def configure_net(net_type, vocab_size, emb_dim, conv_hidden,
lstm_hidden, lstm_layer, bidirectional, use_bert,
bert_type, bert_cache, tokenizer_cache, cuda, aux_device, fix_bert):
assert net_type in ['ff', 'rnn']
net_args = {}
net_args['conv_hidden'] = conv_hidden
net_args['lstm_hidden'] = lstm_hidden
net_args['lstm_layer'] = lstm_layer
net_args['bidirectional'] = bidirectional
if not use_bert:
net_args['vocab_size'] = vocab_size
net_args['emb_dim'] = emb_dim
net = (ExtractSumm(**net_args) if net_type == 'ff'
else PtrExtractSumm(**net_args))
if cuda:
net = net.cuda()
else:
# bert config
net_args['bert_type'] = bert_type
net_args['bert_cache'] = bert_cache
net_args['tokenizer_cache'] = tokenizer_cache
net_args['fix_bert'] = fix_bert
# add aux cuda
added_net_args = dict(net_args)
added_net_args['aux_device'] = aux_device
net = BertPtrExtractSumm(**added_net_args)
return net, net_args
def configure_net(net_type,
vocab_size,
emb_dim,
conv_hidden,
lstm_hidden,
lstm_layer,
bidirectional,
prev_ckpt=None):
assert net_type in ['ff', 'rnn', 'trans_rnn']
net_args = {}
net_args['vocab_size'] = vocab_size
net_args['emb_dim'] = emb_dim
net_args['conv_hidden'] = conv_hidden
net_args['lstm_hidden'] = lstm_hidden
net_args['lstm_layer'] = lstm_layer
net_args['bidirectional'] = bidirectional
if net_type == 'ff':
net = ExtractSumm(**net_args)
elif net_type == 'trans_rnn':
net = TransExtractSumm(**net_args)
else:
net = PtrExtractSumm(**net_args)
if prev_ckpt is not None:
ext_ckpt = load_best_ckpt(prev_ckpt)
net.load_state_dict(ext_ckpt)
return net, net_args
def configure_net(net_type, vocab_size, emb_dim, conv_hidden,
lstm_hidden, lstm_layer, bidirectional):
assert net_type in ['ff', 'rnn']
net_args = {}
net_args['vocab_size'] = vocab_size
net_args['emb_dim'] = emb_dim
net_args['conv_hidden'] = conv_hidden
net_args['lstm_hidden'] = lstm_hidden
net_args['lstm_layer'] = lstm_layer
net_args['bidirectional'] = bidirectional
net = (ExtractSumm(**net_args) if net_type == 'ff'
else PtrExtractSumm(**net_args))
return net, net_args
def configure_net(net_type, vocab_size, emb_dim, conv_hidden,
lstm_hidden, lstm_layer, bidirectional):
assert net_type in ['ff', 'rnn']
net_args = {}
net_args['vocab_size'] = vocab_size
net_args['emb_dim'] = emb_dim
net_args['conv_hidden'] = conv_hidden
net_args['lstm_hidden'] = lstm_hidden
net_args['lstm_layer'] = lstm_layer
net_args['bidirectional'] = bidirectional
net_args['dropoute'] = 0.0 # dropout to remove words from embedding layer (0 = no dropout)
net_args['dropout'] = 0.2 # dropout applied to other layers (0 = no dropout)
net_args['wdrop'] = 0.2 # amount of weight dropout to apply to the RNN hidden to hidden matrix
net_args['dropouth'] = 0.2 # dropout for rnn layers (0 = no dropout)
net = (ExtractSumm(**net_args) if net_type == 'ff'
else PtrExtractSumm(**net_args))
return net, net_args
def configure_net(net_type, vocab_size, emb_dim, conv_hidden,
lstm_hidden, lstm_layer, bidirectional, pe, petrainable, stop):
assert net_type in ['ff', 'rnn', 'nnse']
net_args = {}
net_args['vocab_size'] = vocab_size
net_args['emb_dim'] = emb_dim
net_args['conv_hidden'] = conv_hidden
net_args['lstm_hidden'] = lstm_hidden
net_args['lstm_layer'] = lstm_layer
net_args['bidirectional'] = bidirectional
net_args['pe'] = pe # positional encoding
net_args['petrainable'] = petrainable
net_args['stop'] = stop
if net_type in ['ff', 'rnn']:
net = (ExtractSumm(**net_args) if net_type == 'ff'
else PtrExtractSumm(**net_args))
elif net_type == 'nnse':
net = NNSESumm(**net_args)
return net, net_args