def __init__(self, args, data):
super(B2H, self).__init__()
###Networks
self.word_feature_extractor = args.word_extractor
self.use_char = args.use_char
self.char_feature_extractor = args.char_extractor
self.use_crf = args.use_crf
## Training
self.average_batch_loss = args.average_batch_loss
self.status = args.status
### Hyperparameters
self.cnn_layer = args.cnn_layer
self.iteration = args.iteration
self.batch_size = args.batch_size
self.char_hidden_dim = args.char_hidden_dim
self.hidden_dim = args.hidden_dim
self.dropout = args.dropout
self.lstm_layer = args.lstm_layer
self.bilstm = args.bilstm
self.gpu = args.gpu
self.optimizer = args.optimizer
self.lr = args.lr
self.lr_decay = args.lr_decay
self.clip = args.clip
self.momentum = args.momentum
self.l2 = args.l2
#Dual Network Modules...
self.wordrep = WordRep(args, data)
# information interaction unit
self.inter_unit = B2HInterUnit(
args.word_emb_dim + args.char_hidden_dim, data.llabelset_size, 1,
F.relu)
# component of Dual model
self.H_ner = BiLSTMCRF(args,
input_size=args.char_hidden_dim +
args.word_emb_dim,
tagset_size=data.hlabelset_size,
name='High layer model')
self.B_ner = BiLSTMCRF(args,
input_size=args.char_hidden_dim +
args.word_emb_dim,
tagset_size=data.llabelset_size,
name='bottom layer model')
print("build B2H sequence labeling network...")
class B2H(nn.Module):
def __init__(self, args, data):
super(B2H, self).__init__()
###Networks
self.word_feature_extractor = args.word_extractor
self.use_char = args.use_char
self.char_feature_extractor = args.char_extractor
self.use_crf = args.use_crf
## Training
self.average_batch_loss = args.average_batch_loss
self.status = args.status
### Hyperparameters
self.cnn_layer = args.cnn_layer
self.iteration = args.iteration
self.batch_size = args.batch_size
self.char_hidden_dim = args.char_hidden_dim
self.hidden_dim = args.hidden_dim
self.dropout = args.dropout
self.lstm_layer = args.lstm_layer
self.bilstm = args.bilstm
self.gpu = args.gpu
self.optimizer = args.optimizer
self.lr = args.lr
self.lr_decay = args.lr_decay
self.clip = args.clip
self.momentum = args.momentum
self.l2 = args.l2
#Dual Network Modules...
self.wordrep = WordRep(args, data)
# information interaction unit
self.inter_unit = B2HInterUnit(
args.word_emb_dim + args.char_hidden_dim, data.llabelset_size, 1,
F.relu)
# component of Dual model
self.H_ner = BiLSTMCRF(args,
input_size=args.char_hidden_dim +
args.word_emb_dim,
tagset_size=data.hlabelset_size,
name='High layer model')
self.B_ner = BiLSTMCRF(args,
input_size=args.char_hidden_dim +
args.word_emb_dim,
tagset_size=data.llabelset_size,
name='bottom layer model')
print("build B2H sequence labeling network...")
def calculate_loss(self, word_inputs, word_seq_lengths, char_inputs,
char_seq_lengths, char_seq_recover, batch_hlabel,
batch_llabel, mask):
word_represent = self.wordrep(word_inputs, word_seq_lengths,
char_inputs, char_seq_lengths,
char_seq_recover)
B2HB_outs, B2HB_loss, B2HB_tag_seqs = self.B_ner.calculate_loss(
word_represent, word_seq_lengths, batch_llabel, mask)
low_rep = self.inter_unit(word_represent, B2HB_outs)
# concat high level information to do high level tagging
B2HH_outs, B2HH_loss, B2HH_tag_seqs = self.H_ner.calculate_loss(
low_rep, word_seq_lengths, batch_hlabel, mask)
return B2HB_loss, B2HH_loss, B2HB_tag_seqs, B2HH_tag_seqs
#todo for a while...
def forward(self, word_inputs, word_seq_lengths, char_inputs,
char_seq_lengths, char_seq_recover, mask):
word_represent = self.wordrep(word_inputs, word_seq_lengths,
char_inputs, char_seq_lengths,
char_seq_recover)
B2HB_outs, B2HB_tag_seqs = self.B_ner(word_represent, word_seq_lengths,
mask)
low_rep = self.inter_unit(word_represent, B2HB_outs)
# concat low level information to do low level tagging
# concat high level information to do high level tagging
_, B2HH_tag_seqs = self.H_ner(low_rep, word_seq_lengths, mask)
return B2HB_tag_seqs, B2HH_tag_seqs
def show_model_summary(self, logger):
logger.info(" " + "++" * 20)
logger.info(" Model Network:")
logger.info(" Model use_crf: %s" % (self.use_crf))
logger.info(" Model word extractor: %s" %
(self.word_feature_extractor))
logger.info(" Model use_char: %s" % (self.use_char))
if self.use_char:
logger.info(" Model char extractor: %s" %
(self.char_feature_extractor))
logger.info(" Model char_hidden_dim: %s" %
(self.char_hidden_dim))
logger.info(" " + "++" * 20)
logger.info(" Training:")
logger.info(" Optimizer: %s" % (self.optimizer))
logger.info(" Iteration: %s" % (self.iteration))
logger.info(" BatchSize: %s" % (self.batch_size))
logger.info(" Average batch loss: %s" %
(self.average_batch_loss))
logger.info(" " + "++" * 20)
logger.info(" Hyperparameters:")
logger.info(" Hyper lr: %s" % (self.lr))
logger.info(" Hyper lr_decay: %s" % (self.lr_decay))
logger.info(" Hyper HP_clip: %s" % (self.clip))
logger.info(" Hyper momentum: %s" % (self.momentum))
logger.info(" Hyper l2: %s" % (self.l2))
logger.info(" Hyper hidden_dim: %s" % (self.hidden_dim))
logger.info(" Hyper dropout: %s" % (self.dropout))
logger.info(" Hyper lstm_layer: %s" % (self.lstm_layer))
logger.info(" Hyper bilstm: %s" % (self.bilstm))
logger.info(" Hyper GPU: %s" % (self.gpu))
logger.info("Model SUMMARY END.")
logger.info("++" * 50)
sys.stdout.flush()