def __init__(self, opt):
super(DPGANInstructor, self).__init__(opt)
# generator, discriminator
self.gen = DPGAN_G(cfg.gen_embed_dim, cfg.gen_hidden_dim, cfg.vocab_size, cfg.max_seq_len,
cfg.padding_idx, gpu=cfg.CUDA)
self.dis = DPGAN_D(cfg.gen_embed_dim, cfg.gen_hidden_dim, cfg.vocab_size, cfg.max_seq_len,
cfg.padding_idx, gpu=cfg.CUDA)
self.init_model()
# Optimizer
self.gen_opt = optim.Adam(self.gen.parameters(), lr=cfg.gen_lr)
self.gen_adv_opt = optim.Adam(self.gen.parameters(), lr=cfg.gen_lr)
self.dis_opt = optim.Adam(self.dis.parameters(), lr=cfg.dis_lr)
class DPGANInstructor(BasicInstructor):
def __init__(self, opt):
super(DPGANInstructor, self).__init__(opt)
# generator, discriminator
self.gen = DPGAN_G(cfg.gen_embed_dim, cfg.gen_hidden_dim, cfg.vocab_size, cfg.max_seq_len,
cfg.padding_idx, gpu=cfg.CUDA)
self.dis = DPGAN_D(cfg.gen_embed_dim, cfg.gen_hidden_dim, cfg.vocab_size, cfg.max_seq_len,
cfg.padding_idx, gpu=cfg.CUDA)
self.init_model()
# Optimizer
self.gen_opt = optim.Adam(self.gen.parameters(), lr=cfg.gen_lr)
self.gen_adv_opt = optim.Adam(self.gen.parameters(), lr=cfg.gen_lr)
self.dis_opt = optim.Adam(self.dis.parameters(), lr=cfg.dis_lr)
def _run(self):
# ===PRE-TRAINING===
# TRAIN GENERATOR
if not cfg.gen_pretrain:
self.log.info('Starting Generator MLE Training...')
self.pretrain_generator(cfg.MLE_train_epoch)
if cfg.if_save and not cfg.if_test:
torch.save(self.gen.state_dict(), cfg.pretrained_gen_path)
print('Save pre-trained generator: {}'.format(cfg.pretrained_gen_path))
# # ===TRAIN DISCRIMINATOR====
if not cfg.dis_pretrain:
self.log.info('Starting Discriminator Training...')
self.train_discriminator(cfg.d_step, cfg.d_epoch, 'MLE')
if cfg.if_save and not cfg.if_test:
torch.save(self.dis.state_dict(), cfg.pretrained_dis_path)
print('Save pre-trained discriminator: {}'.format(cfg.pretrained_dis_path))
# ===ADVERSARIAL TRAINING===
self.log.info('Starting Adversarial Training...')
self.log.info('Initial generator: %s' % (self.cal_metrics(fmt_str=True)))
for adv_epoch in range(cfg.ADV_train_epoch):
self.log.info('-----\nADV EPOCH %d\n-----' % adv_epoch)
self.sig.update()
if self.sig.adv_sig:
self.adv_train_generator(cfg.ADV_g_step) # Generator
self.train_discriminator(cfg.ADV_d_step, cfg.ADV_d_epoch, 'ADV') # Discriminator
if adv_epoch % cfg.adv_log_step == 0 or adv_epoch == cfg.ADV_train_epoch - 1:
if cfg.if_save and not cfg.if_test:
self._save('ADV', adv_epoch)
else:
self.log.info('>>> Stop by adv_signal! Finishing adversarial training...')
break
def _test(self):
print('>>> Begin test...')
self._run()
pass
def pretrain_generator(self, epochs):
"""
Max Likelihood Pre-training for the generator
"""
for epoch in range(epochs):
self.sig.update()
if self.sig.pre_sig:
pre_loss = self.train_gen_epoch(self.gen, self.oracle_data.loader, self.mle_criterion, self.gen_opt)
# ===Test===
if epoch % cfg.pre_log_step == 0 or epoch == epochs - 1:
self.log.info(
'[MLE-GEN] epoch %d : pre_loss = %.4f, %s' % (epoch, pre_loss, self.cal_metrics(fmt_str=True)))
if cfg.if_save and not cfg.if_test:
self._save('MLE', epoch)
else:
self.log.info('>>> Stop by pre signal, skip to adversarial training...')
break
def adv_train_generator(self, g_step):
"""
The gen is trained using policy gradients, using the reward from the discriminator.
Training is done for num_batches batches.
"""
discount_rate = 1
total_g_loss = 0
dis_count_list = [discount_rate ** i for i in range(cfg.max_seq_len)]
dis_count_matrix = torch.Tensor(dis_count_list).unsqueeze(0).repeat(cfg.batch_size, 1)
if cfg.CUDA:
dis_count_matrix = dis_count_matrix.cuda()
for step in range(g_step):
inp = self.oracle_data.random_batch()['input']
if cfg.CUDA:
inp = inp.cuda()
gen_sample, gen_sample_log_prob = self.gen.sample_teacher_forcing(inp)
word_reward, sentence_reward = self.dis.getReward(gen_sample)
sentence_reward = sentence_reward.repeat(1, cfg.max_seq_len)
reward_matrix = sentence_reward * word_reward * dis_count_matrix
for i in range(cfg.max_seq_len):
reward_matrix[:, i] = reward_matrix[:, i:].sum(dim=-1)
adv_loss = torch.sum(gen_sample_log_prob * reward_matrix)
self.optimize(self.gen_adv_opt, adv_loss, self.gen)
total_g_loss += adv_loss.item()
# ===Test===
self.log.info(
'[ADV-GEN]: g_loss = %.4f, %s' % (total_g_loss / (g_step * cfg.batch_size), self.cal_metrics(fmt_str=True)))
def train_discriminator(self, d_step, d_epoch, phase='MLE'):
"""
Training the discriminator on real_data_samples (positive) and generated samples from gen (negative).
Samples are drawn d_step times, and the discriminator is trained for d_epoch d_epoch.
"""
# prepare loader for validate
global d_loss, train_acc
pos_val = self.oracle.sample(8 * cfg.batch_size, 4 * cfg.batch_size)
neg_val = self.gen.sample(8 * cfg.batch_size, 4 * cfg.batch_size)
for step in range(d_step):
# prepare loader for training
pos_samples = self.oracle_samples # not re-sample the Oracle data
neg_samples = self.gen.sample(pos_samples.size(0), 4 * cfg.batch_size)
for epoch in range(d_epoch):
# ===Train===
self.train_dis_epoch(self.dis, pos_samples, neg_samples, self.dis_opt)
# ===Test===
pos_reward, neg_reward = self.eval_dis(self.dis, pos_val, neg_val)
self.log.info('[%s-DIS] d_step %d: pos_reward = %.4f, neg_reward = %.4f,' % (
phase, step, pos_reward.item(), neg_reward.item()))
if cfg.if_save and not cfg.if_test:
torch.save(self.dis.state_dict(), cfg.pretrained_dis_path)
def eval_dis(self, model, pos_val, neg_val):
_, pos_reward = model.getReward(pos_val)
_, neg_reward = model.getReward(neg_val)
return torch.mean(pos_reward), torch.mean(neg_reward)
def train_dis_epoch(self, model, pos_samples, neg_samples, optimizer):
num_samples = pos_samples.size(0)
num_batch = num_samples // cfg.batch_size
for i in range(num_batch):
pos_sample = pos_samples[i * cfg.batch_size: (i + 1) * cfg.batch_size]
neg_sample = neg_samples[i * cfg.batch_size: (i + 1) * cfg.batch_size]
_, pos_reward = model.getReward(pos_sample)
_, neg_reward = model.getReward(neg_sample)
loss = -torch.mean(pos_reward) + torch.mean(neg_reward)
self.optimize(optimizer, loss, model)