def test_formality_score(files=None):
if files is None:
files = {
'rule_based': ['./data/Family_Relationships/bpe_outputs/formal.rule_based.bpe'],
'pbmt': ['./data/Family_Relationships/bpe_outputs/formal.pbmt.bpe'],
'nmt_baseline': ['./data/Family_Relationships/bpe_outputs/formal.nmt_baseline.bpe'],
'nmt_copy': ['./data/Family_Relationships/bpe_outputs/formal.nmt_copy.bpe'],
'nmt_combined': ['./data/Family_Relationships/bpe_outputs/formal.nmt_combined.bpe'],
}
embedding_path = './new_exp_fr/embedding/embedding.bpe.big.txt'
embedding, vocab_hash = embedding_api.load_word_embedding(embedding_path)
nn = NNModel(np.array(embedding),mode='eval')
nn.batch_size = 128
nn.build_basic_rnn_model()
eval_log={}
for key in files.keys():
if type(files[key])==type([]):
fm_files=files[key]+'.bpe'
else:
fm_files=[files[key]+'.bpe']
data=preprocess(informal_src_list=[],formal_src_list=fm_files,embedding_path=embedding_path,shuffle=False)
result = nn.predict_prob([t.x for t in data], model_path='./new_exp_fr/classifier/model/1700model.ckpt')
score=0
for s in result:
score+=s[1]
print(key,score/len(data))
eval_log[key]=score/len(data)
return eval_log
def test():
test = pickle.load(open('./new_exp_fr/classifier/test.pkl', 'rb'))
embedding_path = './new_exp_fr/embedding/corpus.fine_tune_embedding.epoch.10'
embedding,vocab_hash = embedding_api.load_word_embedding(embedding_path)
nn = NNModel(np.array(embedding),mode='eval')
nn.build_basic_rnn_model()
nn.evaluate([t.x for t in test],[t.y for t in test],model_path='')
def predict(model_path,file_path='./new_exp_fr/classifier/val.pkl',embedding_path='./new_exp_fr/embedding/corpus.fine_tune_embedding.epoch.10'):
test = pickle.load(open(file_path, 'rb'))
embedding, vocab_hash = embedding_api.load_word_embedding(embedding_path)
nn = NNModel(np.array(embedding),mode='predict')
nn.batch_size=10000
nn.build_basic_rnn_model()
result=nn.predict_prob([t.x for t in test], model_path=model_path)
return test,result
def use_nn_model():
train = pickle.load(open('./new_exp_fr/classifier/train.pkl', 'rb'))
val = pickle.load(open('./new_exp_fr/classifier/val.pkl', 'rb'))
embedding_path = './new_exp_fr/embedding/embedding.bpe.big.txt'
embedding, vocab_hash = embedding_api.load_word_embedding(embedding_path)
nn=NNModel(np.array(embedding),mode='train')
nn.build_basic_rnn_model()
nn.train_model([t.x for t in train],[t.y for t in train],[t.x for t in val],[t.y for t in val],
continue_train=False, previous_model_path='./new_exp_fr/classifier/model/990model.ckpt')
def preprocess(informal_src_list,formal_src_list,embedding_path,output_path=None,shuffle=True):
vectors,vocab_hash=embedding_api.load_word_embedding(embedding_path)
all_data=[]
for src in informal_src_list:
with open(src,'r',encoding='utf-8') as f:
for line in f:
d=Data(line.strip().split(), 0, line.strip())
d.str2index(vocab_hash,with_unk=False)
all_data.append(d)
for src in formal_src_list:
with open(src,'r',encoding='utf-8') as f:
for line in f:
d=Data(line.strip().split(), 1, line.strip())
d.str2index(vocab_hash,with_unk=False)
all_data.append(d)
if shuffle:
random.shuffle(all_data)
if output_path is not None:
pickle.dump(all_data,open(output_path,'wb'),protocol=True)
return all_data
def cal_formality_score_for_each_sentence(output_dir,files=None):
if files is None:
files = {
'rule_based': ['./data/Family_Relationships/bpe_outputs/formal.rule_based.bpe'],
'pbmt': ['./data/Family_Relationships/bpe_outputs/formal.pbmt.bpe'],
'nmt_baseline': ['./data/Family_Relationships/bpe_outputs/formal.nmt_baseline.bpe'],
'nmt_copy': ['./data/Family_Relationships/bpe_outputs/formal.nmt_copy.bpe'],
'nmt_combined': ['./data/Family_Relationships/bpe_outputs/formal.nmt_combined.bpe'],
}
embedding_path = './new_exp_fr/embedding/embedding.bpe.big.txt'
embedding, vocab_hash = embedding_api.load_word_embedding(embedding_path)
nn = NNModel(np.array(embedding),mode='eval')
nn.batch_size = 128
nn.build_basic_rnn_model()
for key in files.keys():
data=preprocess(informal_src_list=[],formal_src_list=files[key],embedding_path=embedding_path,shuffle=False)
result = nn.predict_prob([t.x for t in data], model_path='./new_exp_fr/classifier/model/1700model.ckpt')
base_name=os.path.basename(files[key])
with open(os.path.join(output_dir,base_name+'.formality_score'),'w',encoding='utf-8') as fw:
for r in result:
fw.write(str(r[1])+'\n')
def evaluate_one_formality(input_file_path,is_inf):
embedding_path = './new_exp_fr/embedding/embedding.bpe.big.txt'
embedding, vocab_hash = embedding_api.load_word_embedding(embedding_path)
nn = NNModel(np.array(embedding), mode='eval')
nn.batch_size = 128
nn.build_basic_rnn_model()
if is_inf:
data = preprocess(informal_src_list=[input_file_path], formal_src_list=[], embedding_path=embedding_path,
shuffle=False)
else:
data = preprocess(informal_src_list=[], formal_src_list=[input_file_path], embedding_path=embedding_path,
shuffle=False)
result = nn.predict_prob([t.x for t in data], model_path='./new_exp_fr/classifier/model/1700model.ckpt')
score = 0
if is_inf:
for s in result:
score += s[0]
else:
for s in result:
score += s[1]
print(score / len(data))
return score/len(data)