, M, M, 2), (L, L), (NUM, M), (NUM), (NUM)
        self.traj, self.mat1, self.vec, self.label, self.length = traj, m1, v, label, length

    def __getitem__(self, index):
        traj = self.traj[index].to(device)
        mats1 = self.mat1[index].to(device)
        vector = self.vec[index].to(device)
        label = self.label[index].to(device)
        length = self.length[index].to(device)
        return traj, mats1, vector, label, length

    def __len__(self):  # no use
        return len(self.traj)


class Trainer:
    def __init__(self, model, record):
        # load other parameters
        self.model = model.to(device)
        self.records = record
        self.start_epoch = record['epoch'][-1] if load else 1
        self.num_neg = 10
        self.interval = 1000
        self.batch_size = 1 # N = 1
        self.learning_rate = 3e-3
        self.num_epoch = 100
        self.threshold = np.mean(record['acc_valid'][-1]) if load else 0  # 0 if not update

        # (NUM, M, 3), (NUM, M, M, 2), (L, L), (NUM, M, M), (NUM, M), (NUM) i.e. [*M]
        self.traj, self.mat1, self.mat2s, self.mat2t, self.label, self.len = \
            trajs, mat1, mat2s, mat2t, labels, lens
        # nn.cross_entropy_loss counts target from 0 to C - 1, so we minus 1 here.
        self.dataset = DataSet(self.traj, self.mat1, self.mat2t, self.label-1, self.len)
        self.data_loader = data.DataLoader(dataset=self.dataset, batch_size=self.batch_size, shuffle=False)

    def train(self):
        # set optimizer
        optimizer = optim.Adam(self.model.parameters(), lr=self.learning_rate, weight_decay=0)
        scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1000, gamma=1)

        for t in range(self.num_epoch):
            # settings or validation and test
            valid_size, test_size = 0, 0
            acc_valid, acc_test = [0, 0, 0, 0], [0, 0, 0, 0]

            bar = tqdm(total=part)
            for step, item in enumerate(self.data_loader):
                # get batch data, (N, M, 3), (N, M, M, 2), (N, M, M), (N, M), (N)
                person_input, person_m1, person_m2t, person_label, person_traj_len = item

                # first, try batch_size = 1 and mini_batch = 1

                input_mask = torch.zeros((self.batch_size, max_len, 3), dtype=torch.long).to(device)
                m1_mask = torch.zeros((self.batch_size, max_len, max_len, 2), dtype=torch.float32).to(device)
                for mask_len in range(1, person_traj_len[0]+1):  # from 1 -> len
                    # if mask_len != person_traj_len[0]:
                    #     continue
                    input_mask[:, :mask_len] = 1.
                    m1_mask[:, :mask_len, :mask_len] = 1.

                    train_input = person_input * input_mask
                    train_m1 = person_m1 * m1_mask
                    train_m2t = person_m2t[:, mask_len - 1]
                    train_label = person_label[:, mask_len - 1]  # (N)
                    train_len = torch.zeros(size=(self.batch_size,), dtype=torch.long).to(device) + mask_len

                    prob = self.model(train_input, train_m1, self.mat2s, train_m2t, train_len)  # (N, L)

                    if mask_len <= person_traj_len[0] - 2:  # only training
                        # nn.utils.clip_grad_norm_(self.model.parameters(), 10)
                        prob_sample, label_sample = sampling_prob(prob, train_label, self.num_neg)
                        loss_train = F.cross_entropy(prob_sample, label_sample)
                        loss_train.backward()
                        optimizer.step()
                        optimizer.zero_grad()
                        scheduler.step()

                    elif mask_len == person_traj_len[0] - 1:  # only validation
                        valid_size += person_input.shape[0]
                        # v_prob_sample, v_label_sample = sampling_prob(prob_valid, valid_label, self.num_neg)
                        # loss_valid += F.cross_entropy(v_prob_sample, v_label_sample, reduction='sum')
                        acc_valid += calculate_acc(prob, train_label)

                    elif mask_len == person_traj_len[0]:  # only test
                        test_size += person_input.shape[0]
                        # v_prob_sample, v_label_sample = sampling_prob(prob_valid, valid_label, self.num_neg)
                        # loss_valid += F.cross_entropy(v_prob_sample, v_label_sample, reduction='sum')
                        acc_test += calculate_acc(prob, train_label)

                bar.update(self.batch_size)
            bar.close()

            acc_valid = np.array(acc_valid) / valid_size
            print('epoch:{}, time:{}, valid_acc:{}'.format(self.start_epoch + t, time.time() - start, acc_valid))

            acc_test = np.array(acc_test) / test_size
            print('epoch:{}, time:{}, test_acc:{}'.format(self.start_epoch + t, time.time() - start, acc_test))

            self.records['acc_valid'].append(acc_valid)
            self.records['acc_test'].append(acc_test)
            self.records['epoch'].append(self.start_epoch + t)

            if self.threshold < np.mean(acc_valid):
                self.threshold = np.mean(acc_valid)
                # save the model
                torch.save({'state_dict': self.model.state_dict(),
                            'records': self.records,
                            'time': time.time() - start},
                           'best_stan_win_1000_' + dname + '.pth')

    def inference(self):
        user_ids = []
        for t in range(self.num_epoch):
            # settings or validation and test
            valid_size, test_size = 0, 0
            acc_valid, acc_test = [0, 0, 0, 0], [0, 0, 0, 0]
            cum_valid, cum_test = [0, 0, 0, 0], [0, 0, 0, 0]

            for step, item in enumerate(self.data_loader):
                # get batch data, (N, M, 3), (N, M, M, 2), (N, M, M), (N, M), (N)
                person_input, person_m1, person_m2t, person_label, person_traj_len = item

                # first, try batch_size = 1 and mini_batch = 1

                input_mask = torch.zeros((self.batch_size, max_len, 3), dtype=torch.long).to(device)
                m1_mask = torch.zeros((self.batch_size, max_len, max_len, 2), dtype=torch.float32).to(device)
                for mask_len in range(1, person_traj_len[0] + 1):  # from 1 -> len
                    # if mask_len != person_traj_len[0]:
                    #     continue
                    input_mask[:, :mask_len] = 1.
                    m1_mask[:, :mask_len, :mask_len] = 1.

                    train_input = person_input * input_mask
                    train_m1 = person_m1 * m1_mask
                    train_m2t = person_m2t[:, mask_len - 1]
                    train_label = person_label[:, mask_len - 1]  # (N)
                    train_len = torch.zeros(size=(self.batch_size,), dtype=torch.long).to(device) + mask_len

                    prob = self.model(train_input, train_m1, self.mat2s, train_m2t, train_len)  # (N, L)

                    if mask_len <= person_traj_len[0] - 2:  # only training
                        continue

                    elif mask_len == person_traj_len[0] - 1:  # only validation
                        acc_valid = calculate_acc(prob, train_label)
                        cum_valid += calculate_acc(prob, train_label)

                    elif mask_len == person_traj_len[0]:  # only test
                        acc_test = calculate_acc(prob, train_label)
                        cum_test += calculate_acc(prob, train_label)

                print(step, acc_valid, acc_test)

                if acc_valid.sum() == 0 and acc_test.sum() == 0:
                    user_ids.append(step)


if __name__ == '__main__':
    # load data
    dname = 'NYC'
    file = open('./data/' + dname + '_data.pkl', 'rb')
    file_data = joblib.load(file)
    # tensor(NUM, M, 3), np(NUM, M, M, 2), np(L, L), np(NUM, M, M), tensor(NUM, M), np(NUM)
    [trajs, mat1, mat2s, mat2t, labels, lens, u_max, l_max] = file_data
    mat1, mat2s, mat2t, lens = torch.FloatTensor(mat1), torch.FloatTensor(mat2s).to(device), \
                               torch.FloatTensor(mat2t), torch.LongTensor(lens)

    # the run speed is very flow due to the use of location matrix (also huge memory cost)
    # please use a partition of the data (recommended)
    part = 100
    trajs, mat1, mat2t, labels, lens = \
        trajs[:part], mat1[:part], mat2t[:part], labels[:part], lens[:part]

    ex = mat1[:, :, :, 0].max(), mat1[:, :, :, 0].min(), mat1[:, :, :, 1].max(), mat1[:, :, :, 1].min()

    stan = Model(t_dim=hours+1, l_dim=l_max+1, u_dim=u_max+1, embed_dim=50, ex=ex, dropout=0)
    num_params = 0

    for name in stan.state_dict():
        print(name)

    for param in stan.parameters():
        num_params += param.numel()
    print('num of params', num_params)

    load = False

    if load:
        checkpoint = torch.load('best_stan_win_' + dname + '.pth')
        stan.load_state_dict(checkpoint['state_dict'])
        start = time.time() - checkpoint['time']
        records = checkpoint['records']
    else:
        records = {'epoch': [], 'acc_valid': [], 'acc_test': []}
        start = time.time()

    trainer = Trainer(stan, records)
    trainer.train()
    # trainer.inference()