def control(input_msg):
tagger = Komoran()
dataset = Dataset('nsmc/ratings.txt',
tagger,
max_length=MAX_LENGTH,
batch_size=BATCH_SIZE)
Z_DIM = 40
H_DIM = 300
C_DIM = 2
model = RNN_VAE(dataset.num_words,
H_DIM,
Z_DIM,
C_DIM,
freeze_embeddings=False,
gpu=USE_CUDA,
gpu_id=GPU_ID)
test_data = torch.LongTensor(
dataset.sentence2idxs(tagger.morphs(input_msg))).unsqueeze(1)
model.load_state_dict(torch.load('models/vae_epoch_300_400.bin'))
results = model.controlSentence(test_data, t=0.5)
return (dataset.idxs2sentence(results[0], no_pad=True),
dataset.idxs2sentence(results[1], no_pad=True))
z_dim = h_dim
c_dim = 2
kl_weight_max = 0.4
# Specific hyperparams
beta = 0.1
lambda_c = 0.1
lambda_z = 0.1
lambda_u = 0.1
dataset = SST_Dataset(mbsize=mbsize)
model = RNN_VAE(dataset.n_vocab,
h_dim,
z_dim,
c_dim,
p_word_dropout=0.3,
pretrained_embeddings=dataset.get_vocab_vectors(),
freeze_embeddings=True,
gpu=args.gpu)
# Load pretrained base VAE with c ~ p(c)
model.load_state_dict(torch.load('models/vae.bin'))
def kl_weight(it):
"""
Credit to: https://github.com/kefirski/pytorch_RVAE/
0 -> 1
"""
return (math.tanh((it - 3500) / 1000) + 1) / 2
mb_size = 32
z_dim = 20
h_dim = 64
lr = 1e-3
lr_decay_every = 1000000
n_iter = 20000
log_interval = 1000
z_dim = h_dim
c_dim = 7
dataset = SST_Dataset()
model = RNN_VAE(dataset.n_vocab,
h_dim,
z_dim,
c_dim,
p_word_dropout=0.3,
pretrained_embeddings=dataset.get_vocab_vectors(),
freeze_embeddings=False,
gpu=args.gpu)
def main():
# Annealing for KL term
kld_start_inc = 3000
kld_weight = 0.01
kld_max = 0.15
kld_inc = (kld_max - kld_weight) / (n_iter - kld_start_inc)
trainer = optim.Adam(model.vae_params, lr=lr)
for it in range(n_iter):
mb_size = 50
z_dim = 20
h_dim = args.dimension
lr = 1e-3
lr_decay_every = 1000000
n_iter = 10000
log_interval = 1000
z_dim = h_dim
c_dim = args.num_classes
torch.manual_seed(int(time.time()))
model = RNN_VAE(
dataset.n_vocab, h_dim, z_dim, c_dim, p_word_dropout=0.3,
pretrained_embeddings=dataset.get_vocab_vectors(),
cnn_filters=args.filters, cnn_units=args.units,
freeze_embeddings=args.freeze_emb, gpu=args.gpu
)
if args.gpu:
model.load_state_dict(torch.load('models/{}.bin'
.format(args.model
+ utils.getModelName(args))))
else:
model.load_state_dict(torch.load('models/{}.bin'
.format(args.model
+ utils.getModelName(args)),
map_location=lambda storage, loc: storage)
)
# Samples latent and conditional codes randomly from prior
TEXT.build_vocab(train, max_size=10000, vectors="fasttext.en.300d")
#TEXT.build_vocab(train, max_size=10000)
LABEL.build_vocab(train)
LABEL.vocab.stoi['1']=0
LABEL.vocab.stoi['2']=0
LABEL.vocab.stoi['3']=0
LABEL.vocab.stoi['4']=1
LABEL.vocab.stoi['5']=1
model = RNN_VAE(
len(TEXT.vocab), h_dim, z_dim, c_dim, p_word_dropout=0.3,
pretrained_embeddings=TEXT.vocab.vectors, freeze_embeddings=False,
gpu=args.gpu
)
if args.gpu:
model.load_state_dict(torch.load('models/{}.bin'.format(args.model)))
else:
model.load_state_dict(torch.load('models/{}.bin'.format(args.model), map_location=lambda storage, loc: storage))
for i in range(3):
h_dim = 64
lr = 1e-3
lr_decay_every = 1000000
n_iter = 20000
log_interval = 1000
z_dim = h_dim
c_dim = 1
dataset = SST_Dataset()
torch.manual_seed(int(time.time()))
model = RNN_VAE(dataset.n_vocab,
h_dim,
z_dim,
c_dim,
p_word_dropout=0.3,
pretrained_embeddings=dataset.get_vocab_vectors(),
freeze_embeddings=True,
gpu=args.gpu)
if args.gpu:
model.load_state_dict(torch.load('{}'.format(args.path)))
else:
model.load_state_dict(
torch.load('{}'.format(args.path),
map_location=lambda storage, loc: storage))
# Samples latent and conditional codes randomly from prior
z = model.sample_z_prior(1)
c = model.sample_c_prior(1)
batch_size=BATCH_SIZE)
# # 3. Train Model
# ## 3.1 Define model
#
# 모델은 GRU를 사용한 Variational Recurrent Autoencoder입니다.
Z_DIM = 40
H_DIM = 300
C_DIM = 2
model = RNN_VAE(dataset.num_words,
H_DIM,
Z_DIM,
C_DIM,
freeze_embeddings=False,
gpu=USE_CUDA,
gpu_id=GPU_ID)
def save_base_vae():
if not os.path.exists('models/'):
os.makedirs('models/')
torch.save(model.state_dict(), 'models/vae_epoch_300_complete.bin')
def save_base_vae_iter(iter):
if not os.path.exists('models/'):
os.makedirs('models/')
log_interval = 100
z_dim = h_dim
c_dim = args.num_classes
kl_weight_max = 0.4
# Specific hyperparams
beta = 0.1
lambda_c = 0.1
lambda_z = 0.1
lambda_u = 0.1
model = RNN_VAE(dataset.n_vocab,
h_dim,
z_dim,
c_dim,
p_word_dropout=0.3,
pretrained_embeddings=dataset.get_vocab_vectors(),
cnn_filters=args.filters,
cnn_units=args.units,
freeze_embeddings=args.freeze_emb,
gpu=args.gpu)
# Load pretrained base VAE with c ~ p(c)
model.load_state_dict(
torch.load('models/vae' + utils.getModelName(args) + '.bin'))
def kl_weight(it):
"""
Credit to: https://github.com/kefirski/pytorch_RVAE/
0 -> 1
"""
