def aim(Phen):
f = np.zeros([1, 1])
CV = np.zeros([1, 1])
for i in Phen:
train_model = Train.train(i)
train_model.training()
#print(i)
#print(sum(i==1))
c = sum(i == 1)
if (c == 0):
c = 1
f = np.append(f, np.array([[train_model.acc]]), axis=0)
CV = np.append(CV, np.array([[math.log(c, 64)]]), axis=0)
return f[1:], CV[1:] # 返回目标函数值矩阵
# data_loader
sem_data_loader = ReadSenData(config)
tra_sen_data, dev_sen_data = sem_data_loader.read_data()
word_data_loader = ReadData(config)
tra_data, dev_data = word_data_loader.read_data()
# vocab
if os.path.isfile(config.fact_word_vocab):
tra_fact_word_vocab = read_pkl(config.fact_word_vocab)
tra_fact_char_vocab = read_pkl(config.fact_char_vocab)
if config.read_sen:
tra_fact_sen_vocab = read_pkl(config.fact_sen_vocab)
else:
if not os.path.isdir(config.save_vocab_path):
os.makedirs(config.save_vocab_path)
tra_fact_word_vocab = FactWordVocab([tra_data, dev_data], config)
pickle.dump(tra_fact_word_vocab, open(config.fact_word_vocab, 'wb'))
tra_fact_char_vocab = FactCharVocab([tra_data, dev_data], config)
pickle.dump(tra_fact_char_vocab, open(config.fact_char_vocab, 'wb'))
if config.read_sen:
tra_fact_sen_vocab = FactSenVocab([tra_sen_data, dev_sen_data], config)
pickle.dump(tra_fact_sen_vocab, open(config.fact_sen_vocab, 'wb'))
model = MyBertModel(config)
if config.use_cuda:
model.cuda()
# train
train(model, tra_data, dev_data, tra_fact_word_vocab, config)
if gpu:
config.add_args('Train', 'use_cuda', 'True')
word_data_loader = ReadData(config)
usual_tra_data_set, virus_tra_data_set, usual_dev_data_set, virus_dev_data_set, usual_vat_data, virus_vat_data \
= word_data_loader.read_data(tokenizer)
usual_eval_data, virus_eval_data = word_data_loader.read_eval_data(tokenizer)
# vocab
if os.path.isfile(config.fact_word_src_vocab):
tag_vocab = read_pkl(config.fact_word_tag_vocab)
else:
if not os.path.isdir(config.save_vocab_path):
os.makedirs(config.save_vocab_path)
tag_vocab = FactWordTagVocab([usual_tra_data_set, virus_tra_data_set])
pickle.dump(tag_vocab, open(config.fact_word_tag_vocab, 'wb'))
distinguish_vocab = DistinguishVocab()
bert_model = MyBertModel(config)
distinguish_model = DistinguishModel(config)
if config.use_cuda:
bert_model = bert_model.cuda()
distinguish_model = distinguish_model.cuda()
# train
train(bert_model, distinguish_model, usual_tra_data_set, virus_tra_data_set, usual_dev_data_set, virus_dev_data_set, tag_vocab, config,
distinguish_vocab, tokenizer, usual_eval_data, virus_eval_data, usual_vat_data, virus_vat_data)
from Model import Model
from Train import DataLoader, BiOptimizer, train
from Evaluation import Visualizer
device = None
model = Model(width=8, height=8).to(device)
data_loader = DataLoader(data_dir='../Data/npimage8.npy',
batch_size=128,
device=device)
optim = BiOptimizer(model)
visualizer = Visualizer(model, device)
optim.step()
# for batch in data_loader:
# _ = model(batch[0])
# train.CTR = True
train.train(model, data_loader, optim, 1)
visualizer.visualize()
visualizer.show()
imgs, labels, fn = data
imgs = imgs.to(device)
labels = labels.to(device)
out_hm, out_wh, out_off = self.model(imgs)
# out = self.model(imgs)
# out_hm = torch.cat([o['hm'].squeeze() for o in out], dim=0)
# out_off = torch.cat([o['off'].squeeze() for o in out], dim=0)
# out_wh = torch.cat([o['wh'].squeeze() for o in out], dim=0)
loss_hm = self.focal_loss(out_hm, labels[:, 0])
loss_off = self.regloss(out_off, labels[:, [1,2]])
loss_wh = self.regloss(out_wh, labels[:, [3,4]])
losses = {}
losses['loss_hm'] = loss_hm
losses['loss_wh'] = loss_wh
losses['loss_offset'] = loss_off
loss = loss_hm + 0.1 * loss_wh + loss_off
return losses, loss
if __name__=='__main__':
context = Context(cfg)
train(context)
if gpu:
config.add_args('Train', 'use_cuda', 'True')
word_data_loader = ReadData(config)
kd_tra_data_set, ca_tra_data_set, kd_dev_data_set, ca_dev_data_set = word_data_loader.read_data(
tokenizer)
kd_tra_data, kd_tra_data_gold = kd_tra_data_set[0], kd_tra_data_set[1]
ca_tra_data, ca_tra_data_gold = ca_tra_data_set[0], ca_tra_data_set[1]
kd_dev_data, kd_dev_data_gold = kd_dev_data_set[0], kd_dev_data_set[1]
ca_dev_data, ca_dev_data_gold = ca_dev_data_set[0], ca_dev_data_set[1]
# vocab
if os.path.isfile(config.fact_word_src_vocab):
tag_vocab = read_pkl(config.fact_word_tag_vocab)
else:
if not os.path.isdir(config.save_vocab_path):
os.makedirs(config.save_vocab_path)
tag_vocab = FactWordTagVocab([kd_tra_data, ca_tra_data])
pickle.dump(tag_vocab, open(config.fact_word_tag_vocab, 'wb'))
bert_model = MyBertModel(config)
if config.use_cuda:
bert_model = bert_model.cuda()
# train
train(bert_model, kd_tra_data, ca_tra_data, kd_dev_data, ca_dev_data,
tag_vocab, config, kd_tra_data_gold, ca_tra_data_gold,
kd_dev_data_gold, ca_dev_data_gold, tokenizer)