def compute(self, data, scale_res=True):
res = {}
for method in self.method:
res["ruber_refer_bert_%s" % method] = []
for tmp_data in data:
origin_candidate = tmp_data['candidate']
origin_reference = tmp_data['reference']
if not isinstance(origin_reference, list):
origin_reference = [origin_reference]
tmp_res = {}
for method in self.method:
tmp_res[method] = []
for r in origin_reference:
cand_tokens = torch.tensor(self.tokenizer.encode(origin_candidate)).unsqueeze(0)
ref_tokens = torch.tensor(self.tokenizer.encode(r)).unsqueeze(0)
if torch.cuda.is_available():
cand_tokens = cand_tokens.cuda()
ref_tokens = ref_tokens.cuda()
cand_embed_list = torch.squeeze(self.model(cand_tokens)[0]).detach().cpu().numpy().tolist()
ref_embed_list = torch.squeeze(self.model(ref_tokens)[0]).detach().cpu().numpy().tolist()
for method in self.method:
tmp_res[method].append(
self.score(cand_embed_list=cand_embed_list, ref_embed_list=ref_embed_list, pooling_type=method))
for method in self.method:
res["ruber_refer_bert_%s" % method].append(max(tmp_res[method]))
if scale_res:
for key in res:
res[key] = normalize_score(res[key])
return res
def compute(self, data, scale_res=True):
res = {}
for method in self.method:
res["%s_%s" % (self.name, method)] = []
for k, tmp_data in enumerate(data):
st = time.time()
origin_candidate = tmp_data['candidate']
origin_reference = tmp_data['reference']
if not isinstance(origin_reference, list):
origin_reference = [origin_reference]
tmp_res = {}
for method in self.method:
tmp_res[method] = []
for r in origin_reference:
cand_embed_list = sent2vec(tokenize(origin_candidate, self.tokenizer), self.embed)
ref_embed_list = sent2vec(tokenize(r, self.tokenizer), self.embed)
for method in self.method:
tmp_res[method].append(
self.score(cand_embed_list=cand_embed_list, ref_embed_list=ref_embed_list, pooling_type=method))
for method in self.method:
res["%s_%s" % (self.name, method)].append(max(tmp_res[method]))
if scale_res:
for key in res:
res[key] = normalize_score(res[key])
return res
def compute(self, data, scale_res=True):
ruber_bert_graph = tf.Graph()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
res = {}
with tf.Session(config=config, graph=ruber_bert_graph) as sess:
model = ruber_unrefer_model_bert(hidden_dim=self.model.config.hidden_size)
model.print_parameters()
if tf.train.get_checkpoint_state(self.model_path):
print("Reading model parameters from %s" % (self.model_path))
model.saver.restore(sess, tf.train.latest_checkpoint(self.model_path))
else:
print("Created model with fresh parameters.")
sess.run(tf.global_variables_initializer())
data_ = self.pro_data(data, pooling_type=self.method)
if scale_res:
score_list = normalize_score(self.score(model, sess, data_))
else:
score_list = self.score(model, sess, data_)
res["ruber_unrefer_bert_%s" % self.method] = score_list
return res
def compute(self, data, scale_res=True):
ruber_graph = tf.Graph()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
data = self.pro_data(data)
with tf.Session(config=config, graph=ruber_graph) as sess:
model = ruber_unrefer_model(
vocab_size=self.vocab_size,
embed_dim=self.embed_dim,
hidden_dim=self.hidden_dim)
model.print_parameters()
if tf.train.get_checkpoint_state(self.model_path):
print("Reading model parameters from %s" % (self.model_path))
model.saver.restore(sess, tf.train.latest_checkpoint(self.model_path))
else:
print("Created model with fresh parameters.")
sess.run(tf.global_variables_initializer())
if scale_res:
score_list = normalize_score(self.score(model, sess, data))
else:
score_list = self.score(model, sess, data)
return {"ruber_unrefer": score_list}
def compute(self, data):
ruber_refer_score = self.ruber_refer.compute(data)
ruber_unrefer_score = self.ruber_unrefer.compute(data)
res = {}
res.update(ruber_refer_score)
res.update(ruber_unrefer_score)
for method in self.method:
for key1 in ruber_refer_score:
for key2 in ruber_unrefer_score:
ruber_score = self.hybird_score(ruber_refer_score[key1], ruber_unrefer_score[key2], method=method)
res["%s_%s (with %s, %s)" % (
"ruber_bert" if self.bert else "ruber", method, key1, key2)] = normalize_score(ruber_score)
return res