def train_discriminator(dis, gen, criterion, optimizer, epochs,
dis_adversarial_train_loss, dis_adversarial_train_acc,
args):
"""
Train discriminator
"""
generate_samples(gen, args.batch_size, args.n_samples, NEGATIVE_FILE)
data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, args.batch_size)
for epoch in range(epochs):
correct = 0
total_loss = 0.
for data, target in data_iter:
if args.cuda:
data, target = data.cuda(), target.cuda()
target = target.contiguous().view(-1)
output = dis(data)
pred = output.data.max(1)[1]
correct += pred.eq(target.data).cpu().sum()
loss = criterion(output, target)
total_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
data_iter.reset()
avg_loss = total_loss / len(data_iter)
acc = correct.item() / data_iter.data_num
print("Epoch {}, train loss: {:.5f}, train acc: {:.3f}".format(
epoch, avg_loss, acc))
dis_adversarial_train_loss.append(avg_loss)
dis_adversarial_train_acc.append(acc)
def main():
random.seed(SEED)
np.random.seed(SEED)
# Define Networks
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
discriminator = Discriminator(d_num_class, VOCAB_SIZE, d_emb_dim,
d_filter_sizes, d_num_filters, d_dropout)
target_lstm = TargetLSTM(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
if opt.cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
target_lstm = target_lstm.cuda()
# Generate toy data using target lstm
print('Generating data ...')
generate_samples(target_lstm, BATCH_SIZE, GENERATED_NUM, POSITIVE_FILE)
# Load data from file
gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
# Pretrain Generator using MLE
gen_criterion = nn.NLLLoss(size_average=False)
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
print('Pretrain with MLE ...')
for epoch in range(PRE_EPOCH_NUM):
loss = train_epoch(generator, gen_data_iter, gen_criterion,
gen_optimizer)
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
print('Epoch [%d] True Loss: %f' % (epoch, loss))
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Dsicriminator ...')
for epoch in range(5):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(3):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
print('Epoch [%d], loss: %f' % (epoch, loss))
# Adversarial Training
rollout = Rollout(generator, 0.8)
print('#####################################################')
print('Start Adeversatial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if opt.cuda:
gen_gan_loss = gen_gan_loss.cuda()
gen_criterion = nn.NLLLoss(size_average=False)
if opt.cuda:
gen_criterion = gen_criterion.cuda()
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
for total_batch in range(TOTAL_BATCH):
# Train the generator for one step
for it in range(1):
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# construct the input to the generator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(
torch.cat([zeros, samples.data], dim=1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view((-1, ))
# calculate the reward
rewards = rollout.get_reward(samples, 16, discriminator)
rewards = Variable(torch.Tensor(rewards))
if opt.cuda:
rewards = torch.exp(rewards.cuda()).contiguous().view((-1, ))
prob = generator.forward(inputs)
loss = gen_gan_loss(prob, targets, rewards)
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
print('Batch [%d] True Loss: %f' % (total_batch, loss))
rollout.update_params()
for _ in range(4):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
BATCH_SIZE)
for _ in range(2):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
def train_GAN(conf_data):
"""Training Process for GAN.
Parameters
----------
conf_data: dict
Dictionary containing all parameters and objects.
Returns
-------
conf_data: dict
Dictionary containing all parameters and objects.
"""
seq = conf_data['GAN_model']['seq']
if seq == 1:
pre_epoch_num = conf_data['generator']['pre_epoch_num']
GENERATED_NUM = 10000
EVAL_FILE = 'eval.data'
POSITIVE_FILE = 'real.data'
NEGATIVE_FILE = 'gene.data'
temp = 1 #TODO Determines how many times is the discriminator updated. Take this as a value input
epochs = int(conf_data['GAN_model']['epochs'])
if seq == 0:
dataloader = conf_data['data_learn']
mini_batch_size = int(conf_data['GAN_model']['mini_batch_size'])
data_label = int(conf_data['GAN_model']['data_label'])
cuda = conf_data['cuda']
g_latent_dim = int(conf_data['generator']['latent_dim'])
classes = int(conf_data['GAN_model']['classes'])
w_loss = int(conf_data['GAN_model']['w_loss'])
clip_value = float(conf_data['GAN_model']['clip_value'])
n_critic = int(conf_data['GAN_model']['n_critic'])
lambda_gp = int(conf_data['GAN_model']['lambda_gp'])
log_file = open(conf_data['performance_log'] + "/log.txt", "w+")
#Covert these to parameters of the config data
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
conf_data['Tensor'] = Tensor
LongTensor = torch.cuda.LongTensor if cuda else torch.LongTensor
conf_data['LongTensor'] = LongTensor
FloatTensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
conf_data['FloatTensor'] = FloatTensor
conf_data['epochs'] = epochs
#print ("Just before training")
if seq == 1: #TODO: Change back to 1
target_lstm = TargetLSTM(conf_data['GAN_model']['vocab_size'],
conf_data['generator']['embedding_dim'],
conf_data['generator']['hidden_dim'],
conf_data['cuda'])
if cuda == True:
target_lstm = target_lstm.cuda()
conf_data['target_lstm'] = target_lstm
gen_data_iter = GenDataIter('real.data', mini_batch_size)
generator = conf_data['generator_model']
discriminator = conf_data['discriminator_model']
g_loss_func = conf_data['generator_loss']
d_loss_func = conf_data['discriminator_loss']
optimizer_D = conf_data['discriminator_optimizer']
optimizer_G = conf_data['generator_optimizer']
#print('Pretrain with MLE ...')
for epoch in range(pre_epoch_num): #TODO: Change the range
loss = train_epoch(generator, gen_data_iter, g_loss_func,
optimizer_G, conf_data, 'g')
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
generate_samples(generator, mini_batch_size, GENERATED_NUM,
EVAL_FILE, conf_data)
eval_iter = GenDataIter(EVAL_FILE, mini_batch_size)
loss = eval_epoch(target_lstm, eval_iter, g_loss_func, conf_data)
print('Epoch [%d] True Loss: %f' % (epoch, loss))
dis_criterion = d_loss_func
dis_optimizer = optimizer_D
#TODO: Understand why the below two code line were there ?
# if conf_data['cuda']:
# dis_criterion = dis_criterion.cuda()
#print('Pretrain Dsicriminator ...')
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
mini_batch_size)
for epoch in range(5): #TODO: change back 5
generate_samples(generator, mini_batch_size, GENERATED_NUM,
NEGATIVE_FILE, conf_data)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
mini_batch_size)
for _ in range(3): #TODO: change back 3
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer, conf_data, 'd')
print('Epoch [%d], loss: %f' % (epoch, loss))
conf_data['generator_model'] = generator
conf_data['discriminator_model'] = discriminator
torch.save(conf_data['generator_model'].state_dict(),
conf_data['save_model_path'] + '/Seq/' + 'pre_generator.pt')
torch.save(
conf_data['discriminator_model'].state_dict(),
conf_data['save_model_path'] + '/Seq/' + 'pre_discriminator.pt')
conf_data['rollout'] = Rollout(generator, 0.8)
for epoch in range(epochs):
conf_data['epoch'] = epoch
if seq == 0:
to_iter = dataloader
elif seq == 1: #TODO: Change this back to 1
to_iter = [1]
for i, iterator in enumerate(to_iter):
optimizer_D = conf_data['discriminator_optimizer']
optimizer_G = conf_data['generator_optimizer']
generator = conf_data['generator_model']
discriminator = conf_data['discriminator_model']
g_loss_func = conf_data['generator_loss']
d_loss_func = conf_data['discriminator_loss']
# if aux = 1:
#print ("Reached here --------------> ")
conf_data['iterator'] = i
if seq == 0:
if data_label == 1:
imgs, labels = iterator
else:
imgs = iterator
# Adversarial ground truths
valid = Variable(Tensor(imgs.size(0), 1).fill_(1.0),
requires_grad=False)
conf_data['valid'] = valid
fake = Variable(Tensor(imgs.size(0), 1).fill_(0.0),
requires_grad=False)
conf_data['fake'] = fake
# Configure input
real_imgs = Variable(imgs.type(Tensor))
if data_label == 1:
labels = Variable(labels.type(LongTensor))
# Sample noise as generator input
z = Variable(
Tensor(
np.random.normal(0, 1, (imgs.shape[0], g_latent_dim))))
if classes > 0:
gen_labels = Variable(
LongTensor(np.random.randint(0, classes,
imgs.shape[0])))
conf_data['gen_labels'] = gen_labels
# elif seq == 1: #If yes seqGAN
# # samples = generator.sample(mini_batch_size,conf_data['generator']['sequece_length'])
# # zeros = torch.zeros((mini_batch_size,1)).type(LongTensor)
# # imgs = Variable(torch.cat([zeros,samples.data]),dim=1)[:,:-1].contiguous() #TODO: change imgs to inps all, to make more sense of the code
# # targets = Variable(sample.data).contiguous().view((-1,))
# # rewards = rollout.get_reward(sample,16,discriminator)
# # rewards = Variable(Tensor(rewards))
# # prob = generator.forward(inputs)
# # loss = gen_gan_loss(prob)
# pass
# #optimizer_G
# ---------------------
# Train Discriminator
# ---------------------
optimizer_D.zero_grad()
if seq == 1: #TODO change this back to 1
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
mini_batch_size)
for i in range(
temp
): # TODO: Make this a parameter -> for x updates --> I am read the stored models here as well. Should I reamove this ???
optimizer_D = conf_data['discriminator_optimizer']
optimizer_G = conf_data['generator_optimizer']
generator = conf_data['generator_model']
discriminator = conf_data['discriminator_model']
g_loss_func = conf_data['generator_loss']
d_loss_func = conf_data['discriminator_loss']
if classes <= 0:
#print ("Reached here 2 --------------> ")
if seq == 0:
gen_imgs = generator(z)
# Measure discriminator's ability to classify real from generated samples
#Real images
real_validity = discriminator(real_imgs)
#Fake images
fake_validity = discriminator(gen_imgs.detach())
if seq == 1:
generate_samples(generator, mini_batch_size,
GENERATED_NUM, NEGATIVE_FILE,
conf_data)
dis_data_iter = DisDataIter(POSITIVE_FILE,
NEGATIVE_FILE,
mini_batch_size)
loss = train_epoch(discriminator, dis_data_iter,
d_loss_func, optimizer_D, conf_data,
'd')
conf_data['d_loss'] = loss
#exit()
else:
if seq == 0:
gen_imgs = generator(z, gen_labels)
real_validity = discriminator(real_imgs, labels)
fake_validity = discriminator(gen_imgs.detach(),
labels)
if seq == 0:
conf_data['gen_imgs'] = gen_imgs
if seq == 0:
if w_loss == 0:
real_loss = d_loss_func.loss(real_validity, valid)
fake_loss = d_loss_func.loss(fake_validity, fake)
d_loss = (real_loss + fake_loss) / 2
elif w_loss == 1:
d_loss = -d_loss_func.loss(real_validity,
valid) + d_loss_func.loss(
fake_validity, fake)
if lambda_gp > 0:
conf_data['real_data_sample'] = real_imgs.data
conf_data['fake_data_sample'] = gen_imgs.data
conf_data = compute_gradient_penalty(conf_data)
gradient_penalty = conf_data['gradient_penalty']
d_loss = d_loss + lambda_gp * gradient_penalty
conf_data['d_loss'] = d_loss
d_loss.backward()
optimizer_D.step()
if clip_value > 0:
# Clip weights of discriminator
for p in discriminator.parameters():
p.data.clamp_(-clip_value, clip_value)
# -----------------
# Train Generator
# -----------------
conf_data['generator_model'] = generator
conf_data['discriminator_model'] = discriminator
#Next 4 lines were recently added maybe have to remove this.
conf_data['optimizer_G'] = optimizer_G
conf_data['optimizer_D'] = optimizer_D
conf_data['generator_loss'] = g_loss_func
conf_data['discriminator_loss'] = d_loss_func
if seq == 0:
conf_data['noise'] = z
if n_critic <= 0:
conf_data = training_fucntion_generator(conf_data)
elif n_critic > 0:
# Train the generator every n_critic iterations
if i % n_critic == 0:
conf_data = training_fucntion_generator(conf_data)
#exit()
# print ("------------------ Here (train_GAN.py)")
if seq == 0:
batches_done = epoch * len(dataloader) + i
if batches_done % int(conf_data['sample_interval']) == 0:
if classes <= 0:
# print ("Here")
# print (type(gen_imgs.data[:25]))
# print (gen_imgs.data[:25].shape)
save_image(gen_imgs.data[:25],
conf_data['result_path'] +
'/%d.png' % batches_done,
nrow=5,
normalize=True)
elif classes > 0:
sample_image(10, batches_done, conf_data)
if seq == 0:
log_file.write("[Epoch %d/%d] [D loss: %f] [G loss: %f] \n" %
(epoch, epochs, conf_data['d_loss'].item(),
conf_data['g_loss'].item()))
elif seq == 1:
# print ("Done")
log_file.write(
"[Epoch %d/%d] [D loss: %f] [G loss: %f] \n" %
(epoch, epochs, conf_data['d_loss'], conf_data['g_loss']))
conf_data['generator_model'] = generator
conf_data['discriminator_model'] = discriminator
conf_data['log_file'] = log_file
return conf_data
def main(opt):
cuda = opt.cuda
visualize = opt.visualize
print(f"cuda = {cuda}, visualize = {opt.visualize}")
if visualize:
if PRE_EPOCH_GEN > 0:
pretrain_G_score_logger = VisdomPlotLogger(
'line', opts={'title': 'Pre-train G Goodness Score'})
if PRE_EPOCH_DIS > 0:
pretrain_D_loss_logger = VisdomPlotLogger(
'line', opts={'title': 'Pre-train D Loss'})
adversarial_G_score_logger = VisdomPlotLogger(
'line',
opts={
'title': f'Adversarial G {GD} Goodness Score',
'Y': '{0, 13}',
'X': '{0, TOTAL_BATCH}'
})
if CHECK_VARIANCE:
G_variance_logger = VisdomPlotLogger(
'line', opts={'title': f'Adversarial G {GD} Variance'})
G_text_logger = VisdomTextLogger(update_type='APPEND')
adversarial_D_loss_logger = VisdomPlotLogger(
'line', opts={'title': 'Adversarial Batch D Loss'})
# Define Networks
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, cuda)
n_gen = Variable(torch.Tensor([get_n_params(generator)]))
use_cuda = False
if cuda:
n_gen = n_gen.cuda()
use_cuda = True
print('Number of parameters in the generator: {}'.format(n_gen))
discriminator = LSTMDiscriminator(d_num_class, VOCAB_SIZE,
d_lstm_hidden_dim, use_cuda)
c_phi_hat = AnnexNetwork(d_num_class, VOCAB_SIZE, d_emb_dim,
c_filter_sizes, c_num_filters, d_dropout,
BATCH_SIZE, g_sequence_len)
if cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
c_phi_hat = c_phi_hat.cuda()
# Generate toy data using target lstm
print('Generating data ...')
# Load data from file
gen_data_iter = DataLoader(POSITIVE_FILE, BATCH_SIZE)
gen_criterion = nn.NLLLoss(size_average=False)
gen_optimizer = optim.Adam(generator.parameters())
if cuda:
gen_criterion = gen_criterion.cuda()
# 预训练Generator
# Pretrain Generator using MLE
pre_train_scores = []
if MLE:
print('Pretrain with MLE ...')
for epoch in range(int(np.ceil(PRE_EPOCH_GEN))):
loss = train_epoch(generator, gen_data_iter, gen_criterion,
gen_optimizer, PRE_EPOCH_GEN, epoch, cuda)
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
samples = generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
EVAL_FILE)
eval_iter = DataLoader(EVAL_FILE, BATCH_SIZE)
generated_string = eval_iter.convert_to_char(samples)
print(generated_string)
eval_score = get_data_goodness_score(generated_string, SPACES)
if SPACES == False:
kl_score = get_data_freq(generated_string)
else:
kl_score = -1
freq_score = get_char_freq(generated_string, SPACES)
pre_train_scores.append(eval_score)
print('Epoch [%d] Generation Score: %f' % (epoch, eval_score))
print('Epoch [%d] KL Score: %f' % (epoch, kl_score))
print('Epoch [{}] Character distribution: {}'.format(
epoch, list(freq_score)))
torch.save(
generator.state_dict(),
f"checkpoints/MLE_space_{SPACES}_length_{SEQ_LEN}_preTrainG_epoch_{epoch}.pth"
)
if visualize:
pretrain_G_score_logger.log(epoch, eval_score)
else:
generator.load_state_dict(torch.load(weights_path))
# Finishing training with MLE
if GD == "MLE":
for epoch in range(3 * int(GENERATED_NUM / BATCH_SIZE)):
loss = train_epoch_batch(generator, gen_data_iter, gen_criterion,
gen_optimizer, PRE_EPOCH_GEN, epoch,
int(GENERATED_NUM / BATCH_SIZE), cuda)
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
samples = generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
EVAL_FILE)
eval_iter = DataLoader(EVAL_FILE, BATCH_SIZE)
generated_string = eval_iter.convert_to_char(samples)
print(generated_string)
eval_score = get_data_goodness_score(generated_string, SPACES)
if SPACES == False:
kl_score = get_data_freq(generated_string)
else:
kl_score = -1
freq_score = get_char_freq(generated_string, SPACES)
pre_train_scores.append(eval_score)
print('Epoch [%d] Generation Score: %f' % (epoch, eval_score))
print('Epoch [%d] KL Score: %f' % (epoch, kl_score))
print('Epoch [{}] Character distribution: {}'.format(
epoch, list(freq_score)))
torch.save(
generator.state_dict(),
f"checkpoints/MLE_space_{SPACES}_length_{SEQ_LEN}_preTrainG_epoch_{epoch}.pth"
)
if visualize:
pretrain_G_score_logger.log(epoch, eval_score)
# 预训练Discriminator
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Discriminator ...')
for epoch in range(PRE_EPOCH_DIS):
samples = generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE,
SEQ_LEN)
for _ in range(PRE_ITER_DIS):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer, 1, 1, cuda)
print('Epoch [%d], loss: %f' % (epoch, loss))
if visualize:
pretrain_D_loss_logger.log(epoch, loss)
# 对抗训练
# Adversarial Training
rollout = Rollout(generator, UPDATE_RATE)
print('#####################################################')
print('Start Adversarial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if cuda:
gen_gan_loss = gen_gan_loss.cuda()
gen_criterion = nn.NLLLoss(size_average=False)
if cuda:
gen_criterion = gen_criterion.cuda()
dis_criterion = nn.NLLLoss(size_average=False)
dis_criterion_bce = nn.BCELoss()
dis_optimizer = optim.Adam(discriminator.parameters())
if cuda:
dis_criterion = dis_criterion.cuda()
c_phi_hat_loss = VarianceLoss()
if cuda:
c_phi_hat_loss = c_phi_hat_loss.cuda()
c_phi_hat_optm = optim.Adam(c_phi_hat.parameters())
gen_scores = pre_train_scores
for total_batch in range(TOTAL_BATCH):
# Train the generator for one step
for it in range(G_STEPS):
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# samples has size (BS, sequence_len)
# Construct the input to the generator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(
torch.cat([zeros, samples.data], dim=1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view((-1, ))
if opt.cuda:
inputs = inputs.cuda()
targets = targets.cuda()
# Calculate the reward
rewards = rollout.get_reward(samples, discriminator, VOCAB_SIZE,
cuda)
rewards = Variable(torch.Tensor(rewards))
if cuda:
rewards = torch.exp(rewards.cuda()).contiguous().view((-1, ))
rewards = torch.exp(rewards)
# rewards has size (BS)
prob = generator.forward(inputs)
# prob has size (BS*sequence_len, VOCAB_SIZE)
# 3.a
theta_prime = g_output_prob(prob)
# theta_prime has size (BS*sequence_len, VOCAB_SIZE)
# 3.e and f
c_phi_z_ori, c_phi_z_tilde_ori = c_phi_out(
GD,
c_phi_hat,
theta_prime,
discriminator,
temperature=DEFAULT_TEMPERATURE,
eta=DEFAULT_ETA,
cuda=cuda)
c_phi_z_ori = torch.exp(c_phi_z_ori)
c_phi_z_tilde_ori = torch.exp(c_phi_z_tilde_ori)
c_phi_z = torch.sum(c_phi_z_ori[:, 1]) / BATCH_SIZE
c_phi_z_tilde = -torch.sum(c_phi_z_tilde_ori[:, 1]) / BATCH_SIZE
if opt.cuda:
c_phi_z = c_phi_z.cuda()
c_phi_z_tilde = c_phi_z_tilde.cuda()
c_phi_hat = c_phi_hat.cuda()
# 3.i
grads = []
first_term_grads = []
# 3.h optimization step
# first, empty the gradient buffers
gen_gan_optm.zero_grad()
# first, re arrange prob
new_prob = prob.view((BATCH_SIZE, g_sequence_len, VOCAB_SIZE))
# 3.g new gradient loss for relax
batch_i_grads_1 = gen_gan_loss.forward_reward_grads(
samples, new_prob, rewards, generator, BATCH_SIZE,
g_sequence_len, VOCAB_SIZE, cuda)
batch_i_grads_2 = gen_gan_loss.forward_reward_grads(
samples, new_prob, c_phi_z_tilde_ori[:, 1], generator,
BATCH_SIZE, g_sequence_len, VOCAB_SIZE, cuda)
# batch_i_grads_1 and batch_i_grads_2 should be of length BATCH SIZE of arrays of all the gradients
# # 3.i
batch_grads = batch_i_grads_1
if GD != "REINFORCE":
for i in range(len(batch_i_grads_1)):
for j in range(len(batch_i_grads_1[i])):
batch_grads[i][j] = torch.add(batch_grads[i][j], (-1) *
batch_i_grads_2[i][j])
# batch_grads should be of length BATCH SIZE
grads.append(batch_grads)
# NOW, TRAIN THE GENERATOR
generator.zero_grad()
for i in range(g_sequence_len):
# 3.g new gradient loss for relax
cond_prob = gen_gan_loss.forward_reward(
i, samples, new_prob, rewards, BATCH_SIZE, g_sequence_len,
VOCAB_SIZE, cuda)
c_term = gen_gan_loss.forward_reward(i, samples, new_prob,
c_phi_z_tilde_ori[:, 1],
BATCH_SIZE,
g_sequence_len,
VOCAB_SIZE, cuda)
if GD != "REINFORCE":
cond_prob = torch.add(cond_prob, (-1) * c_term)
new_prob[:, i, :].backward(cond_prob, retain_graph=True)
# 3.h - still training the generator, with the last two terms of the RELAX equation
if GD != "REINFORCE":
c_phi_z.backward(retain_graph=True)
c_phi_z_tilde.backward(retain_graph=True)
gen_gan_optm.step()
# 3.i
if CHECK_VARIANCE:
# c_phi_z term
partial_grads = []
for j in range(BATCH_SIZE):
generator.zero_grad()
c_phi_z_ori[j, 1].backward(retain_graph=True)
j_grads = []
for p in generator.parameters():
j_grads.append(p.grad.clone())
partial_grads.append(j_grads)
grads.append(partial_grads)
# c_phi_z_tilde term
partial_grads = []
for j in range(BATCH_SIZE):
generator.zero_grad()
c_phi_z_tilde_ori[j, 1].backward(retain_graph=True)
j_grads = []
for p in generator.parameters():
j_grads.append(-1 * p.grad.clone())
partial_grads.append(j_grads)
grads.append(partial_grads)
# Uncomment the below code if you want to check gradients
"""
print('1st contribution to the gradient')
print(grads[0][0][6])
print('2nd contribution to the gradient')
print(grads[1][0][6])
print('3rd contribution to the gradient')
print(grads[2][0][6])
"""
#grads should be of length 3
#grads[0] should be of length BATCH SIZE
# 3.j
all_grads = grads[0]
if GD != "REINFORCE":
for i in range(len(grads[0])):
for j in range(len(grads[0][i])):
all_grads[i][j] = torch.add(
torch.add(all_grads[i][j], grads[1][i][j]),
grads[2][i][j])
# all_grads should be of length BATCH_SIZE
c_phi_hat_optm.zero_grad()
var_loss = c_phi_hat_loss.forward(all_grads, cuda) #/n_gen
true_variance = c_phi_hat_loss.forward_variance(
all_grads, cuda)
var_loss.backward()
c_phi_hat_optm.step()
print(
'Batch [{}] Estimate of the variance of the gradient at step {}: {}'
.format(total_batch, it, true_variance[0]))
if visualize:
G_variance_logger.log((total_batch + it), true_variance[0])
# Evaluate the quality of the Generator outputs
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
samples = generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
EVAL_FILE)
eval_iter = DataLoader(EVAL_FILE, BATCH_SIZE)
generated_string = eval_iter.convert_to_char(samples)
print(generated_string)
eval_score = get_data_goodness_score(generated_string, SPACES)
if SPACES == False:
kl_score = get_data_freq(generated_string)
else:
kl_score = -1
freq_score = get_char_freq(generated_string, SPACES)
gen_scores.append(eval_score)
print('Batch [%d] Generation Score: %f' %
(total_batch, eval_score))
print('Batch [%d] KL Score: %f' % (total_batch, kl_score))
print('Epoch [{}] Character distribution: {}'.format(
total_batch, list(freq_score)))
#Checkpoint & Visualize
if total_batch % 10 == 0 or total_batch == TOTAL_BATCH - 1:
torch.save(
generator.state_dict(),
f'checkpoints/{GD}_G_space_{SPACES}_pretrain_{PRE_EPOCH_GEN}_batch_{total_batch}.pth'
)
if visualize:
[G_text_logger.log(line) for line in generated_string]
adversarial_G_score_logger.log(total_batch, eval_score)
# Train the discriminator
batch_G_loss = 0.0
for b in range(D_EPOCHS):
for data, _ in gen_data_iter:
data = Variable(data)
real_data = convert_to_one_hot(data, VOCAB_SIZE, cuda)
real_target = Variable(torch.ones((data.size(0), 1)))
samples = generator.sample(data.size(0),
g_sequence_len) # bs x seq_len
fake_data = convert_to_one_hot(
samples, VOCAB_SIZE, cuda) # bs x seq_len x vocab_size
fake_target = Variable(torch.zeros((data.size(0), 1)))
if cuda:
real_target = real_target.cuda()
fake_target = fake_target.cuda()
real_data = real_data.cuda()
fake_data = fake_data.cuda()
real_pred = torch.exp(discriminator(real_data)[:, 1])
fake_pred = torch.exp(discriminator(fake_data)[:, 1])
D_real_loss = dis_criterion_bce(real_pred, real_target)
D_fake_loss = dis_criterion_bce(fake_pred, fake_target)
D_loss = D_real_loss + D_fake_loss
dis_optimizer.zero_grad()
D_loss.backward()
dis_optimizer.step()
gen_data_iter.reset()
print('Batch [{}] Discriminator Loss at step and epoch {}: {}'.
format(total_batch, b, D_loss.data[0]))
if visualize:
adversarial_D_loss_logger.log(total_batch, D_loss.data[0])
if not visualize:
plt.plot(gen_scores)
plt.ylim((0, 13))
plt.title('{}_after_{}_epochs_of_pretraining'.format(
GD, PRE_EPOCH_GEN))
plt.show()
def main():
random.seed(SEED)
np.random.seed(SEED)
calc_bleu([1, 10, 12])
exit()
# Build up dataset
s_train, s_test = load_from_big_file('../data/train_data_obama.txt')
# idx_to_word: List of id to word
# word_to_idx: Dictionary mapping word to id
idx_to_word, word_to_idx = fetch_vocab(s_train, s_train, s_test)
# TODO: 1. Prepare data for attention model
# input_seq, target_seq = prepare_data(DATA_GERMAN, DATA_ENGLISH, word_to_idx)
global VOCAB_SIZE
VOCAB_SIZE = len(idx_to_word)
save_vocab(CHECKPOINT_PATH + 'metadata.data', idx_to_word, word_to_idx,
VOCAB_SIZE, g_emb_dim, g_hidden_dim, g_sequence_len)
print('VOCAB SIZE:', VOCAB_SIZE)
# Define Networks
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, g_sequence_len,
BATCH_SIZE, opt.cuda)
discriminator = Discriminator(d_num_class, VOCAB_SIZE, d_emb_dim,
d_filter_sizes, d_num_filters, d_dropout)
target_lstm = TargetLSTM(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
if opt.cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
target_lstm = target_lstm.cuda()
# Generate toy data using target lstm
print('Generating data ...')
generate_real_data('../data/train_data_obama.txt', BATCH_SIZE,
GENERATED_NUM, idx_to_word, word_to_idx, POSITIVE_FILE,
TEST_FILE)
# Create Test data iterator for testing
test_iter = GenDataIter(TEST_FILE, BATCH_SIZE)
# generate_samples(target_lstm, BATCH_SIZE, GENERATED_NUM, POSITIVE_FILE, idx_to_word)
# Load data from file
gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
# Pretrain Generator using MLE
# gen_criterion = nn.NLLLoss(size_average=False)
gen_criterion = nn.CrossEntropyLoss()
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
print('Pretrain with MLE ...')
for epoch in range(PRE_EPOCH_NUM):
loss = train_epoch(generator, gen_data_iter, gen_criterion,
gen_optimizer)
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
print('Training Output')
test_predict(generator, test_iter, idx_to_word, train_mode=True)
sys.stdout.flush()
# TODO: 2. Flags to ensure dimension of model input is handled
# generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
"""
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
print('Iterator Done')
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
print('Epoch [%d] True Loss: %f' % (epoch, loss))
"""
print('OUTPUT AFTER PRE-TRAINING')
test_predict(generator, test_iter, idx_to_word, train_mode=True)
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Discriminator ...')
for epoch in range(3):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(3):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
print('Epoch [%d], loss: %f' % (epoch, loss))
sys.stdout.flush()
# Adversarial Training
rollout = Rollout(generator, 0.8)
print('#####################################################')
print('Start Adversarial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if opt.cuda:
gen_gan_loss = gen_gan_loss.cuda()
gen_criterion = nn.NLLLoss(size_average=False)
if opt.cuda:
gen_criterion = gen_criterion.cuda()
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
real_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
for total_batch in range(TOTAL_BATCH):
## Train the generator for one step
for it in range(1):
if real_iter.idx >= real_iter.data_num:
real_iter.reset()
inputs = real_iter.next()[0]
inputs = inputs.cuda()
samples = generator.sample(BATCH_SIZE, g_sequence_len, inputs)
samples = samples.cpu()
rewards = rollout.get_reward(samples, 16, discriminator)
rewards = Variable(torch.Tensor(rewards))
if opt.cuda:
rewards = torch.exp(rewards.cuda()).contiguous().view((-1, ))
prob = generator.forward(inputs)
mini_batch = prob.shape[0]
prob = torch.reshape(
prob,
(prob.shape[0] * prob.shape[1], -1)) #prob.view(-1, g_emb_dim)
targets = copy.deepcopy(inputs).contiguous().view((-1, ))
loss = gen_gan_loss(prob, targets, rewards)
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
"""
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# construct the input to the genrator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(torch.cat([zeros, samples.data], dim = 1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view((-1,))
print('', inputs.shape, targets.shape)
print(inputs, targets)
# calculate the reward
rewards = rollout.get_reward(samples, 16, discriminator)
rewards = Variable(torch.Tensor(rewards))
if opt.cuda:
rewards = torch.exp(rewards.cuda()).contiguous().view((-1,))
prob = generator.forward(inputs)
mini_batch = prob.shape[0]
prob = torch.reshape(prob, (prob.shape[0] * prob.shape[1], -1)) #prob.view(-1, g_emb_dim)
loss = gen_gan_loss(prob, targets, rewards)
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
"""
print('Batch [%d] True Loss: %f' % (total_batch, loss))
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
# generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
# eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
# loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
if len(prob.shape) > 2:
prob = torch.reshape(prob, (prob.shape[0] * prob.shape[1], -1))
predictions = torch.max(prob, dim=1)[1]
predictions = predictions.view(mini_batch, -1)
for each_sen in list(predictions):
print('Train Output:',
generate_sentence_from_id(idx_to_word, each_sen))
test_predict(generator, test_iter, idx_to_word, train_mode=True)
torch.save(generator.state_dict(),
CHECKPOINT_PATH + 'generator.model')
torch.save(discriminator.state_dict(),
CHECKPOINT_PATH + 'discriminator.model')
rollout.update_params()
for _ in range(4):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
BATCH_SIZE)
for _ in range(2):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
gen_pretrain_eval_loss.append(gen_loss)
print("eval loss: {:.5f}\n".format(gen_loss))
print('#####################################################\n\n')
# Pre-train discriminator
print('#####################################################')
print('Start pre-training discriminator...')
print('#####################################################\n')
for i in range(args.d_pretrain_steps):
print("D-Step {}".format(i))
train_discriminator(discriminator, generator, nll_loss, dis_optimizer,
args.dk_epochs, dis_adversarial_train_loss,
dis_adversarial_train_acc, args)
generate_samples(generator, args.batch_size, args.n_samples,
NEGATIVE_FILE)
eval_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, args.batch_size)
dis_loss, dis_acc = eval_discriminator(discriminator, eval_iter,
nll_loss, args)
dis_pretrain_eval_loss.append(dis_loss)
dis_pretrain_eval_acc.append(dis_acc)
print("eval loss: {:.5f}, eval acc: {:.3f}\n".format(
dis_loss, dis_acc))
print('#####################################################\n\n')
# Adversarial training
print('#####################################################')
print('Start adversarial training...')
print('#####################################################\n')
rollout = Rollout(generator, args.update_rate)
for i in range(args.rounds):
print("Round {}".format(i))
def main():
random.seed(SEED)
np.random.seed(SEED)
# Define Networks
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
discriminator = Discriminator(d_num_class, VOCAB_SIZE, d_emb_dim,
d_filter_sizes, d_num_filters, d_dropout)
if opt.cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
# Generate toy data using target lstm 也就是新建一个data，假装他就是真实数据
print('啊啊')
# Load data from file
# 每个iter输出一个data和一个target，其中data是每个point前填个0，每个target是后面添个0
gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
# Pretrain Generator using MLE
gen_criterion = nn.NLLLoss(
reduction='sum'
) #You may use CrossEntropyLoss instead, if you prefer not to add an extra LogSoftmax layer .
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
print('Pretrain with MLE ...')
for epoch in range(PRE_EPOCH_NUM): #PRE_EPOCH_NUM =120
loss = train_epoch(generator, gen_data_iter, gen_criterion,
gen_optimizer) #使得generator的参数更新，使之适应gen_data_iter
print('Epoch [%d] Model Loss: %f' % (epoch, loss)) # 9月1日
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(reduction='sum')
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Discriminator ...')
for epoch in range(5):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(3):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
print('Epoch [%d], loss: %f' % (epoch, loss))
# Adversarial Training
rollout = Rollout(generator, 0.8)
print('#####################################################')
print('Start Adeversatial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if opt.cuda:
gen_gan_loss = gen_gan_loss.cuda()
gen_criterion = nn.NLLLoss(reduction='sum')
if opt.cuda:
gen_criterion = gen_criterion.cuda()
dis_criterion = nn.NLLLoss(reduction='sum')
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
for total_batch in range(TOTAL_BATCH):
## Train the generator for one step
for it in range(1):
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# construct the input to the genrator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(
torch.cat([zeros, samples.data], dim=1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view(
(-1, )) # 这里我不明白，为什么inputs：（batch_size,seq_len)，而targets是一个序列
# calculate the reward 为什么在rollout里reward不能直接直接由generator来sample
rewards = rollout.get_reward(
samples, 16, discriminator) # rewards:(batch_size,seq_len)
rewards = Variable(torch.Tensor(rewards))
rewards = torch.exp(rewards).contiguous().view(
(-1, )) # 这是因为Discriminator的最后一层是log_Softmax
if opt.cuda:
rewards = rewards.cuda()
prob = generator.forward(inputs)
loss = gen_gan_loss(prob, targets, rewards) #这里是点睛之笔
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step() #这里更新的是rollout里面的ori_model吗？
rollout.update_params() #这里理解了的话，基本没问题了
for _ in range(4):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
BATCH_SIZE)
for _ in range(2):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
print('Adversarial Training %d complete \n' % (total_batch))
print('保存模型genetor')
torch.save(generator.state_dict(), PATH_GPU)
gen_gan_optm.step()
rollout.update_params()
g_save_path = os.path.join(m_save_path, c_cat)
if not os.path.exists(g_save_path):
os.mkdir(g_save_path)
g_save_path = os.path.join(g_save_path, 'generator'+str(total_batch)+'.pkl')
# torch.save(generator.state_dict(), g_save_path)
print('mul_loss ',loss.item())
for p in range(4):
NEGATIVE_FILE1 = NEGATIVE_FILE + '\\' + str(total_batch) + '\\gene'
samples_lenth = generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE1,x_info,x_ids,start_id_list,end_id_list,bank_dict)
NEGATIVE_FILEtxt = NEGATIVE_FILE + '\\' + str(total_batch) + '\\gene.txt'
dis_data_iter = DisDataIter(real_data_id1, NEGATIVE_FILEtxt, BATCH_SIZE)
for q in range(2):
total_loss = 0.
total_words = 0.
n = 0
for (data, target) in dis_data_iter:
n+=1
data = Variable(data)
target = Variable(target)
if opt.cuda:
data, target = data.cuda(), target.cuda()
target = target.contiguous().view(-1)
pred = discriminator.forward(data)
loss = dis_criterion(pred, target) # negative log likelihood loss
total_loss += loss.item()
total_words += data.size(0) * data.size(1)
def main(generator_, discriminator, model_index):
#random.seed(SEED)
#np.random.seed(SEED)
perf_dict = {}
true_loss = []
generator_loss = []
disc_loss = []
nb_batch_per_epoch = int(GENERATED_NUM / BATCH_SIZE)
# Define Networks
generator = copy.deepcopy(generator_)
#target_lstm = TargetLSTM(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
if opt.cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
#target_lstm = target_lstm.cuda()
# Generate toy data using target lstm
#print('Generating data ...')
#generate_samples(target_lstm, BATCH_SIZE, GENERATED_NUM, POSITIVE_FILE)
# Load data from file
#gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
# Pretrain Generator using MLE
gen_criterion = nn.NLLLoss(reduction='sum')
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
print('Pretrain with MLE ...')
for epoch in range(PRE_EPOCH_NUM):
loss = train_epoch(generator, gen_data_iter, gen_criterion,
gen_optimizer)
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
generator_loss.append(loss)
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
true_loss.append(loss)
print('Epoch [%d] True Loss: %f' % (epoch, loss))
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(reduction='sum')
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Discriminator ...')
for epoch in range(5):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(3):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
disc_loss.append(loss)
print('Epoch [%d], loss: %f' % (epoch, loss))
# Adversarial Training
rollout = Rollout(generator, 0.8)
print('#####################################################')
print('Start Adeversatial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if opt.cuda:
gen_gan_loss = gen_gan_loss.cuda()
# gen_criterion = nn.NLLLoss(reduction='sum')
# if opt.cuda:
# gen_criterion = gen_criterion.cuda()
# dis_criterion = nn.NLLLoss(reduction='sum')
# dis_optimizer = optim.Adam(discriminator.parameters())
# if opt.cuda:
# dis_criterion = dis_criterion.cuda()
for total_batch in range(TOTAL_BATCH):
## Train the generator for one step
for it in range(1):
print(it)
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# construct the input to the genrator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(
torch.cat([zeros, samples.data], dim=1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view((-1, ))
# calculate the reward
rewards = rollout.get_reward(samples, 16, discriminator)
rewards = Variable(torch.Tensor(rewards))
rewards = torch.exp(rewards).contiguous().view((-1, ))
if opt.cuda:
rewards = rewards.cuda()
prob = generator.forward(inputs)
loss = gen_gan_loss(prob, targets, rewards)
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
true_loss.append(loss)
print('Batch [%d] True Loss: %f' % (total_batch, loss))
loss_gen = eval_epoch(generator, gen_data_iter, gen_criterion)
print('Epoch [%d] Model Loss: %f' % (total_batch, loss_gen))
generator_loss.append(loss_gen)
rollout.update_params()
for _ in range(4):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
BATCH_SIZE)
for _ in range(2):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
disc_loss.append(loss)
perf_dict['true_loss'] = true_loss
perf_dict['generator_loss'] = generator_loss
perf_dict['disc_loss'] = disc_loss
np.save('perf_dict' + str(model_index), perf_dict)
def main():
random.seed(SEED)
np.random.seed(SEED)
# Define Networks
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
discriminator = Discriminator(d_num_class, VOCAB_SIZE, d_emb_dim,
d_filter_sizes, d_num_filters, d_dropout)
target_lstm = TargetLSTM(VOCAB_SIZE, g_emb_dim, g_hidden_dim,
opt.cuda) #和Generator函数只有sample方法不一样；
if opt.cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
target_lstm = target_lstm.cuda()
# Generate toy data using target lstm 也就是新建一个data，假装他就是真实数据
print('Generating data ...')
generate_samples(target_lstm, BATCH_SIZE, GENERATED_NUM,
POSITIVE_FILE) #没经过啥子训练，直接随机一个lstm产出9984*20的矩阵
# Load data from file
# 每个iter输出一个data和一个target，其中data是每个point前填个0，每个target是后面添个0
gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
# Pretrain Generator using MLE
gen_criterion = nn.NLLLoss(
reduction='sum'
) #You may use CrossEntropyLoss instead, if you prefer not to add an extra LogSoftmax layer .
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
print('Pretrain with MLE ...')
for epoch in range(PRE_EPOCH_NUM): #PRE_EPOCH_NUM =120
loss = train_epoch(generator, gen_data_iter, gen_criterion,
gen_optimizer) #使得generator的参数更新，使之适应gen_data_iter
print('Epoch [%d] Model Loss: %f' % (epoch, loss)) # 9月1日
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter,
gen_criterion) # generator有向target_lstm靠近吗？
print('Epoch [%d] True Loss: %f' % (epoch, loss))
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(reduction='sum')
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Discriminator ...')
for epoch in range(5):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(3):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
print('Epoch [%d], loss: %f' % (epoch, loss))
# Adversarial Training 前面的预训练的目的是什么？？？？？？
rollout = Rollout(generator, 0.8)
print('#####################################################')
print('Start Adeversatial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if opt.cuda:
gen_gan_loss = gen_gan_loss.cuda()
gen_criterion = nn.NLLLoss(reduction='sum')
if opt.cuda:
gen_criterion = gen_criterion.cuda()
dis_criterion = nn.NLLLoss(reduction='sum')
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
for total_batch in range(TOTAL_BATCH):
## Train the generator for one step
for it in range(1):
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# construct the input to the genrator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(
torch.cat([zeros, samples.data], dim=1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view(
(-1, )) # 这里我不明白，为什么inputs：（batch_size,seq_len)，而targets是一个序列
# calculate the reward
rewards = rollout.get_reward(
samples, 16, discriminator) # rewards:(batch_size,seq_len)
rewards = Variable(torch.Tensor(rewards))
rewards = torch.exp(rewards).contiguous().view((-1, ))
if opt.cuda:
rewards = rewards.cuda()
prob = generator.forward(inputs)
loss = gen_gan_loss(prob, targets, rewards)
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1: #
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
print('Batch [%d] True Loss: %f' % (total_batch, loss))
rollout.update_params() #这里理解了的话，基本没问题了
for _ in range(4):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
BATCH_SIZE)
for _ in range(2):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
def main():
random.seed(SEED)
np.random.seed(SEED)
track_blue = []
# Build up dataset
s_train, s_test = load_from_big_file('obama_speech',g_sequence_len)
# idx_to_word: List of id to word
# word_to_idx: Dictionary mapping word to id
idx_to_word, word_to_idx = fetch_vocab(s_train, s_train, s_test)
# input_seq, target_seq = prepare_data(DATA_GERMAN, DATA_ENGLISH, word_to_idx)
global VOCAB_SIZE
VOCAB_SIZE = len(idx_to_word)
save_vocab(CHECKPOINT_PATH + 'metadata.data', idx_to_word, word_to_idx, VOCAB_SIZE, g_emb_dim, g_hidden_dim)
print('VOCAB SIZE:', VOCAB_SIZE)
# Define Networks
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
discriminator = Discriminator(d_num_class, VOCAB_SIZE, d_emb_dim, d_filter_sizes, d_num_filters, d_dropout)
if opt.cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
# Generate toy data using target lstm
print('Generating data ...')
# Generate samples either from sentences file or lstm
# Sentences file will be structured input sentences
# LSTM based is BOG approach
generate_real_data('obama_speech', BATCH_SIZE, GENERATED_NUM, idx_to_word, word_to_idx,
POSITIVE_FILE, TEST_FILE)
# generate_samples(target_lstm, BATCH_SIZE, GENERATED_NUM, POSITIVE_FILE, idx_to_word)
# generate_samples(target_lstm, BATCH_SIZE, 10, TEST_FILE, idx_to_word)
# Create Test data iterator for testing
test_iter = GenDataIter(TEST_FILE, BATCH_SIZE)
#test_predict(generator, test_iter, idx_to_word, train_mode=True)
# Load data from file
gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
lines = read_file(POSITIVE_FILE)
refrences = []
for line in lines:
phrase = []
for char in line:
phrase.append(idx_to_word[char])
refrences.append(' '.join(phrase))
#refrences.append(phrase)
# Pretrain Generator using MLE
gen_criterion = nn.NLLLoss(size_average=False)
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
print('Pretrain with MLE ...')
for epoch in range(PRE_EPOCH_NUM):
loss = train_epoch(generator, gen_data_iter, gen_criterion, gen_optimizer)
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
sys.stdout.flush()
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
if track_training:
lines = read_file(EVAL_FILE)
hypotheses = []
for line in lines:
phrase = []
for char in line:
phrase.append(idx_to_word[char])
hypotheses.append(' '.join(phrase))
#hypotheses.append(phrase)
bleu_score=get_moses_multi_bleu(hypotheses, refrences, lowercase=True)
track_blue.append(bleu_score)
print(track_blue)
# generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
# eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
# loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
# print('Epoch [%d] True Loss: %f' % (epoch, loss))
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Discriminator ...')
for epoch in range(5):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(3):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion, dis_optimizer)
print('Epoch [%d], loss: %f' % (epoch, loss))
#sys.stdout.flush()
# Adversarial Training
rollout = Rollout(generator, 0.8)
print('#####################################################')
print('Start Adversarial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if opt.cuda:
gen_gan_loss = gen_gan_loss.cuda()
gen_criterion = nn.NLLLoss(size_average=False)
if opt.cuda:
gen_criterion = gen_criterion.cuda()
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
for total_batch in range(TOTAL_BATCH):
## Train the generator for one step
for it in range(1):
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# construct the input to the genrator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(torch.cat([zeros, samples.data], dim=1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view((-1,))
# calculate the reward
rewards = rollout.get_reward(samples, 16, discriminator)
rewards = Variable(torch.Tensor(rewards))
rewards = torch.exp(rewards).contiguous().view((-1,))
if opt.cuda:
rewards = rewards.cuda()
prob = generator.forward(inputs)
# print('SHAPE: ', prob.shape, targets.shape, rewards.shape)
loss = gen_gan_loss(prob, targets, rewards)
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
# print('GEN PRED DIM: ', prob.shape)
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
# generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
# eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
# loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
# print('Batch [%d] True Loss: %f' % (total_batch, loss))
# predictions = torch.max(prob, dim=1)[1]
# predictions = predictions.view(BATCH_SIZE, -1)
# # print('PRED SHAPE:' , predictions.shape)
# for each_sen in list(predictions):
# print('Training Output:', generate_sentence_from_id(idx_to_word, each_sen, DEBUG_FILE))
#
# test_predict(generator, test_iter, idx_to_word, train_mode=True)
loss_gen = eval_epoch(generator, gen_data_iter, gen_criterion)
print('Epoch [%d] Model Loss: %f' % (total_batch, loss_gen))
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
#show_some_generated_sequences(idx_to_word, 10, EVAL_FILE)
sys.stdout.flush()
if track_training:
lines = read_file(EVAL_FILE)
hypotheses = []
for line in lines:
phrase = []
for char in line:
phrase.append(idx_to_word[char])
hypotheses.append(' '.join(phrase))
# hypotheses.append(phrase)
bleu_score = get_moses_multi_bleu(hypotheses, refrences, lowercase=True)
track_blue.append(bleu_score)
print(track_blue)
torch.save(generator.state_dict(), CHECKPOINT_PATH + 'generator_seqgan.model')
torch.save(discriminator.state_dict(), CHECKPOINT_PATH + 'discriminator_seqgan.model')
rollout.update_params()
for _ in range(4):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(2):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion, dis_optimizer)
track_blue = np.array(track_blue)
np.save(ROOT_PATH + 'track_blue_seqgan2.npy', track_blue)
plt.plot(track_blue)
plt.show()
def PG_BLEU(generator_):
def read_file(data_file):
with open(data_file, 'r') as f:
lines = f.readlines()
lis = []
for line in lines:
l = line.strip().split(' ')
l = [int(s) for s in l]
lis.append(l)
return lis
lines = read_file(POSITIVE_FILE)
train_dataset = AmazonReviewGloveDataset('train_small.csv')
vocab = train_dataset.indexer.index_to_word[1::]
del train_dataset
refrences = []
for line in lines:
phrase = []
for char in line:
phrase.append(vocab[char])
#refrences.append(' '.join(phrase))
refrences.append(phrase)
hypotheses = []
for line in lines[:3]:
phrase = []
for char in line:
phrase.append(vocab[char])
#hypotheses.append(' '.join(phrase))
hypotheses.append(phrase)
BLEUscore = nltk.translate.bleu_score.sentence_bleu(
refrences, hypotheses[0])
nb_batch_per_epoch = int(GENERATED_NUM / BATCH_SIZE)
# Define Networks
generator = copy.deepcopy(generator_)
discriminator = Discriminator(d_num_class, VOCAB_SIZE, d_emb_dim,
d_filter_sizes, d_num_filters, d_dropout)
#target_lstm = TargetLSTM(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
if opt.cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
#target_lstm = target_lstm.cuda()
# Generate toy data using target lstm
#print('Generating data ...')
#generate_samples(target_lstm, BATCH_SIZE, GENERATED_NUM, POSITIVE_FILE)
# Load data from file
#gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
# Pretrain Generator using MLE
gen_criterion = nn.NLLLoss(reduction='sum')
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
print('Pretrain with MLE ...')
for epoch in range(40):
loss = train_epoch(generator, gen_data_iter, gen_criterion,
gen_optimizer)
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
generator_loss.append(loss)
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
true_loss.append(loss)
print('Epoch [%d] True Loss: %f' % (epoch, loss))
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(reduction='sum')
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Discriminator ...')
for epoch in range(1):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(1):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
disc_loss.append(loss)
print('Epoch [%d], loss: %f' % (epoch, loss))
# Adversarial Training
rollout = bleu_Rollout(generator, 0.8, refrences, vocab)
print('#####################################################')
print('Start Adeversatial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if opt.cuda:
gen_gan_loss = gen_gan_loss.cuda()
for total_batch in range(TOTAL_BATCH):
## Train the generator for one step
#nb_batch_per_epoch
for it in range(1):
#print(it)
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# construct the input to the genrator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(
torch.cat([zeros, samples.data], dim=1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view((-1, ))
# calculate the reward
rewards = rollout.get_reward(samples, 2)
rewards = Variable(torch.Tensor(rewards))
rewards = torch.exp(rewards).contiguous().view((-1, ))
if opt.cuda:
rewards = rewards.cuda()
prob = generator.forward(inputs)
loss = gen_gan_loss(prob, targets, rewards)
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
true_loss.append(loss)
print('Batch [%d] True Loss: %f' % (total_batch, loss))
loss_gen = eval_epoch(generator, gen_data_iter, gen_criterion)
print('Epoch [%d] Model Loss: %f' % (total_batch, loss_gen))
generator_loss.append(loss_gen)
rollout.update_params()
for _ in range(4):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM,
NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE,
BATCH_SIZE)
for _ in range(2):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion,
dis_optimizer)
disc_loss.append(loss)
perf_dict_pgbleu['true_loss'] = true_loss
perf_dict_pgbleu['generator_loss'] = generator_loss
perf_dict_pgbleu['disc_loss'] = disc_loss
np.save('perf_dict', perf_dict)
def main():
random.seed(SEED)
np.random.seed(SEED)
# Build up dataset
s_train, s_test = load_from_big_file('../data/train_data_obama.txt')
# idx_to_word: List of id to word
# word_to_idx: Dictionary mapping word to id
idx_to_word, word_to_idx = fetch_vocab(s_train, s_train, s_test)
# TODO: 1. Prepare data for attention model
# input_seq, target_seq = prepare_data(DATA_GERMAN, DATA_ENGLISH, word_to_idx)
global VOCAB_SIZE
VOCAB_SIZE = len(idx_to_word)
print('VOCAB SIZE:' , VOCAB_SIZE)
# Define Networks
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, g_sequence_len, BATCH_SIZE, opt.cuda)
discriminator = Discriminator(d_num_class, VOCAB_SIZE, d_emb_dim, d_filter_sizes, d_num_filters, d_dropout)
target_lstm = TargetLSTM(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
if opt.cuda:
generator = generator.cuda()
discriminator = discriminator.cuda()
target_lstm = target_lstm.cuda()
# Generate toy data using target lstm
print('Generating data ...')
generate_samples(target_lstm, BATCH_SIZE, GENERATED_NUM, POSITIVE_FILE, idx_to_word)
# Load data from file
gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
# Pretrain Generator using MLE
gen_criterion = nn.NLLLoss(size_average=False)
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
print('Pretrain with MLE ...')
for epoch in range(PRE_EPOCH_NUM):
loss = train_epoch(generator, gen_data_iter, gen_criterion, gen_optimizer)
print('Epoch [%d] Model Loss: %f'% (epoch, loss))
# TODO: 2. Flags to ensure dimension of model input is handled
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
print('Gen sampled')
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
print('Iterator Done')
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
print('Epoch [%d] True Loss: %f' % (epoch, loss))
# Pretrain Discriminator
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
print('Pretrain Dsicriminator ...')
for epoch in range(3):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(3):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion, dis_optimizer)
print('Epoch [%d], loss: %f' % (epoch, loss))
# Adversarial Training
rollout = Rollout(generator, 0.8)
print('#####################################################')
print('Start Adeversatial Training...\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if opt.cuda:
gen_gan_loss = gen_gan_loss.cuda()
gen_criterion = nn.NLLLoss(size_average=False)
if opt.cuda:
gen_criterion = gen_criterion.cuda()
dis_criterion = nn.NLLLoss(size_average=False)
dis_optimizer = optim.Adam(discriminator.parameters())
if opt.cuda:
dis_criterion = dis_criterion.cuda()
for total_batch in range(TOTAL_BATCH):
## Train the generator for one step
for it in range(1):
samples = generator.sample(BATCH_SIZE, g_sequence_len)
# construct the input to the genrator, add zeros before samples and delete the last column
zeros = torch.zeros((BATCH_SIZE, 1)).type(torch.LongTensor)
if samples.is_cuda:
zeros = zeros.cuda()
inputs = Variable(torch.cat([zeros, samples.data], dim = 1)[:, :-1].contiguous())
targets = Variable(samples.data).contiguous().view((-1,))
# calculate the reward
rewards = rollout.get_reward(samples, 16, discriminator)
rewards = Variable(torch.Tensor(rewards))
if opt.cuda:
rewards = torch.exp(rewards.cuda()).contiguous().view((-1,))
prob = generator.forward(inputs)
prob = torch.reshape(prob, (prob.shape[0] * prob.shape[1], -1)) #prob.view(-1, g_emb_dim)
loss = gen_gan_loss(prob, targets, rewards)
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
print('Batch [%d] True Loss: %f' % (total_batch, loss))
if total_batch % 10 == 0 or total_batch == TOTAL_BATCH - 1:
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, EVAL_FILE)
eval_iter = GenDataIter(EVAL_FILE, BATCH_SIZE)
loss = eval_epoch(target_lstm, eval_iter, gen_criterion)
predictions = torch.max(prob, dim=1)[1]
predictions = predictions.view(BATCH_SIZE, -1)
# print('PRED SHAPE:' , predictions.shape)
for each_sen in list(predictions):
print('Sample Output:', generate_sentence_from_id(idx_to_word, each_sen, DEBUG_FILE))
sys.stdout.flush()
torch.save(generator.state_dict(), './experiment_3_10000/generator.model')
torch.save(discriminator.state_dict(), './experiment_3_10000/discriminator.model')
rollout.update_params()
for _ in range(4):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
dis_data_iter = DisDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(2):
loss = train_epoch(discriminator, dis_data_iter, dis_criterion, dis_optimizer)
