def setup_network(self, n_features):
if self.net_type == 'lstm':
self.network = CharLSTM(self.vocab_size, self.net_hidden,
self.batch_size)
elif self.net_type == 'irnn':
self.network = CharIRNN(self.vocab_size, self.net_hidden,
self.batch_size)
else:
error("Unknown net_type")
self.reset_accum_loss()
n_epochs = args.epochs
n_units = args.rnn_size
batchsize = args.batchsize
bprop_len = args.seq_length
grad_clip = args.grad_clip
train_data, words, vocab = load_data(args)
pickle.dump(vocab, open('%s/vocab.bin' % args.data_dir, 'wb'))
if len(args.init_from) > 0:
model = pickle.load(open(args.init_from, 'rb'))
else:
if args.net == 'lstm':
model = CharLSTM(len(vocab), n_units)
elif args.net == 'irnn':
model = CharIRNN(len(vocab), n_units)
else:
error("unknown net")
state = model.make_initial_state(n_units, batchsize=batchsize)
if args.gpu >= 0:
cuda.init()
model.to_gpu()
if args.optimizer == 'rmsprop':
optimizer = optimizers.RMSprop(lr=args.learning_rate,
alpha=args.decay_rate)
elif args.optimizer == 'adam':
optimizer = optimizers.Adam()
elif args.optimizers == 'adagrad':
