def run(args):
parameters = RnnParameterData(loc_emb_size=args.loc_emb_size, uid_emb_size=args.uid_emb_size, tim_emb_size=args.tim_emb_size,
app_size=args.app_size, app_encoder_size=args.app_encoder_size,acc_threshold=args.acc_threshold,
hidden_size=args.hidden_size, dropout_p=args.dropout_p, top_size=args.top_size,
data_name=args.data_name, lr=args.learning_rate,users_start=args.users_start,users_end=args.users_end,
lr_step=args.lr_step, lr_decay=args.lr_decay, L2=args.L2, rnn_type=args.rnn_type, loss_alpha=args.loss_alpha,
optim=args.optim, lr_schedule=args.lr_schedule, attn_type=args.attn_type, app_emb_mode=args.app_emb_mode,
clip=args.clip, epoch_max=args.epoch_max, history_mode=args.history_mode, loss_beta=args.loss_beta,
model_mode=args.model_mode, data_path=args.data_path, save_path=args.save_path, baseline_mode=args.baseline_mode,
loc_emb_negative=args.loc_emb_negative, loc_emb_batchsize=args.loc_emb_batchsize, input_mode=args.input_mode, test_start = args.test_start)
"""metric"""
T = 10
T0 = T
lr = parameters.lr
metrics = {'train_loss': [], 'valid_loss': [], 'avg_app_auc': [], 'avg_app_map': [], 'avg_app_precision': [], 'avg_app_recall': [], 'avg_loc_top1': [], 'avg_loc_top5': [],'avg_loc_top10': [], 'avg_uid_top1': [], 'avg_uid_top10': [], 'valid_acc': {}}
if args.candidate == 'loc_entropy':
candidate = candidate_entropy(parameters.data_neural, start=parameters.users_start, end=parameters.users_end, mode=args.reverse_mode)
elif args.candidate == 'app_number':
candidate = candidate_number(parameters.data_neural, start=parameters.users_start, end=parameters.users_end, mode=args.reverse_mode)
else:
candidate_tmp = [u for u in parameters.data_neural.keys()]
candidate = candidate_tmp[parameters.users_start:parameters.users_end]
parameters.uid_size = max(candidate) + 1
print('Candidate:',candidate[:5])
print('Candidate:{} max user_id:{} min user_id:{}'.format(args.candidate, max(candidate), min(candidate)))
app_number = static_app(parameters,candidate)
print('App: max app number used:{} min appp number userd:{} average appp number userd:{}'.format(max(app_number), min(app_number), np.mean(app_number)))
"""baseline"""
if parameters.baseline_mode == 'App':
print("======Run baseline: beyasian_prediction==========")
auc, users_auc, f1, users_f1, precision, users_precision, recall, users_recall = beyasian_prediction(parameters,candidate=candidate)
print('==> NB auc: {:.4f} map: {:.4f} precision: {:.4f} recall: {:.4f}'.format(float(auc), float(f1),float(precision), float(recall)))
print("======Run baseline: history average==========")
auc, users_auc, f1, users_f1, precision, users_precision, recall, users_recall = history_average(parameters,candidate=candidate)
print('==> HA auc: {:.4f} map: {:.4f} precision: {:.4f} recall: {:.4f}'.format(float(auc), float(f1),float(precision), float(recall)))
print("======Run baseline: most recent==========")
auc, users_auc, f1, users_f1, precision, users_precision, recall, users_recall = most_recently(parameters,candidate=candidate)
json.dump(users_f1, open("users_f1_MRU_App.json","w"))
print('==> MR auc: {:.4f} map: {:.4f} precision: {:.4f} recall: {:.4f}'.format(float(auc), float(f1),float(precision), float(recall)))
print("======Run baseline: most popular==========")
auc, users_auc, f1, users_f1, precision, users_precision, recall, users_recall = most_popular(parameters,candidate=candidate)
print('==> MF auc: {:.4f} map: {:.4f} precision: {:.4f} recall: {:.4f}'.format(float(auc), float(f1),float(precision), float(recall)))
elif parameters.baseline_mode == 'Loc':
print("======Run Loc baseline: one order markov==========")
avg_acc_top1,avg_acc_top5,avg_acc_top10, user_acc = markov(parameters,candidate=candidate)
print('==> Markov acc@1: {:.4f} acc@5: {:.4f} acc@10: {:.4f}'.format(float(avg_acc_top1), float(avg_acc_top5),float(avg_acc_top10)))
print("======Run Loc baseline: most recently==========")
avg_acc_top1,user_acc = most_recently_loc(parameters,candidate=candidate)
json.dump(user_acc, open("avg_acc_top1_MRU_Loc.json","w"))
print('==> MR acc@1: {:.4f} '.format(float(avg_acc_top1)))
print("======Run Loc baseline: most popular==========")
avg_acc_top1,avg_acc_top5,avg_acc_top10, user_acc = most_popular_loc(parameters,candidate=candidate)
print('==> MF acc@1: {:.4f} acc@5: {:.4f} acc@10: {:.4f}'.format(float(avg_acc_top1), float(avg_acc_top5),float(avg_acc_top10)))
elif parameters.baseline_mode == None:
print("================Run models=============")
"""get loc embedding graph"""
loc_old2newid = {}
loc_old2newid,loc_graph = get_loc_graph(parameters, candidate)
parameters.loc_size = len(loc_old2newid)
parameters.uid_size = max(candidate)+1
#Model Training
print('Split training and testing data', datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
if 'Topk' in parameters.model_mode:
print('using topk model')
user_topk = generate_input_topk(parameters,'train',loc_old2newid, mode2=None, candidate=candidate)
else:
user_topk = None
if parameters.input_mode == 'short':
data_train, train_idx = generate_input(parameters, 'train', loc_old2newid, user_topk, mode2=parameters.history_mode,candidate=candidate)
data_test, test_idx = generate_input(parameters, 'test', loc_old2newid, user_topk, mode2=parameters.history_mode,candidate=candidate)
elif parameters.input_mode == 'short_history':
data_train, train_idx = generate_input_history(parameters, 'train', loc_old2newid, user_topk, mode2=parameters.history_mode,candidate=candidate)
data_test, test_idx = generate_input_history(parameters, 'test', loc_old2newid, user_topk, mode2=parameters.history_mode,candidate=candidate)
elif parameters.input_mode == 'long':
data_train, train_idx = generate_input_long_history(parameters, 'train', loc_old2newid, user_topk, mode2=parameters.history_mode,candidate=candidate)
data_test, test_idx = generate_input_long_history(parameters, 'test', loc_old2newid, user_topk, mode2=parameters.history_mode,candidate=candidate)
#print('Generating Line first order similarity')
#loc_emb_data_1 = SmapleDataset(loc_graph, 0) #the edge with link, return one pair with label=1
#loc_emb_data_loaer_1 = DataLoader(loc_emb_data_1, shuffle=True, batch_size=parameters.loc_emb_batchsize, num_workers=4)
#print('Generating Line second order similarity')
#loc_emb_data_2 = SmapleDataset(loc_graph, parameters.loc_emb_negative) #negative sample, return 5 pairs with label=-1
#loc_emb_data_loaer_2 = DataLoader(loc_emb_data_2, shuffle=True, batch_size=parameters.loc_emb_batchsize//6, num_workers=4)
#print('data len:', len(loc_emb_data_1))
#T1 = int(len(loc_emb_data_1)//parameters.loc_emb_batchsize)
"""Model Init"""
print('Model Init!', datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
for mi in range(1):
#"""Pre App"""
if parameters.model_mode in ['AppPre']:
model = AppPre(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreGtr']:
model = AppPreGtr(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUser']:
model = AppPreUser(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUserCGtr']:
model = AppPreUserCGtr(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUserRGtr']:
model = AppPreUserRGtr(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUserPGtr']:
model = AppPreUserPGtr(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUserCGtrTopk']:
model = AppPreUserCGtrTopk(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUserCGtrHis']:
model = AppPreUserCGtrHis(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUserCGtrHisAttn']:
model = AppPreUserCGtrHisAttn(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUserPGtrTopk']:
model = AppPreUserPGtrTopk(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocRUserCGtrHis']:
model = AppPreLocRUserCGtrHis(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocCUserCGtrHis']:
model = AppPreLocCUserCGtrHis(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocRUserCGtr']:
model = AppPreLocRUserCGtr(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocRUserCGtrTopk']:
model = AppPreLocRUserCGtrTopk(parameters=parameters).cuda()
criterion = AppLoss(parameters=parameters).cuda()
#"""Pre Loc"""
elif parameters.model_mode in ['LocPre']:
model = LocPre(parameters=parameters).cuda()
criterion = LocLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['LocPreGt','LocPreGtr']:
model = LocPreGtr(parameters=parameters).cuda()
criterion = LocLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['LocPreUser']:
model = LocPreUser(parameters=parameters).cuda()
criterion = LocLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['LocPreUserGt', 'LocPreUserGtr']:
model = LocPreUserGtr(parameters=parameters).cuda()
criterion = LocLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['LocPreUserGtTopk', 'LocPreUserGtrTopk']:
model = LocPreUserGtrTopk(parameters=parameters).cuda()
criterion = LocLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['LocPreUserGtRec', 'LocPreUserGtrRec']:
model = LocPreUserGtrRec(parameters=parameters).cuda()
criterion = LocLoss(parameters=parameters).cuda()
#"""Iden user"""
elif parameters.model_mode in ['UserIden']:
model = UserIden(parameters=parameters).cuda()
criterion = nn.NLLLoss().cuda()
#"""Pre App, Loc and user"""
elif parameters.model_mode in ['AppPreUserIden']:
model = AppPreUserIden(parameters=parameters).cuda()
criterion = AppUserLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocPreGtr']:
model = AppPreLocPreGtr(parameters=parameters).cuda()
criterion = AppLocLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreUserIdenGt','AppPreUserIdenGtr']:
model = AppPreUserIdenGtr(parameters=parameters).cuda()
criterion = AppUserLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocUserIdenGtr']:
model = AppPreLocUserIdenGtr(parameters=parameters).cuda()
criterion = AppUserLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocPreUserIden']:
model = AppPreLocPreUserIden(parameters=parameters).cuda()
criterion = AppLocUserLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocPreUserIdenGt','AppPreLocPreUserIdenGtr']:
model = AppPreLocPreUserIdenGtr(parameters=parameters).cuda()
criterion = AppLocUserLoss(parameters=parameters).cuda()
elif parameters.model_mode in ['AppPreLocPreUserIdenGtrLinear']:
model = AppPreLocPreUserIdenGtrLinear(parameters=parameters).cuda()
criterion = AppLocUserLoss(parameters=parameters).cuda()
#"""For embedding"""
elif parameters.model_mode in ['LocEmbed', 'LocPreUserGtrLocEmb']:
line_1st = Line_1st(parameters.loc_size, parameters.loc_emb_size).cuda()
line_2nd = Line_2nd(parameters.loc_size, parameters.loc_emb_size).cuda()
model = LocPreUserGtrLocEmb(parameters=parameters,line_1st=line_1st,line_2nd=line_2nd,alpha=1).cuda()
criterion = nn.NLLLoss().cuda()
T0 = T-1
print(model)
params = list(model.parameters())
k = 0
for i in params:
l = 1
for j in i.size():
l *= j
k = k + l
print("The number of parameters:" + str(k))
"Forward network with randomly initialization "
for epoch in range(1):
optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=lr, weight_decay=parameters.L2)
#scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=parameters.lr_step, factor=parameters.lr_decay, threshold=1e-3)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'max', patience=parameters.lr_step,factor=parameters.lr_decay, threshold=1e-3)
#scheduler = optim.lr_scheduler.MultiStepLR(optimizer, [10,20,30], gamma=parameters.lr_decay)
#prediction_all[epoch] = {}
#train dataset but test model without feedback
model, avg_train_loss, prediction = run_simple(data_train, train_idx, 'train_test', parameters.input_mode, lr, parameters.clip, model, optimizer, criterion, parameters.model_mode, parameters.test_start, parameters.acc_threshold)
print('========================>Epoch:{:0>2d} lr:{}<=================='.format(epoch,lr))
print('==>Train Epoch:{:0>2d} Loss:{:.4f} lr:{}'.format(epoch, avg_train_loss, lr))
metrics['train_loss'].append(avg_train_loss)
#prediction_all[epoch]['train'] = prediction
avg_loss, avg_acc, users_acc, prediction = run_simple(data_test, test_idx, 'test', parameters.input_mode, lr, parameters.clip, model,optimizer, criterion, parameters.model_mode, parameters.test_start, parameters.acc_threshold)
print('==>Test Loss:{:.4f}'.format(avg_loss))
print('==>Test Acc App_AUC:{:.4f} App_map:{:.4f} App_Precision:{:.4f} App_Recall:{:.4f} '.format(avg_acc['app_auc'], avg_acc['app_map'], avg_acc['app_precision'], avg_acc['app_recall']))
print(' Loc_top1:{:.4f} Loc_top5:{:.4f} Loc_top10:{:.4f}'.format(avg_acc['loc_top1'],avg_acc['loc_top5'], avg_acc['loc_top10']))
print(' Uid_top1:{:.4f} Uid_top10:{:.4f}'.format(avg_acc['uid_top1'], avg_acc['uid_top10']))
metrics['valid_loss'].append(avg_loss) #total average loss
metrics['valid_acc'][epoch] = users_acc #accuracy for each user
metrics['avg_app_auc'].append(0) #total average accuracy
metrics['avg_app_map'].append(0)
metrics['avg_app_precision'].append(0)
metrics['avg_app_recall'].append(0)
metrics['avg_loc_top1'].append(0)
metrics['avg_loc_top5'].append(0)
metrics['avg_loc_top10'].append(0)
metrics['avg_uid_top1'].append(0)
metrics['avg_uid_top10'].append(0)
#prediction_all[epoch]['test'] = prediction
st = time.time()
start_time = time.time()
for epoch in range(1, parameters.epoch):
#prediction_all[epoch] = {}
if epoch%T < T0:
#optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=lr, weight_decay=parameters.L2)
#scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', patience=parameters.lr_step, factor=parameters.lr_decay, threshold=1e-3)
model, avg_train_loss, prediction = run_simple(data_train, train_idx, 'train', parameters.input_mode, lr, parameters.clip, model, optimizer, criterion, parameters.model_mode, parameters.test_start, parameters.acc_threshold)
print('========================>Epoch:{:0>2d} lr:{}<=================='.format(epoch,lr))
print('==>Train Epoch:{:0>2d} Loss:{:.4f} lr:{}'.format(epoch, avg_train_loss, lr))
metrics['train_loss'].append(avg_train_loss)
#prediction_all[epoch]['train'] = prediction
avg_loss, avg_acc, users_acc, prediction = run_simple(data_test, test_idx, 'test', parameters.input_mode, lr, parameters.clip, model,optimizer, criterion, parameters.model_mode, parameters.test_start, parameters.acc_threshold)
print('==>Test Loss:{:.4f}'.format(avg_loss))
print('==>Test Acc App_AUC:{:.4f} App_map:{:.4f} App_Precision:{:.4f} App_Recall:{:.4f} '.format(avg_acc['app_auc'], avg_acc['app_map'], avg_acc['app_precision'], avg_acc['app_recall']))
print(' Loc_top1:{:.4f} Loc_top5:{:.4f} Loc_top10:{:.4f}'.format(avg_acc['loc_top1'],avg_acc['loc_top5'], avg_acc['loc_top10']))
print(' Uid_top1:{:.4f} Uid_top10:{:.4f}'.format(avg_acc['uid_top1'], avg_acc['uid_top10']))
metrics['valid_loss'].append(avg_loss) #total average loss
metrics['valid_acc'][epoch] = users_acc #accuracy for each user
metrics['avg_app_auc'].append(avg_acc['app_auc']) #total average accuracy
metrics['avg_app_map'].append(avg_acc['app_map'])
metrics['avg_app_precision'].append(avg_acc['app_precision'])
metrics['avg_app_recall'].append(avg_acc['app_recall'])
metrics['avg_loc_top1'].append(avg_acc['loc_top1'])
metrics['avg_loc_top5'].append(avg_acc['loc_top5'])
metrics['avg_loc_top10'].append(avg_acc['loc_top10'])
metrics['avg_uid_top1'].append(avg_acc['uid_top1'])
metrics['avg_uid_top10'].append(avg_acc['uid_top10'])
#prediction_all[epoch]['test'] = prediction
save_name_tmp = 'ep_' + str(epoch) + '_' + str(start_time) + '.m'
torch.save(model.state_dict(), parameters.save_path + 'tmp/' + save_name_tmp)
if parameters.lr_schedule == 'Loss':
if 'AppPre' in parameters.model_mode:
scheduler.step(avg_acc['app_map'])
elif 'LocPre' in parameters.model_mode:
scheduler.step(avg_acc['loc_top1'])
elif 'UserIden' in parameters.model_mode:
scheduler.step(avg_acc['uid_top1'])
lr_last = lr
lr = optimizer.param_groups[0]['lr']
if lr_last > lr:
if 'AppPre' in parameters.model_mode:
load_epoch = np.argmax(metrics['avg_app_map'])
elif 'LocPre' in parameters.model_mode:
load_epoch = np.argmax(metrics['avg_loc_top1'])
else:
load_epoch = np.argmax(metrics['avg_uid_top1'])
load_name_tmp = 'ep_' + str(load_epoch) + '_' + str(start_time) + '.m'
model.load_state_dict(torch.load(parameters.save_path + 'tmp/' + load_name_tmp))
print('load epoch={} model state'.format(load_epoch)) #lr decreased
if epoch == 1:
print('single epoch time cost:{}'.format(time.time() - start_time))
if lr <= 0.9 * 1e-6:
break
else:
optimizer1 = optim.Adam(filter(lambda p: p.requires_grad, line_1st.parameters()), lr=1e-3)
run_embedding(line_1st, loc_emb_data_loaer_1, 1, optimizer1, 0, epoch, parameters.epoch)
optimizer2 = optim.Adam(filter(lambda p: p.requires_grad, line_2nd.parameters()), lr=1e-3)
run_embedding(line_2nd, loc_emb_data_loaer_2, 2, optimizer2, parameters.loc_emb_negative, epoch, parameters.epoch)
line = Line(line_1st, line_2nd, alpha=1,name='epoch'+str(epoch)) #the ration of 1st and 2nd
line.save_emb()
overhead = time.time() - start_time
if 'AppPre' in parameters.model_mode:
load_epoch = np.argmax(metrics['avg_app_map'])
print('==>Test Best Epoch:{:0>2d} App_AUC:{:.4f} app_map:{:.4f} App_Precision:{:.4f} App_Recall:{:.4f} '.format(load_epoch, metrics['avg_app_auc'][load_epoch], metrics['avg_app_map'][load_epoch], metrics['avg_app_precision'][load_epoch], metrics['avg_app_recall'][load_epoch]))
elif 'LocPre' in parameters.model_mode:
load_epoch = np.argmax(metrics['avg_loc_top1'])
print('==>Test Best Epoch:{:0>2d} Loc_Top1:{:.4f} Loc_top10:{:.4f}'.format(load_epoch, metrics['avg_loc_top1'][load_epoch], metrics['avg_loc_top10'][load_epoch]))
else:
load_epoch = np.argmax(metrics['avg_uid_top1'])
print('==>Test Best Epoch:{:0>2d} Uid_Top1:{:.4f} Uid_top10:{:.4f}'.format(load_epoch, metrics['avg_uid_top1'][load_epoch], metrics['avg_uid_top10'][load_epoch]))
load_name_tmp = 'ep_' + str(load_epoch) + '_' + str(start_time) + '.m'
model.load_state_dict(torch.load(parameters.save_path + 'tmp/' + load_name_tmp))
save_name = args.model_mode + '_' + str(args.users_start) + '-' + str(args.users_end) + '_' + str(args.uid_emb_size) + '_' + \
str(args.hidden_size) + '_' + str(args.top_size)+ '_' + \
str(metrics['avg_app_auc'][load_epoch])[:6] + '_' + str(metrics['avg_app_map'][load_epoch])[:6] + '_' + \
str(metrics['avg_app_precision'][load_epoch])[:6]+ '_' + str(metrics['avg_app_recall'][load_epoch])[:6] + '_' + \
str(args.process_name) + '_' + str(args.loss_alpha)[:4] + '_' + str(args.loss_beta)[:4] + '_' + str(overhead/60)[:5]
#save_name = args.model_mode + '_' + str(args.users_start) + '-' + str(args.users_end) + '_' + str(args.app_encoder_size) + '_' + \
# str(args.hidden_size) + '_' + str(metrics['avg_app_map'][load_epoch])[:6] + '_' + \
# str(metrics['avg_loc_top1'][load_epoch])[:6]+ '_' + str(metrics['avg_uid_top1'][load_epoch])[:6] + '_' + \
# str(args.process_name) + '_' + str(args.loss_alpha)[:4] + '_' + str(args.loss_beta)[:4] + '_' + str(overhead/60)[:5]
json.dump(metrics['valid_acc'][load_epoch], open("users_acc_"+save_name+".json","w"))
"""saving embedding"""
if parameters.model_mode in ['LocPreUserGtrLocEmb','LocPreUserGtr']:
model.save_emb()
"""precess visualization"""
for p in range(1):
fig = plt.figure(dpi=300)
if args.plot_mode == 'both':
ax1 = plt.subplot(221)
plt.plot(metrics['train_loss'],'r-',label='train_loss')
plt.plot(metrics['valid_loss'],'b-',label='test_loss')
plt.legend(loc='best')
ax2 = plt.subplot(222)
plt.plot(metrics['avg_app_auc'],'g-',label='test_app_auc')
plt.plot(metrics['avg_app_map'],'y-',label='test_app_map')
plt.legend(loc='best')
ax3 = plt.subplot(223)
plt.plot(metrics['avg_loc_top1'],'g-',label='test_loc_top1')
plt.plot(metrics['avg_loc_top10'],'y-',label='test_loc_top10')
plt.legend(loc='best')
ax4 = plt.subplot(224)
plt.plot(metrics['avg_uid_top1'],'g-',label='test_uid_top1')
plt.plot(metrics['avg_uid_top10'],'y-',label='test_uid_top10')
plt.legend(loc='best')
else:
ax1 = plt.subplot(211)
plt.plot(metrics['train_loss'],'r-',label='train_loss')
plt.plot(metrics['valid_loss'],'b-',label='test_loss')
plt.legend(loc='best')
ax2 = plt.subplot(212)
if args.plot_mode == 'App':
plt.plot(metrics['avg_app_auc'],'g-',label='test_app_auc')
plt.plot(metrics['avg_app_map'],'y-',label='test_app_map')
elif args.plot_mode == 'Loc':
plt.plot(metrics['avg_loc_top1'],'g-',label='test_loc_top1')
plt.plot(metrics['avg_loc_top10'],'y-',label='test_loc_top10')
elif args.plot_mode == 'User':
plt.plot(metrics['avg_uid_top1'],'g-',label='test_uid_top1')
plt.plot(metrics['avg_uid_top10'],'y-',label='test_uid_top10')
elif args.plot_mode == 'Loc_emb':
plt.plot(metrics['avg_loc_emb_P'],'g-',label='avg_loc_emb_P')
plt.plot(metrics['avg_loc_emb_R'],'y-',label='avg_loc_emb_R')
plt.legend(loc='best')
plt.savefig(save_name + '.png')
precess = np.zeros([10,len(metrics['train_loss'])])
precess[0,:]=np.array(metrics['train_loss'])
precess[1,:]=np.array(metrics['valid_loss'])
precess[2,:]=np.array(metrics['avg_app_auc'])
precess[3,:]=np.array(metrics['avg_app_map'])
precess[4,:]=np.array(metrics['avg_loc_top1'])
precess[5,:]=np.array(metrics['avg_loc_top10'])
def run(args):
parameters = RnnParameterData(loc_emb_size=args.loc_emb_size,
uid_emb_size=args.uid_emb_size,
voc_emb_size=args.voc_emb_size,
tim_emb_size=args.tim_emb_size,
hidden_size=args.hidden_size,
dropout_p=args.dropout_p,
data_name=args.data_name,
lr=args.learning_rate,
lr_step=args.lr_step,
lr_decay=args.lr_decay,
L2=args.L2,
rnn_type=args.rnn_type,
optim=args.optim,
attn_type=args.attn_type,
clip=args.clip,
epoch_max=args.epoch_max,
history_mode=args.history_mode,
model_mode=args.model_mode,
data_path=args.data_path,
save_path=args.save_path)
argv = {
'loc_emb_size': args.loc_emb_size,
'uid_emb_size': args.uid_emb_size,
'voc_emb_size': args.voc_emb_size,
'tim_emb_size': args.tim_emb_size,
'hidden_size': args.hidden_size,
'dropout_p': args.dropout_p,
'data_name': args.data_name,
'learning_rate': args.learning_rate,
'lr_step': args.lr_step,
'lr_decay': args.lr_decay,
'L2': args.L2,
'act_type': 'selu',
'optim': args.optim,
'attn_type': args.attn_type,
'clip': args.clip,
'rnn_type': args.rnn_type,
'epoch_max': args.epoch_max,
'history_mode': args.history_mode,
'model_mode': args.model_mode
}
print('*' * 15 + 'start training' + '*' * 15)
print('model_mode:{} history_mode:{} users:{}'.format(
parameters.model_mode, parameters.history_mode, parameters.uid_size))
if parameters.model_mode in ['simple', 'simple_long']:
model = TrajPreSimple(parameters=parameters).cuda()
elif parameters.model_mode == 'attn_avg_long_user':
model = TrajPreAttnAvgLongUser(parameters=parameters).cuda()
elif parameters.model_mode == 'attn_local_long':
model = TrajPreLocalAttnLong(parameters=parameters).cuda()
if args.pretrain == 1:
model.load_state_dict(
torch.load("../pretrain/" + args.model_mode + "/res.m"))
if 'max' in parameters.model_mode:
parameters.history_mode = 'max'
elif 'avg' in parameters.model_mode:
parameters.history_mode = 'avg'
else:
parameters.history_mode = 'whole'
criterion = nn.NLLLoss().cuda()
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=parameters.lr,
weight_decay=parameters.L2)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
'max',
patience=parameters.lr_step,
factor=parameters.lr_decay,
threshold=1e-3)
lr = parameters.lr
metrics = {
'train_loss': [],
'valid_loss': [],
'accuracy': [],
'valid_acc': {}
}
candidate = parameters.data_neural.keys()
avg_acc_markov, users_acc_markov = markov(parameters, candidate)
metrics['markov_acc'] = users_acc_markov
if 'long' in parameters.model_mode:
long_history = True
else:
long_history = False
if long_history is False:
data_train, train_idx = generate_input_history(
parameters.data_neural,
'train',
mode2=parameters.history_mode,
candidate=candidate)
data_test, test_idx = generate_input_history(
parameters.data_neural,
'test',
mode2=parameters.history_mode,
candidate=candidate)
elif long_history is True:
if parameters.model_mode == 'simple_long':
data_train, train_idx = generate_input_long_history2(
parameters.data_neural, 'train', candidate=candidate)
data_test, test_idx = generate_input_long_history2(
parameters.data_neural, 'test', candidate=candidate)
else:
data_train, train_idx = generate_input_long_history(
parameters.data_neural, 'train', candidate=candidate)
data_test, test_idx = generate_input_long_history(
parameters.data_neural, 'test', candidate=candidate)
print('users:{} markov:{} train:{} test:{}'.format(
len(candidate), avg_acc_markov,
len([y for x in train_idx for y in train_idx[x]]),
len([y for x in test_idx for y in test_idx[x]])))
SAVE_PATH = args.save_path
tmp_path = 'checkpoint/'
os.mkdir(SAVE_PATH + tmp_path)
for epoch in range(parameters.epoch):
st = time.time()
if args.pretrain == 0:
model, avg_loss = run_simple(data_train, train_idx, 'train', lr,
parameters.clip, model, optimizer,
criterion, parameters.model_mode)
print('==>Train Epoch:{:0>2d} Loss:{:.4f} lr:{}'.format(
epoch, avg_loss, lr))
metrics['train_loss'].append(avg_loss)
avg_loss, avg_acc, users_acc = run_simple(data_test, test_idx, 'test',
lr, parameters.clip, model,
optimizer, criterion,
parameters.model_mode)
print('==>Test Acc:{:.4f} Loss:{:.4f}'.format(avg_acc, avg_loss))
metrics['valid_loss'].append(avg_loss)
metrics['accuracy'].append(avg_acc)
metrics['valid_acc'][epoch] = users_acc
save_name_tmp = 'ep_' + str(epoch) + '.m'
torch.save(model.state_dict(), SAVE_PATH + tmp_path + save_name_tmp)
scheduler.step(avg_acc)
lr_last = lr
lr = optimizer.param_groups[0]['lr']
if lr_last > lr:
load_epoch = np.argmax(metrics['accuracy'])
load_name_tmp = 'ep_' + str(load_epoch) + '.m'
model.load_state_dict(
torch.load(SAVE_PATH + tmp_path + load_name_tmp))
print('load epoch={} model state'.format(load_epoch))
if epoch == 0:
print('single epoch time cost:{}'.format(time.time() - st))
if lr <= 0.9 * 1e-5:
break
if args.pretrain == 1:
break
mid = np.argmax(metrics['accuracy'])
avg_acc = metrics['accuracy'][mid]
load_name_tmp = 'ep_' + str(mid) + '.m'
model.load_state_dict(torch.load(SAVE_PATH + tmp_path + load_name_tmp))
save_name = 'res'
json.dump({
'args': argv,
'metrics': metrics
},
fp=open(SAVE_PATH + save_name + '.rs', 'w'),
indent=4)
metrics_view = {'train_loss': [], 'valid_loss': [], 'accuracy': []}
for key in metrics_view:
metrics_view[key] = metrics[key]
json.dump({
'args': argv,
'metrics': metrics_view
},
fp=open(SAVE_PATH + save_name + '.txt', 'w'),
indent=4)
torch.save(model.state_dict(), SAVE_PATH + save_name + '.m')
for rt, dirs, files in os.walk(SAVE_PATH + tmp_path):
for name in files:
remove_path = os.path.join(rt, name)
os.remove(remove_path)
os.rmdir(SAVE_PATH + tmp_path)
return avg_acc
def run(args):
parameters = RnnParameterData(
loc_emb_size=args.loc_emb_size,
uid_emb_size=args.uid_emb_size, #500，40
cid_emb_size=args.cid_emb_size,
tim_emb_size=args.tim_emb_size, #50，10
hidden_size=args.hidden_size,
dropout_p=args.dropout_p, #500，0.3
data_name=args.data_name,
lr=args.learning_rate,
lr_step=args.lr_step,
lr_decay=args.lr_decay,
L2=args.L2,
optim=args.optim,
clip=args.clip,
epoch_max=args.epoch_max,
data_path=args.data_path,
save_path=args.save_path)
argv = {
'loc_emb_size': args.loc_emb_size,
'uid_emb_size': args.uid_emb_size,
'cid_emb_size': args.cid_emb_size,
'tim_emb_size': args.tim_emb_size,
'hidden_size': args.hidden_size,
'dropout_p': args.dropout_p,
'data_name': args.data_name,
'learning_rate': args.learning_rate,
'lr_step': args.lr_step,
'lr_decay': args.lr_decay,
'L2': args.L2,
'act_type': 'selu',
'optim': args.optim,
'clip': args.clip,
'epoch_max': args.epoch_max
}
auxiliary_rate = 0.05
model = PG2Net(parameters=parameters, weight=weight,
weight_cid=weight_cid).cuda()
if args.pretrain == 1:
model.load_state_dict(
torch.load("../pretrain/" + args.model_mode + "/res.m"))
criterion = nn.NLLLoss().cuda()
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=parameters.lr,
weight_decay=parameters.L2)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
'max',
patience=parameters.lr_step,
factor=parameters.lr_decay,
threshold=1e-3)
lr = parameters.lr
#Metrics
metrics = {
'train_loss': [],
'valid_loss': [],
'accuracy': [],
'valid_acc': {}
}
candidate = parameters.data_neural.keys() #937 users
data_train, train_idx = generate_input_long_history(parameters.data_neural,
'train',
candidate=candidate)
data_test, test_idx = generate_input_long_history(parameters.data_neural,
'test',
candidate=candidate)
print('users:{} markov:{} train:{} test:{}'.format(
len(candidate), avg_acc_markov,
len([y for x in train_idx for y in train_idx[x]]),
len([y for x in test_idx for y in test_idx[x]])))
SAVE_PATH = args.save_path
msg = 'users:{} markov:{} train:{} test:{}'.format(
len(candidate), avg_acc_markov,
len([y for x in train_idx for y in train_idx[x]]),
len([y for x in test_idx for y in test_idx[x]]))
with open(SAVE_PATH + "result.txt", "a") as file:
file.write(msg + "\n")
file.close()
tmp_path = 'checkpoint/'
if not SAVE_PATH + tmp_path:
os.mkdir(SAVE_PATH + tmp_path)
#Computing time similarity
time_sim_matrix = caculate_time_sim(parameters.data_neural) #(48,48)
#Import time and category relationship
poi_cid_tim = pickle.load(open('cid_time.pkl', 'rb'),
encoding='iso-8859-1')
#Import the spatial distance between each location
poi_distance_matrix = pickle.load(open('distance.pkl', 'rb'),
encoding='iso-8859-1')
for epoch in range(parameters.epoch):
pred = []
st = time.time()
if args.pretrain == 0:
model, avg_loss, pred = run_simple(
pred, data_train, train_idx, auxiliary_rate, 'train', lr,
parameters.clip, model, optimizer, criterion,
parameters.model_mode, time_sim_matrix, poi_distance_matrix,
poi_cid_tim)
print('auxiliary_rate:{}'.format(auxiliary_rate))
msg = 'auxiliary_rate:{}'.format(auxiliary_rate)
with open(SAVE_PATH + "result.txt", "a") as file:
file.write(msg + "\n")
file.close()
print('==>Train Epoch:{:0>2d} Loss:{:.4f} lr:{}'.format(
epoch, avg_loss, lr))
msg = '==>Train Epoch:{:0>2d} Loss:{:.4f} lr:{}'.format(
epoch, avg_loss, lr)
with open(SAVE_PATH + "result.txt", "a") as file:
file.write(msg + "\n")
file.close()
metrics['train_loss'].append(avg_loss)
avg_loss, avg_acc, users_acc, pred = run_simple(
pred, data_test, test_idx, auxiliary_rate, 'test', lr,
parameters.clip, model, optimizer, criterion,
parameters.model_mode, time_sim_matrix, poi_distance_matrix,
poi_cid_tim)
#print('==>Test Acc:{:.4f} Loss:{:.4f}'.format(avg_acc, avg_loss))
print(
'==>Rec@1:{:.4f} Rec@5:{:.4f} Rec@10:{:.4f} NDCG@1:{:.4f} NDCG@5:{:.4f} NDCG@10:{:.4f} Loss:{:.4f}'
.format(avg_acc[0], avg_acc[1], avg_acc[2], avg_acc[3], avg_acc[4],
avg_acc[5], avg_loss))
msg = '==>Rec@1:{:.4f} Rec@5:{:.4f} Rec@10:{:.4f} NDCG@1:{:.4f} NDCG@5:{:.4f} NDCG@10:{:.4f} Loss:{:.4f}'.format(
avg_acc[0], avg_acc[1], avg_acc[2], avg_acc[3], avg_acc[4],
avg_acc[5], avg_loss)
with open(SAVE_PATH + "result.txt", "a") as file:
file.write(msg + "\n")
file.close()
pickle.dump(pred, open("{}_our_nyc_loc.pkl".format(epoch), 'wb'))
metrics['valid_loss'].append(avg_loss)
metrics['accuracy'].append(avg_acc[0])
metrics['valid_acc'][epoch] = users_acc
save_name_tmp = 'ep_' + str(epoch) + '.m'
torch.save(model.state_dict(), SAVE_PATH + tmp_path + save_name_tmp)
scheduler.step(avg_acc[0])
lr_last = lr
lr = optimizer.param_groups[0]['lr']
if lr_last > lr:
load_epoch = np.argmax(metrics['accuracy'])
load_name_tmp = 'ep_' + str(load_epoch) + '.m'
model.load_state_dict(
torch.load(SAVE_PATH + tmp_path + load_name_tmp))
auxiliary_rate += 0.05
print('load epoch={} model state'.format(load_epoch))
msg = 'load epoch={} model state'.format(load_epoch)
with open(SAVE_PATH + "result.txt", "a") as file:
file.write(msg + "\n")
file.close()
if epoch == 0:
print('single epoch time cost:{}'.format(time.time() - st))
msg = 'single epoch time cost:{}'.format(time.time() - st)
with open(SAVE_PATH + "result.txt", "a") as file:
file.write(msg + "\n")
file.close()
if lr <= 0.9 * 1e-7:
break
if args.pretrain == 1:
break
mid = np.argmax(metrics['accuracy'])
avg_acc = metrics['accuracy'][mid]
load_name_tmp = 'ep_' + str(mid) + '.m'
print("Best model:", SAVE_PATH + tmp_path + load_name_tmp)
return avg_acc