def main(path_global,path_user,path_session,user, path_tmpfile0,epoch,Date_sess, path_userData,joining=False):
Date_sess = Date_sess.replace('/','')
config_train=Config_train()
path_tmpfile = os.path.join(path_tmpfile0,'feature')
path_session_user=os.path.join(path_session,user)
if not os.path.exists(path_session_user):
print(path_session_user+' is not exist')
return
try:
datapath = modules.get_sessionfile(path_session_user)
users = [user]
r_all, r_sc_all, r_time_all, D_user = data_initial(path_global,path_user,users,path_tmpfile0,path_userData)
D_other = [r_all] + r_sc_all + r_time_all
iter = modules.iterData(datapath, D_user, D_other, r_sc_all, r_time_all,batch_size=config_train.train_batch_size, rate_skip = config_train.skip_rate,rate_skip_neg=1-r_all,config_global=config_model,joining=joining)
data = next(iter)
step = 0
while data!='__STOP__':
nb_tmpfiles = len(os.listdir(path_tmpfile))
while nb_tmpfiles>32:
time.sleep(1)
nb_tmpfiles = len(os.listdir(path_tmpfile))
np.save(os.path.join(path_tmpfile,user+'-'+epoch+'-'+Date_sess+'-'+str(step)+'.npy'),data)
data = next(iter)
step += 1
with open(os.path.join(path_tmpfile0,'trainedlist.txt'),'a+') as f:
f.write(user+'-'+epoch+'-'+Date_sess+'-'+str(step)+'.npy'+'\n')
except:
print('failed to get tmpfile %s'%path_session_user)
shutil.rmtree(path_session_user)
def train(path_global,
path_user,
path_session,
resultpath,
model,
joining=False):
config_train = Config_train()
config_model = modelconfig()
config_train.feature_dim = config_model.get_nb_features()[0]
if joining:
config_train.feature_dim += 1
tf.reset_default_graph()
if model == 'lr':
X_holder, y_holder, learning_rate, predict_y, loss, optimizer, train_op, grads, accuracy = simple_lr(
config_train)
elif model == 'lr-dense':
X_holder, y_holder, learning_rate, predict_y, loss, optimizer, train_op, grads, accuracy = simple_lr_dense(
config_train)
global_step = tf.train.get_or_create_global_step()
train_op = tf.group(train_op, [tf.assign_add(global_step, 1)])
saver = tf.train.Saver(max_to_keep=10)
session = tf.Session()
path_ckpt = os.path.join(resultpath, config_train.CKPT_path)
if not os.path.exists(path_ckpt):
os.mkdir(path_ckpt)
ckpt_file = tf.train.latest_checkpoint(path_ckpt)
if ckpt_file:
saver.restore(session, ckpt_file)
else:
init = tf.global_variables_initializer()
session.run(init)
learning_rate_ = config_train.learning_rate
step = 0
completed = 0
data_train = os.listdir(path_session)
usernb = 0
for path in data_train:
usernb += 1
datapath = modules.get_sessionfile(os.path.join(path_session, path))
print('training on ' + path)
time0 = time.time()
users = path
r_all, r_sc_all, r_time_all, D_user = data_initial(
path_global, path_user, users, Sc, resultpath)
D_other = [r_all] + r_sc_all + r_time_all
iter = modules.iterData(datapath,
D_user,
D_other,
r_sc_all,
r_time_all,
batch_size=config_train.train_batch_size,
rate_skip=config_train.skip_rate,
rate_skip_neg=1 - r_all,
config_global=config_model,
joining=joining)
epoch = 0
loss_ = 0
while epoch < config_train.epochs:
data = next(iter)
if data == '__STOP__':
iter = modules.iterData(
datapath,
D_user,
D_other,
r_sc_all,
r_time_all,
batch_size=config_train.train_batch_size,
rate_skip=1 - r_all,
config_global=config_model,
joining=joining)
epoch += 1
continue
x0, y0, _ = data
y0 = np.array(y0)
y0 = np.reshape(y0, (len(y0), 1))
if step % config_train.step_saveckpt == 0:
saver.save(session,
os.path.join(path_ckpt, 'model.ckpt'),
global_step=global_step)
print('loss in step-{} is {}'.format(step, loss_))
_, loss_, = session.run([train_op, loss],
feed_dict={
X_holder: x0,
y_holder: y0,
learning_rate: learning_rate_
})
step += 1
completed += 1
KK = len(data_train) - completed
print(
'training used time %0.2f mins and needs about %0.2f mins for the rest users'
% ((time.time() - time0) / 60, (time.time() - time0) / 60 * KK))
def test(path_global,
path_user,
path_session,
resultpath,
model,
joining=False):
config_train = Config_train()
config_model = modelconfig()
config_train.feature_dim = config_model.get_nb_features()[0]
if joining:
config_train.feature_dim += 1
tf.reset_default_graph()
if model == 'lr':
X_holder, y_holder, learning_rate, predict_y, loss, optimizer, train_op, grads, accuracy = simple_lr(
config_train)
elif model == 'lr-dense':
X_holder, y_holder, learning_rate, predict_y, loss, optimizer, train_op, grads, accuracy = simple_lr_dense(
config_train)
saver = tf.train.Saver(max_to_keep=10)
session = tf.Session()
path_ckpt = os.path.join(resultpath, config_train.CKPT_path)
if not os.path.exists(path_ckpt):
print('no model')
return
ckpt_file = tf.train.latest_checkpoint(path_ckpt)
if ckpt_file:
saver.restore(session, ckpt_file)
else:
print('no model')
return
ckpt_backup = os.path.join(resultpath, "ckpt_backup")
if not os.path.exists(ckpt_backup):
os.mkdir(ckpt_backup)
tmpfile = ckpt_file
cfile = tmpfile + ".*"
cmdstr = "cp " + cfile + " " + ckpt_backup
os.system(cmdstr)
idx = ckpt_file.find('model.ckpt-') + len('model.ckpt-')
if not os.path.exists(os.path.join(resultpath, 'test')):
os.mkdir(os.path.join(resultpath, 'test'))
predictpath0 = os.path.join(resultpath, 'test',
'predict' + ckpt_file[idx:])
learning_rate_ = config_train.learning_rate
data_train = os.listdir(path_session)
auc = 0
auc_each = []
predictpaths = []
for path in data_train:
predictpath = predictpath0 + '-' + path + '.txt'
datapath = modules.get_sessionfile(os.path.join(path_session, path))
users = path
r_all, r_sc_all, r_time_all, D_user = data_initial(
path_global, path_user, users, Sc, resultpath)
D_other = [r_all] + r_sc_all + r_time_all
iter = modules.iterData(datapath,
D_user,
D_other,
r_sc_all,
r_time_all,
batch_size=config_train.testlines,
rate_skip=config_train.skip_rate,
rate_skip_neg=0,
config_global=config_model,
joining=joining)
data = next(iter)
X, y, user = data
y = np.array(y)
y = np.reshape(y, (len(y), 1))
y_p = []
batch_size = config_train.train_batch_size
i = 0
while i * batch_size < len(y):
X_test = X[i * batch_size:(i + 1) * batch_size]
y_test = y[i * batch_size:(i + 1) * batch_size]
y_p0 = session.run(predict_y,
feed_dict={
X_holder: X_test,
y_holder: y_test,
learning_rate: learning_rate_
})
y_p.append(y_p0)
i += 1
y_p = np.concatenate(y_p)
tmp = [
'\t'.join([user[ii], str(int(y[ii][0])),
str(y_p[ii][0])]) for ii in range(len(y))
]
with open(predictpath, 'w') as f:
f.write('\n'.join(tmp))
auctmp = modules.getAUC(predictpath)
auc += auctmp
auc_each.append('%0.4f' % auctmp)
predictpaths.append(predictpath)
print('testing on file %s with auc = %0.4f and ckptfile is %s' %
(os.path.join(path_session, path), auctmp, ckpt_file))
auc = auc / len(data_train)
print('testing on path %s with average auc = %0.4f and ckptfile is %s' %
(path_session, auc, ckpt_file))
if os.path.exists(os.path.join(resultpath, 'test-auc-ckpt.txt')):
with open(os.path.join(resultpath, 'test-auc-ckpt.txt'), 'r') as f:
stmp = f.read().strip().split('\n')
auc0 = float(stmp[-1].split('\t')[0])
else:
stmp = []
auc0 = 0
if auc > auc0:
stmp.append('%0.4f' % auc + '\t' + '\t'.join(auc_each) + '\t' +
tmpfile)
with open(os.path.join(resultpath, 'test-auc-ckpt.txt'), 'w') as f:
f.write('\n'.join(stmp))
else:
tmpfile = tmpfile.replace('/ckpt/', '/ckpt_backup/')
os.remove(tmpfile + ".meta")
os.remove(tmpfile + ".index")
os.remove(tmpfile + ".data-00000-of-00001")
return auc, predictpaths