def load_model(path, args=None):
if not args:
if not os.path.exists(path):
raise Exception("Model " + path + " does not exist")
with open(path + "/args", "r") as f:
args = pickle.load(f)
model = dynet.Model()
RNNModel = rnnlm.get_lm(args.model)
vocab = load_vocab(args)
if args.s2s:
print "loading s2s..."
s2s = seq2seq.get_s2s(args.s2s_type).load(model, args.s2s)
pron_dict = util.PronDict(model, s2s)
#print "getting prons for train data"
#pron_dict.add_prons(train_data)
#print "getting prons for valid data"
#pron_dict.add_prons(valid_data)
lm = RNNModel(model, vocab, pron_dict, s2s, args)
else:
lm = RNNModel(model, vocab, args)
if not args:
lm.m.load(path + "/params")
else:
lm.m.load(path)
return lm, vocab
def __init__(self, model, vocab, args):
self.m = model
self.vocab = vocab
self.args = args
if args.s2s:
print "loading s2s..."
self.s2s = seq2seq.get_s2s(args.s2s_type).load(model, args.s2s)
self.rnn = args.rnn(args.layers, self.s2s.args.hidden_dim * 2,
args.hidden_dim, model)
self.R = model.add_parameters((vocab.size, args.hidden_dim))
self.bias = model.add_parameters((vocab.size, ))
## model-specific parameters
parser.add_argument("--beam_size", default=3, type=int)
args = parser.parse_args()
print "ARGS:", args
BEGIN_TOKEN = '<s>'
END_TOKEN = '<e>'
# define model
model = dynet.Model()
sgd = dynet.SimpleSGDTrainer(model)
S2SModel = seq2seq.get_s2s(args.model)
if args.load:
print "Loading model..."
s2s = S2SModel.load(model, args.load)
src_vocab = s2s.src_vocab
tgt_vocab = s2s.tgt_vocab
else:
print "fresh model. getting vocab...",
src_reader = util.get_reader(args.reader_mode)(args.train,
mode=args.reader_mode,
begin=BEGIN_TOKEN,
end=END_TOKEN)
src_vocab = util.Vocab.load_from_corpus(src_reader,
remake=args.rebuild_vocab,
src_or_tgt="src")
src_vocab.START_TOK = src_vocab[BEGIN_TOKEN]
else:
if args.percent_valid > 1: cutoff = args.percent_valid
else: cutoff = int(len(train_data)*(args.percent_valid))
valid_data = train_data[-cutoff:]
train_data = train_data[:-cutoff]
#valid_data = train_data[-int(len(train_data)*(args.percent_valid+args.percent_test)):-int(len(train_data)*args.percent_test)]
#test_data = train_data[-int(len(train_data)*args.percent_test):]
#train_data = train_data[:-int(len(train_data)*(args.percent_valid+args.percent_test))]
print "finished loading data"
RNNModel = rnnlm.get_lm(args.model)
if args.s2s:
print "loading s2s..."
s2s = seq2seq.get_s2s(args.s2s_type).load(model, args.s2s)
pron_dict = util.PronDict(model, s2s)
print "getting prons for train data"
pron_dict.add_prons(train_data)
print "getting prons for valid data"
pron_dict.add_prons(valid_data)
lm = RNNModel(model, vocab, pron_dict, s2s, args)
else:
lm = RNNModel(model, vocab, args)
if args.load:
print "loading params..."
model.load(args.load)