blm_model_seq = SparseSeqLM(blm_model, True, args.lm_droprate, False)
SL_map = {'vanilla':Vanilla_SeqLabel, 'lm-aug': SeqLabel}
seq_model = SL_map[args.seq_model](flm_model_seq, blm_model_seq, len(c_map), args.seq_c_dim, args.seq_c_hid, args.seq_c_layer, len(gw_map), args.seq_w_dim, args.seq_w_hid, args.seq_w_layer, len(y_map), args.seq_droprate, unit=args.seq_rnn_unit)
pw.info('Loading pre-trained models from {}.'.format(args.load_seq))
seq_file = wrapper.restore_checkpoint(args.load_seq)['model']
seq_model.load_state_dict(seq_file)
seq_model.to(device)
crit = CRFLoss(y_map)
decoder = CRFDecode(y_map)
evaluator = eval_wc(decoder, 'f1')
pw.info('Constructing dataset.')
train_dataset, test_dataset, dev_dataset = [SeqDataset(tup_data, flm_map['\n'], blm_map['\n'], gw_map['<\n>'], c_map[' '], c_map['\n'], y_map['<s>'], y_map['<eof>'], len(y_map), args.batch_size) for tup_data in [train_data, test_data, dev_data]]
pw.info('Constructing optimizer.')
param_dict = filter(lambda t: t.requires_grad, seq_model.parameters())
optim_map = {'Adam' : optim.Adam, 'Adagrad': optim.Adagrad, 'Adadelta': optim.Adadelta, 'SGD': functools.partial(optim.SGD, momentum=0.9)}
if args.lr > 0:
optimizer=optim_map[args.update](param_dict, lr=args.lr)
else:
optimizer=optim_map[args.update](param_dict)
pw.info('Saving configues.')
pw.save_configue(args)
pw.info('Setting up training environ.')
args.seq_w_layer,
len(y_map),
args.seq_droprate,
unit=args.seq_rnn_unit)
seq_model.rand_init()
seq_model.load_pretrained_word_embedding(torch.FloatTensor(emb_array))
seq_config = seq_model.to_params()
seq_model.to(device)
crit = CRFLoss(y_map)
decoder = CRFDecode(y_map)
evaluator = eval_wc(decoder, args.eval_type)
pw.info('Constructing dataset')
train_dataset, test_dataset, dev_dataset = [
SeqDataset(tup_data, gw_map['<\n>'], c_map[' '], c_map['\n'],
y_map['<s>'], y_map['<eof>'], len(y_map), args.batch_size)
for tup_data in [train_data, test_data, dev_data]
]
pw.info('Constructing optimizer')
param_dict = filter(lambda t: t.requires_grad, seq_model.parameters())
optim_map = {
'Adam': optim.Adam,
'Adagrad': optim.Adagrad,
'Adadelta': optim.Adadelta,
'SGD': functools.partial(optim.SGD, momentum=0.9)
}
if args.lr > 0:
optimizer = optim_map[args.update](param_dict, lr=args.lr)
else:
param.requires_grad = False
print("set ", name, "as False")
#print(name)
seq_config = seq_model.to_params()
seq_model.to(device)
crit = CRFLoss(y_map)
decoder = CRFDecode(y_map)
evaluator = eval_wc(decoder, args.eval_type, inv_w_map, inv_y_map)
models = [seq_model]
pw.info('Constructing dataset')
train_dataset = SeqDataset(train_data, gw_map['<\n>'], c_map[' '],
c_map['\n'], y_map['<s>'], y_map['<eof>'],
len(y_map), args.batch_size)
test_dataset, dev_dataset = [
SeqDataset(tup_data, gw_map['<\n>'], c_map[' '], c_map['\n'],
y_map['<s>'], y_map['<eof>'], len(y_map), args.batch_size)
for tup_data in [test_data, dev_data]
]
pw.info('Constructing optimizer')
#param_dict = filter(lambda t: t.requires_grad, seq_model.parameters())
#params = list(seq_model.parameters()) + list(crit.parameters())
params = list(seq_model.parameters())
param_dict = filter(lambda t: t.requires_grad, params)
optim_map = {
gw_map['<\n>'],
c_map[' '],
c_map['\n'],
y_map['<s>'],
y_map['<eof>'],
len(y_map),
args.batch_size,
y_unk=y_map['<unk>'],
y_O=y_map['O'],
if_shuffle=False)
test_dataset, dev_dataset = [
SeqDataset(tup_data,
gw_map['<\n>'],
c_map[' '],
c_map['\n'],
y_map['<s>'],
y_map['<eof>'],
len(y_map),
args.batch_size,
if_shuffle=False) for tup_data in [test_data, dev_data]
]
pw.info('Constructing optimizer')
#param_dict = filter(lambda t: t.requires_grad, seq_model.parameters())
#params = list(seq_model.parameters()) + list(crit.parameters())
params = list(seq_model.parameters())
param_dict = filter(lambda t: t.requires_grad, params)
optim_map = {
'Adam': optim.Adam,