def eval(self, model, poolsize, K, dataset):
"""
simple validation in a code pool.
:param model: Trained Model
:param poolsize: poolsize - size of the code pool, if -1, load the whole test set
:param K: Top K results
:param dataset: which dataset to evaluate on
:return: Accuracy, MRR, MAP, nDCG
"""
if self.valid_set is None: # load evaluation dataset
if dataset == "train":
self.valid_set = CodeSearchDataset(self.path, self.conf,
"train") # load train set
elif dataset == "val":
self.valid_set = CodeSearchDataset(self.path, self.conf,
"val") # load val set
else:
self.valid_set = CodeSearchDataset(self.path, self.conf,
"test") # load test set
data_loader = torch.utils.data.DataLoader(dataset=self.valid_set,
batch_size=poolsize,
shuffle=True,
drop_last=True,
num_workers=1)
model = model.eval()
accs, mrrs, maps, ndcgs = [], [], [], []
for qts, codes, _, qbs, _ in data_loader:
qts, codes, qbs = gVar(qts), gVar(codes), gVar(qbs)
code_repr = model.code_encoding(codes)
if self.conf['use_qb']:
qb_repr = model.qb_encoding(qbs)
else:
qb_repr = None
for i in range(poolsize):
qt = gVar(qts[i].expand(poolsize, -1))
qt_repr = model.qt_encoding(qt)
sims = model.score_qt_code_qb(qt_repr, code_repr,
qb_repr).data.cpu().numpy()
# sims = model.combine_qt_and_code(qt_repr, code_repr).data.cpu().numpy()
# sims = F.cosine_similarity(torch.concat(qt_repr,code_repr)).data.cpu().numpy()
# n_results = K
negsims = np.negative(sims)
predict = np.argsort(negsims)
# predict = np.argpartition(negsims, kth=n_results-1)
# predict = predict[:n_results]
predict = [int(k) for k in predict]
real = [i] # index of positive sample
accs.append(ACC(real, predict))
mrrs.append(MRR(real, predict))
maps.append(MAP(real, predict))
ndcgs.append(NDCG(real, predict))
logger.info('ACC={}, MRR={}, MAP={}, nDCG={}'.format(
np.mean(accs), np.mean(mrrs), np.mean(maps), np.mean(ndcgs)))
return np.mean(accs), np.mean(mrrs), np.mean(maps), np.mean(ndcgs)
def repr_code(self,model):
vecs=None
use_set = CodeSearchDataset(self.conf['workdir'],
self.conf['use_names'],self.conf['name_len'],
self.conf['use_apis'],self.conf['api_len'],
self.conf['use_tokens'],self.conf['tokens_len'])
data_loader = torch.utils.data.DataLoader(dataset=use_set, batch_size=1000,
shuffle=False, drop_last=False, num_workers=1)
for names,apis,toks in data_loader:
names, apis, toks = gVar(names), gVar(apis), gVar(toks)
reprs = model.code_encoding(names,apis,toks).data.cpu().numpy()
vecs=reprs if vecs is None else np.concatenate((vecs, reprs),0)
vecs = normalize(vecs)
save_vecs(vecs,self.path+self.conf['use_codevecs'])
return vecs
def train(self, model):
model.train()
log_every = self.conf['log_every']
valid_every = self.conf['valid_every']
save_every = self.conf['save_every']
batch_size = self.conf['batch_size']
nb_epoch = self.conf['nb_epoch']
train_set = CodeSearchDataset(
self.path, self.conf['train_name'], self.conf['name_len'],
self.conf['train_api'], self.conf['api_len'],
self.conf['train_tokens'], self.conf['tokens_len'],
self.conf['train_desc'], self.conf['desc_len'])
data_loader = torch.utils.data.DataLoader(
dataset=train_set,
batch_size=self.conf['batch_size'],
shuffle=True,
drop_last=True,
num_workers=1)
val_loss = {'loss': 1., 'epoch': 0}
for epoch in range(self.conf['reload'] + 1, nb_epoch):
itr = 1
losses = []
for names, apis, toks, good_descs, bad_descs in data_loader:
names, apis, toks, good_descs, bad_descs = gVar(names), gVar(
apis), gVar(toks), gVar(good_descs), gVar(bad_descs)
loss = model(names, apis, toks, good_descs, bad_descs)
losses.append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
if itr % log_every == 0:
logger.info('epo:[%d/%d] itr:%d Loss=%.5f' %
(epoch, nb_epoch, itr, np.mean(losses)))
losses = []
itr = itr + 1
# if epoch and epoch % valid_every == 0:
# logger.info("validating..")
# acc1, mrr, map, ndcg = self.eval(model,1000,1)
if epoch and epoch % save_every == 0:
self.save_model(model, epoch)
def train(self, model):
log_every = self.model_params['log_every']
save_every = self.model_params['save_every']
batch_size = self.model_params['batch_size']
nb_epoch = self.model_params['nb_epoch']
train_set = CodeSearchDataset(self.path,
self.model_params['train_name'],
self.model_params['name_len'],
self.model_params['train_api'],
self.model_params['api_len'],
self.model_params['train_tokens'],
self.model_params['tokens_len'],
self.model_params['train_desc'],
self.model_params['desc_len'],
load_in_memory=True)
data_loader = torch.utils.data.DataLoader(dataset=train_set,
batch_size=batch_size,
shuffle=True,
drop_last=True,
num_workers=4,
pin_memory=True)
for epoch in range(self.model_params['reload'] + 1, nb_epoch):
epoch_loss = []
losses = []
for itr, (names, apis, toks, good_descs,
bad_descs) in enumerate(data_loader, start=1):
names, apis, toks, good_descs, bad_descs = gVar(names), gVar(
apis), gVar(toks), gVar(good_descs), gVar(bad_descs)
loss = model.train()(names, apis, toks, good_descs, bad_descs)
losses.append(loss.item())
epoch_loss.append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
if itr % log_every == 0:
logger.info('epo:[{}/{}] itr:{} Loss={:.5f}'.format(
epoch, nb_epoch, itr, np.mean(losses)))
losses = []
if epoch and epoch % save_every == 0:
self.save_model_epoch(model, epoch)
logger.info('[SUMMARY] epo:[{}/{}] Loss={:.5f}'.format(
epoch, nb_epoch, np.mean(epoch_loss)))
def repr_code(self, model, norm=True):
logging.info("Start Code Representation")
use_set = CodeSearchDataset(self.model_params['workdir'],
self.model_params['use_names'],
self.model_params['name_len'],
self.model_params['use_apis'],
self.model_params['api_len'],
self.model_params['use_tokens'],
self.model_params['tokens_len'],
load_in_memory=True)
data_loader = torch.utils.data.DataLoader(dataset=use_set,
batch_size=1000,
shuffle=False,
drop_last=False,
num_workers=2,
pin_memory=True)
vecs = []
logging.debug("Calculating code vectors")
for itr, (names, apis, toks) in enumerate(data_loader, start=1):
names, apis, toks = gVar(names), gVar(apis), gVar(toks)
reprs = model.eval().code_encoding(names, apis,
toks).data.cpu().numpy()
vecs.append(reprs)
if itr % 100 == 0:
logger.info('itr:{}/{}'.format(itr, len(use_set) // 1000))
logging.debug("Concatenating all vectors")
vecs = np.concatenate(vecs, 0)
if norm:
logger.debug("Normalizing...")
vecs = normalize(vecs)
logging.debug("Writing to disk - vectors")
save_vecs(vecs, self.path + self.model_params['use_codevecs'])
return vecs
def eval(self, model, poolsize, K):
"""
simple validation in a code pool.
@param: poolsize - size of the code pool, if -1, load the whole test set
"""
def ACC(real,predict):
sum=0.0
for val in real:
try:
index=predict.index(val)
except ValueError:
index=-1
if index!=-1:
sum=sum+1
return sum/float(len(real))
def MAP(real,predict):
sum=0.0
for id,val in enumerate(real):
try:
index=predict.index(val)
except ValueError:
index=-1
if index!=-1:
sum=sum+(id+1)/float(index+1)
return sum/float(len(real))
def MRR(real,predict):
sum=0.0
for val in real:
try:
index=predict.index(val)
except ValueError:
index=-1
if index!=-1:
sum=sum+1.0/float(index+1)
return sum/float(len(real))
def NDCG(real,predict):
dcg=0.0
idcg=IDCG(len(real))
for i,predictItem in enumerate(predict):
if predictItem in real:
itemRelevance=1
rank = i+1
dcg+=(math.pow(2,itemRelevance)-1.0)*(math.log(2)/math.log(rank+1))
return dcg/float(idcg)
def IDCG(n):
idcg=0
itemRelevance=1
for i in range(n):
idcg+=(math.pow(2,itemRelevance)-1.0)*(math.log(2)/math.log(i+2))
return idcg
if self.valid_set is None: #load test dataset
self.valid_set=CodeSearchDataset(self.path,
self.conf['valid_name'],self.conf['name_len'],
self.conf['valid_api'],self.conf['api_len'],
self.conf['valid_tokens'],self.conf['tokens_len'],
self.conf['valid_desc'],self.conf['desc_len'])
data_loader = torch.utils.data.DataLoader(dataset=self.valid_set, batch_size=poolsize,
shuffle=True, drop_last=True, num_workers=1)
accs,mrrs,maps,ndcgs=[],[],[],[]
for names, apis, toks, descs, _ in tqdm(data_loader):
names, apis, toks = gVar(names), gVar(apis), gVar(toks)
code_repr=model.code_encoding(names, apis, toks)
for i in range(poolsize):
desc=gVar(descs[i].expand(poolsize,-1))
desc_repr=model.desc_encoding(desc)
n_results = K
sims = F.cosine_similarity(code_repr, desc_repr).data.cpu().numpy()
negsims=np.negative(sims)
predict=np.argsort(negsims)#predict = np.argpartition(negsims, kth=n_results-1)
predict = predict[:n_results]
predict = [int(k) for k in predict]
real=[i]
accs.append(ACC(real,predict))
mrrs.append(MRR(real,predict))
maps.append(MAP(real,predict))
ndcgs.append(NDCG(real,predict))
logger.info('ACC={}, MRR={}, MAP={}, nDCG={}'.format(np.mean(accs),np.mean(mrrs),np.mean(maps),np.mean(ndcgs)))
return np.mean(accs),np.mean(mrrs),np.mean(maps),np.mean(ndcgs)
def train(self, model):
"""
Trains an initialized model
:param model: Initialized model
:return: None
"""
log_every = self.conf['log_every']
valid_every = self.conf['valid_every']
batch_size = self.conf['batch_size']
nb_epoch = self.conf['nb_epoch']
max_patience = self.conf['patience']
train_set = CodeSearchDataset(self.path, self.conf, "train")
data_loader = torch.utils.data.DataLoader(dataset=train_set,
batch_size=batch_size,
shuffle=True,
drop_last=True,
num_workers=1)
# MRR for the Best Saved model, if reload > 0, else -1
if self.conf['reload'] > 0:
_, max_mrr, _, _ = self.eval(model, 50, 10, "val")
else:
max_mrr = -1
patience = 0
for epoch in range(self.conf['reload'] + 1, nb_epoch):
itr = 1
losses = []
model = model.train()
for qts, good_codes, bad_codes, good_qbs, bad_qbs in data_loader:
qts, good_codes, bad_codes, good_qbs, bad_qbs = gVar(
qts), gVar(good_codes), gVar(bad_codes), gVar(
good_qbs), gVar(bad_qbs)
loss = model(qts, good_codes, bad_codes, good_qbs, bad_qbs)
losses.append(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
if itr % log_every == 0:
logger.info('epo:[%d/%d] itr:%d Loss=%.5f' %
(epoch, nb_epoch, itr, np.mean(losses)))
losses = []
itr = itr + 1
if epoch % valid_every == 0:
logger.info("validating..")
acc1, mrr, map, ndcg = self.eval(model, 50, 10, "val")
if mrr > max_mrr:
self.save_model(model, epoch)
patience = 0
print("Model improved. Saved model at %d epoch" % epoch)
max_mrr = mrr
else:
print("Model didn't improve for ", patience + 1, " epochs")
patience += 1
if patience >= max_patience:
logger.info("Patience Limit Reached. Stopping Training")
break
def eval(self, model, poolsize, K, test_all=True):
"""
simple validation in a code pool.
@param: poolsize - size of the code pool, if -1, load the whole test set
"""
def ACC(real, predict):
_sum = 0.0
for val in real:
if val in predict:
_sum += 1
return _sum / float(len(real))
def MAP(real, predict):
_sum = 0.0
for _id, val in enumerate(real, start=1):
try:
index = predict.index(val) + 1
except ValueError:
continue
else:
_sum += _id / float(index)
return _sum / float(len(real))
def MRR(real, predict):
_sum = 0.0
for val in real:
try:
index = predict.index(val) + 1
except ValueError:
continue
else:
_sum += 1.0 / float(index)
return _sum / float(len(real))
def NDCG(real, predict):
dcg = 0.0
idcg = IDCG(len(real))
for i, predictItem in enumerate(predict):
if predictItem in real:
itemRelevance = 1
rank = i + 1
dcg += (math.pow(2, itemRelevance) -
1.0) * (math.log(2) / math.log(rank + 1))
return dcg / float(idcg)
def IDCG(n):
idcg = 0
itemRelevance = 1
for i in range(n):
idcg += (math.pow(2, itemRelevance) - 1.0) * (math.log(2) /
math.log(i + 2))
return idcg
# load test dataset
if self.validation_set is None:
self.validation_set = CodeSearchDataset(
self.path,
self.model_params['valid_name'],
self.model_params['name_len'],
self.model_params['valid_api'],
self.model_params['api_len'],
self.model_params['valid_tokens'],
self.model_params['tokens_len'],
self.model_params['valid_desc'],
self.model_params['desc_len'],
load_in_memory=True)
data_loader = torch.utils.data.DataLoader(dataset=self.validation_set,
batch_size=poolsize,
shuffle=False,
drop_last=True,
num_workers=1,
pin_memory=True)
accs, mrrs, maps, ndcgs = [], [], [], []
for names, apis, toks, descs, _ in tqdm(data_loader):
names, apis, toks = gVar(names), gVar(apis), gVar(toks)
code_repr = model.eval().code_encoding(names, apis, toks)
for it in range(poolsize):
desc = gVar(descs[it].expand(poolsize, -1))
desc_repr = model.eval().desc_encoding(desc)
n_results = K
sims = F.cosine_similarity(code_repr,
desc_repr).data.cpu().numpy()
negsims = np.negative(sims)
prediction = np.argpartition(negsims, kth=n_results - 1)
prediction = prediction[:n_results]
# sort the codes by their sim, simulate a ranking
prediction = [
y for _, y in sorted(zip(negsims[prediction], prediction))
]
real_value = [it]
accs.append(ACC(real_value, prediction))
mrrs.append(MRR(real_value, prediction))
maps.append(MAP(real_value, prediction))
ndcgs.append(NDCG(real_value, prediction))
mean_acc = np.mean(accs)
mean_mrr = np.mean(mrrs)
mean_map = np.mean(maps)
mean_ndcg = np.mean(ndcgs)
logger.info('ACC={}, MRR={}, MAP={}, nDCG={}'.format(
mean_acc, mean_mrr, mean_map, mean_ndcg))
return mean_acc, mean_mrr, mean_map, mean_ndcg