def __init__(self,
num_emb,
emb_dim,
hidden_dim,
seq_len,
batch_size,
use_cuda,
test_mode=False):
super(Generator_attention, self).__init__()
# Constants Initialization
self.SOS_Index = 0
self.EOS_Index = 1
self.PAD_Index = 2
# Embeddings
self.emb = nn.Embedding(num_emb, emb_dim)
self.model = transformer.Transformer(self.emb,
self.PAD_Index,
self.emb.num_embeddings,
max_seq_len=max(seq_len, seq_len))
self.test_mode = test_mode
if not test_mode:
self.data_loader = GenDataIter('inshorts_test/real.data',
batch_size)
self.data_loader.reset()
"""
def calc_bleu(self):
data_file = '../seq_gan_with_attention/real.data'
# metadata = load_vocab(self.CHECKPOINT_PATH + '/seq_gan/metadata.data')
# idx_to_word = metadata['idx_to_word']
train_data_iter = GenDataIter(data_file, 1)
reference = self.get_references(train_data_iter)
# pred = self.test_predict(self.seq_gan, train_data_iter, self.seq_gan_metadata)
train_data_iter = GenDataIter(data_file, 1)
candidates_ga = self.test_predict(self.gen_attention, train_data_iter,
self.seq_gan_metadata)
train_data_iter = GenDataIter(data_file, 1)
candidates_sg = self.test_predict(self.seq_gan, train_data_iter,
self.seq_gan_metadata)
# train_data_iter = GenDataIter(data_file, 1)
candidates_ao = []
for sentence in reference:
s = " ".join(sentence)
# s = "kindergarten is great"
sent = test_attention_only.test(s)
candidates_ao.append(sent)
# candidates_sg = self.test_predict(self.seq_gan, train_data_iter, self.seq_gan_metadata)
# CALC BLEU
# references =
# candidates_words = []
# for id in candidates:
# candidates_words.append(idx_to_word[])
# load
# candidatates_id to words
# [0, 12] => [['a', 'aaa']]
sen_score = 0
for sent in candidates_ga:
sen_score += sentence_bleu(reference, sent)
print("@@@@@@@@@@@@@@@@", sen_score) # / len(candidates_ga))
# print('Individual 3-gram: %f' % corpus_bleu(reference, candidates_ga, weights=(1, 0, 0, 0)))
print('Individual 3-gram: %f' % sen_score)
# print('Individual 4-gram: %f' % corpus_bleu(reference, candidates_ga, weights=(1, 0, 0, 0)))
print('Individual 3-gram: %f' %
corpus_bleu(reference, candidates_sg, weights=(1, 0, 0, 0)))
# print('Individual 4-gram: %f' % corpus_bleu(reference, candidates_sg, weights=(1, 0, 0, 0)))
print('Individual 3-gram: %f' %
corpus_bleu(reference, candidates_ao, weights=(1, 0, 0, 0)))
def MLE(epochs, generator_):
generator_loss_mle = []
true_loss_mle = []
generator = copy.deepcopy(generator_)
if opt.cuda:
generator = generator.cuda()
gen_criterion = nn.NLLLoss(reduction='sum')
gen_optimizer = optim.Adam(generator.parameters())
if opt.cuda:
gen_criterion = gen_criterion.cuda()
for epoch in range(epochs):
loss = train_epoch(generator, gen_data_iter, gen_criterion,
gen_optimizer)
print('Epoch [%d] Model Loss: %f' % (epoch, loss))
generator_loss_mle.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_mle.append(loss)
print('Epoch [%d] True Loss: %f' % (epoch, loss))
return true_loss_mle, generator_loss_mle, generator
def demo():
idx_to_word, word_to_idx, VOCAB_SIZE = load_vocab(CHECKPOINT_PATH)
test_iter = GenDataIter(TEST_FILE, BATCH_SIZE)
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
generator = generator.cuda()
generator.load_state_dict(torch.load(CHECKPOINT_PATH + 'generator.model'))
test_predict(generator, test_iter, idx_to_word)
def demo(self, model, metadata, test_sentence):
# idx_to_word, word_to_idx, VOCAB_SIZE = load_vocab(CHECKPOINT_PATH)
self.test_sentence = test_sentence
self.max_test_sentence_len = 10
# Get padded sentencels
self.padded_sentence = pad_sentences(self.test_sentence,
self.max_test_sentence_len)
# print(self.padded_sentence)
# Get ids of padded sentence
padded_sent_ids = get_ids(self.padded_sentence,
metadata['idx_to_word'],
metadata['word_to_idx'],
metadata['vocab_size'])
# print(padded_sent_ids)
# print("&"*80, padded_sent_ids)
# return [["NONE"]]
# Write to temporary file
out_file = self.TEST_FILE
fp = open(out_file, "w")
fp.writelines(["%s " % item for item in padded_sent_ids])
fp.close()
test_iter = GenDataIter(self.TEST_FILE, 1)
return self.test_predict(model, test_iter, metadata)
def demo():
idx_to_word, word_to_idx, VOCAB_SIZE = load_vocab(CHECKPOINT_PATH)
test_iter = GenDataIter(TEST_FILE, 1)
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, g_sequence_len,
BATCH_SIZE, opt.cuda)
generator = generator.cuda()
generator.load_state_dict(
torch.load(CHECKPOINT_PATH + 'generator.model',
map_location={'cuda:1': 'cpu'}))
return test_predict(generator, test_iter, idx_to_word)
def demo():
metadata = load_vocab(CHECKPOINT_PATH)
idx_to_word=metadata['idx_to_word']
word_to_idx=metadata['word_to_idx']
VOCAB_SIZE=metadata['vocab_size']
test_iter = GenDataIter(TEST_FILE, BATCH_SIZE)
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
if opt.cuda:
generator = generator.cuda()
generator.load_state_dict(torch.load(CHECKPOINT_PATH + 'generator_seqgan.model'))
#test_predict(generator, test_iter, idx_to_word)
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE)
show_some_generated_sequences(idx_to_word, 100, NEGATIVE_FILE)
def training_fucntion_generator(conf_data):
"""Training Process for generator network.
Parameters
----------
conf_data: dict
Dictionary containing all parameters and objects.
Returns
-------
conf_data: dict
Dictionary containing all parameters and objects.
"""
PRE_EPOCH_NUM = 2
seq = conf_data['GAN_model']['seq']
BATCH_SIZE = 64
GENERATED_NUM = 10000
EVAL_FILE = 'eval.data'
POSITIVE_FILE = 'real.data'
NEGATIVE_FILE = 'gene.data'
classes = int(conf_data['GAN_model']['classes'])
w_loss = int(conf_data['GAN_model']['w_loss'])
g_loss_func = conf_data['generator_loss']
epoch = conf_data['epoch']
epochs = conf_data['epochs']
generator = conf_data['generator_model']
discriminator = conf_data['discriminator_model']
optimizer_G = conf_data['generator_optimizer']
mini_batch_size = (conf_data['GAN_model']['mini_batch_size'])
optimizer_G.zero_grad()
# Generate a batch of images
if seq == 0:
valid = conf_data['valid']
gen_imgs = conf_data['gen_imgs']
z = conf_data['noise']
if classes <= 0:
#gen_imgs = generator(z)
# Loss measures generator's ability to fool the discriminator
validity = discriminator(gen_imgs)
elif classes > 0:
gen_labels = conf_data['gen_labels']
#gen_imgs = generator(z,gen_labels)
validity = discriminator(gen_imgs, gen_labels)
if w_loss == 1:
g_loss = -g_loss_func.loss(validity,valid)
elif w_loss == 0:
g_loss = g_loss_func.loss(validity,valid)
conf_data['g_loss'] = g_loss
g_loss.backward()
optimizer_G.step()
elif seq == 1:
#print ("Reached Here 3 ---------> ")
gen_gan_loss = GANLoss()
rollout = conf_data['rollout']
target_lstm = conf_data['target_lstm']
for it in range(1):
samples = generator.sample(mini_batch_size, conf_data['generator']['sequece_length'])
# construct the input to the genrator, add zeros before samples and delete the last column
zeros = torch.zeros((mini_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 conf_data['cuda']:
rewards = torch.exp(rewards.cuda()).contiguous().view((-1,))
prob = generator.forward(inputs)
rewards = rewards.contiguous().view(-1,)
loss = gen_gan_loss(prob, targets, rewards)
optimizer_G.zero_grad()
loss.backward()
optimizer_G.step()
#TODO : Change back. Uncomment and indent till line above to rollout
#if total_batch % 1 == 0 or total_batch == TOTAL_BATCH - 1:
generate_samples(generator, mini_batch_size, GENERATED_NUM, EVAL_FILE,conf_data)
#print ("Reached Here 4 ---------> ")
eval_iter = GenDataIter(EVAL_FILE, mini_batch_size)
#print ("Reached Here 5 ---------> ")
loss = eval_epoch(target_lstm, eval_iter, g_loss_func,conf_data)
conf_data['g_loss']= loss
#print ("Reached Here 6 ---------> ")
#print('Batch [%d] True Loss: %f' % (total_batch, loss))
rollout.update_params()
#g_loss = g_loss_func.loss(validity, valid)
# print ("[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]" % (epoch, epochs, conf_data['iterator'], 5,
# conf_data['d_loss'].item(), g_loss.item()))
if seq == 0:
print ("[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]" % (epoch, epochs, conf_data['iterator'], len(conf_data['data_learn']),
conf_data['d_loss'].item(), g_loss.item()))
elif seq == 1:
print("[Epoch %d/%d] [Batch %d] [D loss: %f] [G loss: %f]"% (epoch, epochs, conf_data['iterator'],
conf_data['d_loss'], conf_data['g_loss']))
#print ("Done")
conf_data['generator_model'] = generator
conf_data['generator_optimizer'] = optimizer_G
conf_data['discriminator_model'] = discriminator
conf_data['generator_loss'] = g_loss_func
if seq == 1:
conf_data['rollout'] = rollout
return conf_data
class Generator_attention(nn.Module):
"""Generator """
def __init__(self,
num_emb,
emb_dim,
hidden_dim,
seq_len,
batch_size,
use_cuda,
test_mode=False):
super(Generator_attention, self).__init__()
# Constants Initialization
self.SOS_Index = 0
self.EOS_Index = 1
self.PAD_Index = 2
# Embeddings
self.emb = nn.Embedding(num_emb, emb_dim)
self.model = transformer.Transformer(self.emb,
self.PAD_Index,
self.emb.num_embeddings,
max_seq_len=max(seq_len, seq_len))
self.test_mode = test_mode
if not test_mode:
self.data_loader = GenDataIter('inshorts_test/real.data',
batch_size)
self.data_loader.reset()
"""
self.num_emb = num_emb
self.emb_dim = emb_dim
self.hidden_dim = hidden_dim
self.use_cuda = use_cuda
self.emb = nn.Embedding(num_emb, emb_dim)
self.lstm = nn.LSTM(emb_dim, hidden_dim, batch_first=True)
self.lin = nn.Linear(hidden_dim, num_emb)
self.softmax = nn.LogSoftmax()
self.init_params()
"""
def forward(self, x):
"""
Args:
x: (batch_size, seq_len), sequence of tokens generated by generator
"""
return self.model(x, x)
"""
emb = self.emb(x)
h0, c0 = self.init_hidden(x.size(0))
output, (h, c) = self.lstm(emb, (h0, c0))
pred = self.softmax(self.lin(output.contiguous().view(-1, self.hidden_dim)))
return pred
"""
def step(self, x, h, c):
"""
Args:
x: (batch_size, 1), sequence of tokens generated by generator
h: (1, batch_size, hidden_dim), lstm hidden state
c: (1, batch_size, hidden_dim), lstm cell state
"""
emb = self.emb(x)
output, (h, c) = self.lstm(emb, (h, c))
pred = F.softmax(self.lin(output.view(-1, self.hidden_dim)))
return pred, h, c
def init_hidden(self, batch_size):
h = Variable(torch.zeros((1, batch_size, self.hidden_dim)))
c = Variable(torch.zeros((1, batch_size, self.hidden_dim)))
if self.use_cuda:
h, c = h.cuda(), c.cuda()
return h, c
def init_params(self):
for param in self.parameters():
param.data.uniform_(-0.05, 0.05)
def sample(self, batch_size, seq_len, x=torch.tensor([])):
if self.test_mode:
print('In Test mode')
return None
if self.data_loader.idx >= self.data_loader.data_num:
self.data_loader.reset()
if len(x.shape) > 1:
input_seq = x
else:
input_seq = self.data_loader.next()[0]
input_seq = input_seq.cuda()
sampled_output = transformer.sample_output(self.model, input_seq,
self.EOS_Index,
self.PAD_Index,
input_seq.shape[1])
return sampled_output
"""
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)
dis_adversarial_train_acc = []
dis_adversarial_eval_loss = []
dis_adversarial_eval_acc = []
# Generate toy data using target LSTM
print('#####################################################')
print('Generating data ...')
print('#####################################################\n\n')
generate_samples(target_lstm, args.batch_size, args.n_samples,
POSITIVE_FILE)
# Pre-train generator using MLE
print('#####################################################')
print('Start pre-training generator with MLE...')
print('#####################################################\n')
gen_data_iter = GenDataIter(POSITIVE_FILE, args.batch_size)
for i in range(args.g_pretrain_steps):
print("G-Step {}".format(i))
train_generator_MLE(generator, gen_data_iter, nll_loss, gen_optimizer,
args.gk_epochs, gen_pretrain_train_loss, args)
generate_samples(generator, args.batch_size, args.n_samples,
NEGATIVE_FILE)
eval_iter = GenDataIter(NEGATIVE_FILE, args.batch_size)
gen_loss = eval_generator(target_lstm, eval_iter, nll_loss, args)
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...')
def INIT_TRAIN():
opinion = str(input("expected opinion: e.g. pos neg neu \n"))
REF_SENT = str(input("enter one sentence representing the basic information of generated paragraph \n"))
logging.basicConfig(level=logging.DEBUG, filename='train_init.log')
VOCAB_SIZE, word_index_dict, index_word_dict,g_sequence_len = text_process(TRAIN_LOC)
# Define Networks
generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim)
discriminator_general = DiscriminatorGeneral(d_emb_dim, VOCAB_SIZE, general_D_hidden_dim_top, general_D_hidden_dim_base, d_dropout, BATCH_SIZE,general_D_hidden_dim_out)
discriminator_special = DiscriminatorSpecial(feature_vec_len,hidden_dim_mlp,hidden_dim_out_layer)
if torch.cuda.is_available():
generator = generator.cuda()
discriminator_general = discriminator_general.cuda()
discriminator_special = discriminator_special.cuda()
# Load data from file
gen_data_iter = GenDataIter(POSITIVE_FILE, BATCH_SIZE)
print('Pretrain Generator:')
for epoch in range(PRE_EPOCH_NUM):
time_before = time.time()
loss = pretrain_generator_epoch(generator, gen_data_iter)
time_behind = time.time()
print('Pretrain Generator: Epoch [%d] ; Loss: %f ; Time: %f'% (epoch, loss, time_behind-time_before))
logging.info('Pretrain Generator: Epoch [%d] ; Loss: %f ; Time: %f'% (epoch, loss, time_behind-time_before))
print('Pretrain Discriminator-General:')
for epoch in range(3):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE,index_word_dict,g_sequence_len,"general")
dis_data_iter = DisGenDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(3):
time_before = time.time()
loss = discriminator_gen_epoch(discriminator_general, dis_data_iter)
time_behind = time.time()
print('Pretrain Discriminator-General: Epoch [%d] ; Loss: %f ; Time: %f' % (epoch, loss, time_behind-time_before))
logging.info('Pretrain Discriminator-General: Epoch [%d] ; Loss: %f ; Time: %f' % (epoch, loss, time_behind-time_before))
print('Pretrain Discriminator-Special:')
for epoch in range(5):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE_SPE,index_word_dict,g_sequence_len,"special")
dis_data_iter = DisSpeDataIter(NEGATIVE_FILE_SPE, BATCH_SIZE, REF_SENT, g_sequence_len, opinion, length_control, synthetic_num)
time_before = time.time()
loss = discriminator_spe_epoch(discriminator_special, dis_data_iter)
time_behind = time.time()
print('Pretrain Discriminator-Special: Epoch [%d] ; Loss: %f ; Time: %f' % (epoch, loss, time_behind-time_before))
logging.info('Pretrain Discriminator-Special: Epoch [%d] ; Loss: %f ; Time: %f' % (epoch, loss, time_behind-time_before))
# Adversarial Training
rollout = Rollout(generator, 0.8, theta)
print('#####################################################')
print('Start Adeversatial Training\n')
gen_gan_loss = GANLoss()
gen_gan_optm = optim.Adam(generator.parameters())
if torch.cuda.is_available():
gen_gan_loss = gen_gan_loss.cuda()
gen_criterion = nn.NLLLoss(reduction='sum')
if torch.cuda.is_available():
gen_criterion = gen_criterion.cuda()
(k,distance_left,distance_right) = dis_data_iter.get_para()
for total_batch in range(TOTAL_TRAIN_NUM):
## 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
time_before = time.time()
rewards = rollout.get_reward(samples, 16, discriminator_general, discriminator_special,k,distance_left,distance_right,REF_SENT,index_word_dict)
rewards = Variable(torch.Tensor(rewards))
rewards = torch.exp(rewards).contiguous().view((-1,))
if torch.cuda.is_available():
rewards = rewards.cuda()
prob = generator.forward(inputs)
loss = gen_gan_loss(prob, targets, rewards)
time_behind = time.time()
print("Train Generator: Epoch [%d] ; Loss: %f ; Time: %f"%(total_batch,loss, time_behind-time_before))
logging.info('Train Generator: Epoch [%d] ; Loss: %f ; Time: %f' % (total_batch, loss, time_behind-time_before))
gen_gan_optm.zero_grad()
loss.backward()
gen_gan_optm.step()
rollout.update_params()
for epoch in range(2):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE,index_word_dict,g_sequence_len,"general")
dis_data_iter = DisGenDataIter(POSITIVE_FILE, NEGATIVE_FILE, BATCH_SIZE)
for _ in range(2):
time_before = time.time()
loss = discriminator_gen_epoch(discriminator_general, dis_data_iter)
time_behind = time.time()
print('Train Discriminator-General: Epoch [%d] ; Loss: %f ; Time: %f' % (epoch, loss, time_behind-time_before))
logging.info('Pretrain Discriminator-General: Epoch [%d] ; Loss: %f ; Time: %f' % (epoch, loss, time_behind-time_before))
for epoch in range(1):
generate_samples(generator, BATCH_SIZE, GENERATED_NUM, NEGATIVE_FILE_SPE,index_word_dict,g_sequence_len,"special")
dis_data_iter = DisSpeDataIter(NEGATIVE_FILE_SPE, BATCH_SIZE, REF_SENT, g_sequence_len, opinion, length_control, synthetic_num)
time_before = time.time()
loss = discriminator_spe_epoch(discriminator_special, dis_data_iter)
time_behind = time.time()
print('Train Discriminator-Special: Epoch [%d] ; Loss: %f ; Time: %f' % (epoch, loss, time_behind-time_before))
logging.info('Train Discriminator-Special: Epoch [%d] ; Loss: %f ; Time: %f' % (epoch, loss, time_behind-time_before))
torch.save(discriminator_general, 'DiscriminatorGeneral.pkl')
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)
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)
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():
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)
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)
if prob.is_cuda:
one_hot = one_hot.cuda()
loss = torch.masked_select(prob, one_hot)
loss = loss * reward
loss = -torch.sum(loss)
return loss
target_lstm = TargetLSTM(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
if opt.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)
original_generator = Generator(VOCAB_SIZE, g_emb_dim, g_hidden_dim, opt.cuda)
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_)
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 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 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)
if opt.cuda:
generator = generator.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)
torch.save(generator.state_dict(), CHECKPOINT_PATH + 'generator_mle.model')
track_blue = np.array(track_blue)
np.save(ROOT_PATH + 'track_blue_mle3.npy', track_blue)
plt.plot(track_blue)
plt.show()