def load_models(load_path):
model_args = json.load(open(os.path.join(load_path, 'options.json'), 'r'))
vars(args).update(model_args)
autoencoder = Seq2Seq(emsize=args.emsize,
nhidden=args.nhidden,
ntokens=args.ntokens,
nlayers=args.nlayers,
noise_r=args.noise_r,
hidden_init=args.hidden_init,
dropout=args.dropout,
gpu=args.cuda)
gan_gen = MLP_G(ninput=args.z_size,
noutput=args.nhidden,
layers=args.arch_g)
gan_disc = MLP_D(ninput=args.nhidden, noutput=1, layers=args.arch_d)
autoencoder = autoencoder.cuda()
gan_gen = gan_gen.cuda()
gan_disc = gan_disc.cuda()
word2idx = json.load(open(os.path.join(args.save, 'vocab.json'), 'r'))
idx2word = {v: k for k, v in word2idx.items()}
print('Loading models from {}'.format(args.save))
loaded = torch.load(os.path.join(args.save, "model.pt"))
autoencoder.load_state_dict(loaded.get('ae'))
gan_gen.load_state_dict(loaded.get('gan_g'))
gan_disc.load_state_dict(loaded.get('gan_d'))
return model_args, idx2word, autoencoder, gan_gen, gan_disc
print(gan_gen)
print(gan_disc)
print(classifier)
optimizer_ae = optim.SGD(autoencoder.parameters(), lr=args.lr_ae)
optimizer_gan_g = optim.Adam(gan_gen.parameters(), lr=args.lr_gan_g, betas=(args.beta1, 0.999))
optimizer_gan_d = optim.Adam(gan_disc.parameters(), lr=args.lr_gan_d, betas=(args.beta1, 0.999))
# classify
optimizer_classify = optim.Adam(classifier.parameters(), lr=args.lr_classify, betas=(args.beta1, 0.999))
criterion_ce = nn.CrossEntropyLoss()
if args.cuda:
autoencoder = autoencoder.cuda()
gan_gen = gan_gen.cuda()
gan_disc = gan_disc.cuda()
classifier = classifier.cuda()
criterion_ce = criterion_ce.cuda()
# In[ ]:
def save_model():
print("Saving models")
with open('{}/autoencoder_model.pt'.format(args.outf), 'wb') as f:
torch.save(autoencoder.state_dict(), f)
with open('{}/gan_gen_model.pt'.format(args.outf), 'wb') as f:
torch.save(gan_gen.state_dict(), f)
with open('{}/gan_disc_model.pt'.format(args.outf), 'wb') as f:
torch.save(gan_disc.state_dict(), f)
errG_l = []
errD_real_l = []
errD_fake_l = []
input = torch.FloatTensor(batchSize, 1, s_sample, s_sample, s_sample)
noise = torch.FloatTensor(batchSize, nz, 1, 1, 1,
device=device).normal_(0, 1)
fixed_noise = torch.FloatTensor(batchSize, nz, 1, 1, 1).normal_(0, 1)
one = torch.FloatTensor([1])
#one = torch.tensor(1, dtype=torch.float)
mone = one * -1
#torch.cuda.empty_cache()
if cuda == True:
netD.cuda()
netG.cuda()
input = input.cuda()
one, mone = one.cuda(), mone.cuda()
noise, fixed_noise = noise.cuda(), fixed_noise.cuda()
optimizerD = optim.Adam(netD.parameters(),
lr=opt.lrD,
betas=(opt.beta1, opt.beta2))
optimizerG = optim.Adam(netG.parameters(),
lr=opt.lrG,
betas=(opt.beta1, opt.beta2))
#experiments/ch256_lr0005/optG_41.pth
if opt.load_opt == True:
def main():
state_dict = torch.load(args.ae_model)
with open(args.ae_args) as f:
ae_args = json.load(f)
corpus = Corpus(args.data_file,
args.dict_file,
vocab_size=ae_args['vocab_size'])
autoencoder = Seq2Seq(emsize=ae_args['emsize'],
nhidden=ae_args['nhidden'],
ntokens=ae_args['ntokens'],
nlayers=ae_args['nlayers'],
noise_radius=ae_args['noise_radius'],
hidden_init=ae_args['hidden_init'],
dropout=ae_args['dropout'],
gpu=args.cuda)
autoencoder.load_state_dict(state_dict)
for param in autoencoder.parameters():
param.requires_grad = False
# save arguments
with open(os.path.join(out_dir, 'args.json'), 'w') as f:
json.dump(vars(args), f)
log.info('[Data and AE model loaded.]')
gan_gen = MLP_G(ninput=args.nhidden,
noutput=args.nhidden,
layers=args.arch_g)
gan_disc = MLP_D(ninput=2 * args.nhidden, noutput=1, layers=args.arch_d)
optimizer_gan_g = optim.Adam(gan_gen.parameters(),
lr=args.lr_gan_g,
betas=(args.beta1, 0.999))
optimizer_gan_d = optim.Adam(gan_disc.parameters(),
lr=args.lr_gan_d,
betas=(args.beta1, 0.999))
criterion_ce = nn.CrossEntropyLoss()
if args.cuda:
autoencoder = autoencoder.cuda()
gan_gen = gan_gen.cuda()
gan_disc = gan_disc.cuda()
criterion_ce = criterion_ce.cuda()
one = to_gpu(args.cuda, torch.FloatTensor([1]))
mone = one * -1
train_pairs = BatchGen(corpus.get_chunks(size=2), args.batch_size)
def train_gan_g(batch):
gan_gen.train()
gan_gen.zero_grad()
source, _ = batch
source = to_gpu(args.cuda, Variable(source))
source_hidden = autoencoder(source, noise=False, encode_only=True)
fake_hidden = gan_gen(source_hidden)
errG = gan_disc(source_hidden, fake_hidden)
# loss / backprop
errG.backward(one)
optimizer_gan_g.step()
return errG
def train_gan_d(batch):
# clamp parameters to a cube
for p in gan_disc.parameters():
p.data.clamp_(-args.gan_clamp, args.gan_clamp)
gan_disc.train()
gan_disc.zero_grad()
# positive samples ----------------------------
# generate real codes
source, target = batch
source = to_gpu(args.cuda, Variable(source))
target = to_gpu(args.cuda, Variable(target))
# batch_size x nhidden
source_hidden = autoencoder(source, noise=False, encode_only=True)
target_hidden = autoencoder(target, noise=False, encode_only=True)
# loss / backprop
errD_real = gan_disc(source_hidden, target_hidden)
errD_real.backward(one)
# negative samples ----------------------------
# loss / backprop
fake_hidden = gan_gen(source_hidden)
errD_fake = gan_disc(source_hidden.detach(), fake_hidden.detach())
errD_fake.backward(mone)
optimizer_gan_d.step()
errD = -(errD_real - errD_fake)
return errD, errD_real, errD_fake
niter = 0
start_time = datetime.now()
for t in range(args.updates):
niter += 1
# train discriminator/critic
for i in range(args.niters_gan_d):
# feed a seen sample within this epoch; good for early training
errD, errD_real, errD_fake = \
train_gan_d(next(train_pairs))
# train generator
for i in range(args.niters_gan_g):
errG = train_gan_g(next(train_pairs))
if niter % args.log_interval == 0:
eta = str((datetime.now() - start_time) / (t + 1) *
(args.updates - t - 1)).split('.')[0]
log.info('[{}/{}] Loss_D: {:.6f} (real: {:.6f} '
'fake: {:.6f}) Loss_G: {:.6f} ETA: {}'.format(
niter, args.updates,
errD.data.cpu()[0],
errD_real.data.cpu()[0],
errD_fake.data.cpu()[0],
errG.data.cpu()[0], eta))
if niter % args.save_interval == 0:
save_model(gan_gen, out_dir, 'gan_gen_model_{}.pt'.format(t))
save_model(gan_disc, out_dir, 'gan_disc_model_{}.pt'.format(t))
print(gan_disc)
optimizer_ae = optim.SGD(autoencoder.parameters(), lr=args.lr_ae)
optimizer_gan_g = optim.Adam(gan_gen.parameters(),
lr=args.lr_gan_g,
betas=(args.beta1, 0.999))
optimizer_gan_d = optim.Adam(gan_disc.parameters(),
lr=args.lr_gan_d,
betas=(args.beta1, 0.999))
criterion_ce = nn.CrossEntropyLoss()
if args.cuda:
autoencoder = autoencoder.cuda()
gan_gen = gan_gen.cuda()
gan_disc = gan_disc.cuda()
criterion_ce = criterion_ce.cuda()
###############################################################################
# Training code
###############################################################################
def save_model():
print("Saving models")
with open('./output/{}/autoencoder_model.pt'.format(args.outf), 'wb') as f:
torch.save(autoencoder.state_dict(), f)
with open('./output/{}/gan_gen_model.pt'.format(args.outf), 'wb') as f:
torch.save(gan_gen.state_dict(), f)
with open('./output/{}/gan_disc_model.pt'.format(args.outf), 'wb') as f:
torch.save(gan_disc.state_dict(), f)
lr=args.lr_gan_d,
betas=(args.beta1, 0.999))
optimizer_G = optim.Adam(G.parameters(),
lr=args.lr_gan_g,
betas=(args.beta1, 0.999))
logger.info(char_ae)
logger.info(word_ae)
logger.info(D)
logger.info(G)
if torch.cuda.is_available():
logger.info("Running on GPU")
char_ae = char_ae.cuda()
word_ae = word_ae.cuda()
D = D.cuda()
G = G.cuda()
else:
logger.info("Running on CPU")
###############################################################################
# Training code
###############################################################################
def validate_disc(data_batches):
# Turn on evaluation mode which disables dropout.
char_ae.eval()
word_ae.eval()
D.eval()
total_correct = 0