def pre_train(pre_train_path):
model = train.pre_train()
loss, auc = train.test(model)
with open('logs/log.txt', 'a') as f:
f.write('# {}\n'.format(conf.model_type))
f.write('base_loss={};\n'.format(loss))
f.write('base_auc={};\n'.format(auc))
torch.save(model.state_dict(), pre_train_path)
def me_train(model, me_path):
me_model = train.meta_train(model)
torch.save(me_model.state_dict(), me_path)
loss, auc = train.test(me_model, warm=False)
me_warm = train.meta_test(me_path, model_name='me', with_normal=True)
with open('logs/log.txt', 'a') as f:
f.write('me_loss={};\n'.format(loss))
f.write('me_auc={};\n'.format(auc))
f.write('me_warm={};\n'.format(me_warm))
def best_auc_func(params):
model = train.new_model(pre_train_path)
train_ctw.meta_train(model, params)
loss, auc = train.test(model, warm=False)
global best_auc
global best_params
if auc > best_auc:
torch.save(model.state_dict(), ours_path)
best_auc = auc
best_params = params
print('best:{}'.format(best_auc))
return {'loss': -auc, 'status': STATUS_OK}
def save_tables():
for model_type in ['FM', 'DNN', 'WD', 'DeepFM']:
# for model_type in ['DNN', 'WD', 'DeepFM']:
conf.model_type = model_type
pre_train_path1 = "models/pre-trained-{}-{}.pth".format(
model_type, conf.cvr_flag)
me_path1 = "models/me-{}-{}.pth".format(model_type, conf.cvr_flag)
ours_path1 = "models/ours-{}-{}.pth".format(model_type, conf.cvr_flag)
conf.table_path = 'logs/table_base_cold_{}.csv'.format(model_type)
train.test(pre_train_path1, warm=False, save_table=True)
conf.table_path = 'logs/table_me_cold_{}.csv'.format(model_type)
train.test(me_path1, warm=False, save_table=True)
conf.table_path = 'logs/table_ours_cold_{}.csv'.format(model_type)
train.test(ours_path1, warm=False, save_table=True)
conf.table_path = 'logs/table_base_warm_{}.csv'.format(model_type)
train.meta_test(pre_train_path1,
with_normal=False,
save_table=True,
base=True)
conf.table_path = 'logs/table_me_warm_{}.csv'.format(model_type)
train.meta_test(me_path1, with_normal=False, save_table=True)
conf.table_path = 'logs/table_ours_warm_{}.csv'.format(model_type)
train.meta_test(ours_path1, with_normal=False, save_table=True)
print('{}-------------saved!!!!!!!!!!!!!!!!!!!!!'.format(model_type))
def ours_train(ours_path):
params = models[conf.model_type]
fmin(best_auc_func, params, algo=tpe.suggest, max_evals=max_evals, show_progressbar=False)
loss, auc = train.test(ours_path, warm=False)
ours_warm = train.meta_test(ours_path, model_name='ours', with_normal=False)
with open('logs/log.txt', 'a') as f:
f.write('ours_loss={};\n'.format(loss))
f.write('ours_auc={};\n'.format(auc))
f.write('ours_warm={};\n'.format(ours_warm))
global best_params
with open('logs/best.txt', 'a') as f:
f.write('# {}\n'.format(conf.model_type))
f.write('best_{}={};\n'.format(conf.model_type, best_params))
def hyper_param_analysis(rhos, alphas, path):
ps = {
'alpha': None,
'amsgrad': False,
'batch_n_ID': 50,
'gamma': 1.0,
'lr': 0.01,
'p_lr': 0.0001,
'p_lr_decay': 1.0,
'rho': None,
'weight_decay': 1e-8
}
# if os.path.exists(path):
# os.remove(path)
with open(path, 'a') as f:
f.write('# {}\n'.format(conf.model_type))
for rho in rhos:
for alpha in alphas:
ps['rho'] = rho
ps['alpha'] = alpha
# path
cold_path = "models/mer-trained-rho({})-alpha({}).pth".format(
rho, alpha)
warm_path = 'models/mer-tested-rho({})-alpha({}).pth'.format(
rho, alpha)
# cold
model = train.new_model(pre_train_path)
model = train_ctw.meta_train(model, ps, const.train_n_epoch)
torch.save(model.state_dict(), cold_path)
cold_loss, cold_auc = train.test(cold_path, warm=False)
# warm
ours_warm = train.meta_test(cold_path, 'ours', False)
warm_losses, warm_aucs = zip(*ours_warm['ours'])
losses = ',' + ','.join(map(str, warm_losses))
aucs = ',' + ','.join(map(str, warm_aucs))
torch.save(model.state_dict(), warm_path)
msg = ','.join(
[str(rho),
str(alpha),
str(cold_loss),
str(cold_auc)]) + losses + aucs + '\n'
print(msg)
f.write(msg)
