def epoch(self, epoch_index):
for p, avg_p in zip(self.trainer.generator.parameters(),
self.my_g_clone):
avg_p.mul_(self.old_weight).add_((1.0 - self.old_weight) * p.data)
if epoch_index % self.output_snapshot_ticks == 0:
z = next(self.sample_fn(self.samples_count))
gen_input = cudize(z)
original_param = self.flatten_params(self.trainer.generator)
self.load_params(self.my_g_clone, self.trainer.generator)
dest = os.path.join(
self.checkpoints_dir,
SaverPlugin.last_pattern.format(
'smooth_generator',
'{:06}'.format(self.trainer.cur_nimg // 1000)))
torch.save(
{
'cur_nimg': self.trainer.cur_nimg,
'model': self.trainer.generator.state_dict()
}, dest)
out = generate_samples(self.trainer.generator, gen_input)
self.load_params(original_param, self.trainer.generator)
frequency = self.max_freq * out.shape[2] / self.seq_len
images = self.get_images(frequency, epoch_index, out)
for i, image in enumerate(images):
imwrite(
os.path.join(self.checkpoints_dir,
'{}_{}.png'.format(epoch_index, i)), image)
def main():
model = nn.Sequential(nn.Linear(D_in, H), nn.ReLU(), nn.Linear(H, D_out),
nn.Softmax(dim=1))
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=0.01)
lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, decay)
transform = AdditiveGaussianNoise(std=0.2, shape=(D_in, ), seed=0)
train_data = PolynomialDataset(points=D_in, transform=transform)
train_loader = DataLoader(
train_data,
batch_size=3,
shuffle=True, #
# num_workers=2,
)
# test_data = PolynomialDataset(points=D_in, repeat=85, transform=transform)
samples = generate_samples(seed=1234)
test_data = list(islice(samples, 256))
test_inputs = torch.stack([x for x, l in test_data])
test_labels = torch.tensor([l for x, l in test_data])
validation_loader = train_loader
train_step_ = partial(train_step, model, loss_fn, optimizer, lr_scheduler,
train_loader)
fmt_str = "{:5d} {:.1f} {:7.3f} {:7.3f} {:.1f}"
col_str = "{:>5} {:>4} {:>7} {:>7} {:>4}"
print(col_str.format("epoch", "acc", "loss", "valloss", "lr"))
for epoch in range(1000):
loss = train_step_()
if epoch % 100 != 0:
continue
test_accuracy = top1_accuracy(model, test_inputs, test_labels)
validation_loss = validation_step(model, loss_fn, validation_loader)
print("{:5d} {:.1f} {:7.3f} {:7.3f} {:.1f}".format(
epoch,
100 * test_accuracy,
loss,
validation_loss,
math.log(lr_scheduler.get_last_lr()[0]),
))
def main(length=40, num_epochs=20):
'''
Build and train LSTM network to solve XOR problem
'''
X_train, y_train, X_test, y_test = generate_samples(length=length)
model = build_model()
history = model.fit(X_train,
y_train,
epochs=num_epochs,
batch_size=32,
validation_split=0.10,
shuffle=False)
# Evaluate model on test set
preds = model.predict(X_test)
preds = np.round(preds[:, 0]).astype('float32')
acc = (np.sum(preds == y_test) / len(y_test)) * 100
print('Accuracy: {:.2f}%'.format(acc))
# Plotting loss and accuracy
model_plot(history)
return
classifier = Classifier()
sys.exit("Need to train a classifier!")
# TODO: train classifier
# directories for generated samples
dir_results = model_prefix + 'results'
dir_samples = model_prefix + 'samples'
create_directories(dir_results, dir_samples)
########## TEST MODE ##########
if args.epochs is None:
# load generator
G = torch.load(model_prefix + 'generator.pt').to(device)
print('Generator loaded!')
# generate samples
generate_samples(G, dir_samples, args.batchsize, num_samples=4096)
sampleloader = get_sample_loader(dir_samples, args.batchsize,
image_size)
print('Samples generated!')
# compute fid score with test set
fid_score = get_fid_score(classifier, sampleloader, testloader)
sys.exit("FID score from test set: " + str(fid_score))
########## TRAIN MODE ##########
try:
if not args.create:
G = torch.load(model_prefix + 'generator.pt').to(device)
D = torch.load(model_prefix + 'discriminator.pt').to(device)
Model = torch.load(model_prefix + 'model.pt').to(device)
print('Model loaded!')
else:
model = G(3, 3, 8)
device = 0
##################################################################
# define model's specific weights
cpm2d = './weights/cpm/stb_hpm2d.pth'
cpm3d = './weights/cpm/stb_hpm3d.pth'
weights = "./weights/pix2pix/stb_net_G.pth"
# init paramters
evaluate = Evaluator(model, weights,cpm2d, cpm3d, device)
opt = edict()
opt.dataroot = "./datasets/stb-dataset/test"
opt.isTrain = False
dataset = STBdataset(opt)
# generate samples
output_path = os.path.join(TEMP_PATH, "STB")
generate_samples(model , evaluate, dataset, output_path)
# run evaluation
samples_dir = [os.path.join(output_path, i) for i in os.listdir(output_path)]
p = Pool()
r = list(tqdm(p.imap(run_js, samples_dir),
total=len(samples_dir)))
p.close()
p.join()
#os.system(f"node test.js {output_path}")
run_evaluate(output_path)
##################################################################3
# opt.dataroot = "./datasets/rhd-dataset/test"
# cpm2d = "./weights/cpm/rhd_hpm2d.pth"
# weights = "./weights/pix2pix/rhd_net_G.pth"
from revers import Rjmcmc
from blrMoveFactory import BlrMoveFactory
from stationaryDistributionFactory import StationaryDistributionFactory
from mcmcFactory import McmcFactory
from plotter import Plotter
np.random.seed(0)
line_points_x = [0, 5, 10, 15]
line_points_y = [5, 0, 0, 5]
var = 3
n = 100
xs, ys = utils.generate_samples(line_points_x, line_points_y, n, var)
stats = StationaryDistributionFactory(xs, ys)
moves = BlrMoveFactory()
mcmcs = McmcFactory(stats)
first_sample = [1, 0, 5, 5, 0, 14, 5]
if not stats.get_stationary(1).pdf(first_sample) > 0:
raise Exception("First sample has zero probality")
rjmcmc = Rjmcmc(moves, mcmcs, stats)
samples = rjmcmc.sample(20000, (1, first_sample))
plotter = Plotter(xs, ys, line_points_x, line_points_y, samples)
def epoch(self, epoch_index):
gen_input = Variable(self.sample_fn(self.samples_count)).cuda()
out = generate_samples(self.trainer.G, gen_input)
for proc in self.output_postprocessors:
proc(out, self.trainer.cur_nimg // 1000)
import time
from utils import generate_samples, get_numbers_list
from bubble_sort import bubble_sort
from merge_sort import merge_sort
from radix_sort import radix_sort, radix_group_numbers
from selection_sort import selection_sort
#first we need to generate a hugh amount of data
generate_samples(100000)
lst = get_numbers_list('data.txt')
start_time = time.time()
radix_group_numbers(lst)
end_time = time.time()
print 'radix sort took', end_time - start_time, 'seconds'
start_time = time.time()
merge_sort(lst)
end_time = time.time()
print 'merge sort took', end_time - start_time, 'seconds'
start_time = time.time()
bubble_sort(lst)
end_time = time.time()
## Cat
image1_ = (1 - io.imread("data/shapes/cat_1.jpg")[:, :, 0] / 255)
image1 = resize(image1_, (200, 200), mode="reflect",
anti_aliasing=False).astype("bool") * 1
## Rabbit
image2_ = (1 -
io.imread("data/shapes/rabbit.png")[:, :, 1] / 255).astype("bool")
image2 = resize(image2_, (image1.shape[0], image1.shape[1]),
mode="reflect",
anti_aliasing=False).astype("bool") * 2
image = np.zeros((600, 600))
image[30:30 + n_x, 10:10 + n_y] = image1
image[320:320 + n_x, 350:350 + n_y] = image2[:, ::-1]
data = generate_samples(image, n_y, n_samples=60000, random_state=43)
X = data[0]
Y = data[1]
### Fit GMMs
from sklearn.model_selection import train_test_split
X_train, X_test = train_test_split(X, random_state=54)
Y_train, Y_test = train_test_split(Y, random_state=54)
n_components = 20
gmm_x = GaussianMixture(n_components=n_components, random_state=3).fit(X_train)
gmm_y = GaussianMixture(n_components=n_components,
random_state=34).fit(Y_train)
from utils import generate_samples
torch.manual_seed(0)
D_in = 10
H = 100 # TODO question: why do H=3 and H=4 produce bad results?
D_out = 3
decay = 0.999
learning_rate = 1e-1
loss_fn = nn.CrossEntropyLoss()
model = nn.Sequential(nn.Linear(D_in, H), nn.ReLU(), nn.Linear(H, D_out),
nn.Softmax(dim=1))
# model.apply(lambda m: hasattr(m, "weight") and nn.init.normal_(m.weight))
test_data = list(islice(generate_samples(seed=1234), 256))
test_inputs = torch.stack([x for x, l in test_data])
test_labels = torch.tensor([l for x, l in test_data])
def top1_accuracy() -> float:
with torch.no_grad():
preds = model(test_inputs)
pred_labels = preds.detach().argmax(dim=-1)
return (pred_labels == test_labels).sum().numpy() / len(test_labels)
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=0.01)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, decay)
def main(unused_argv):
config_train = training_config()
config_gen = generator_config()
config_dis = discriminator_config()
np.random.seed(config_train.seed)
assert config_train.start_token == 0
gen_data_loader = Gen_Data_loader(config_gen.gen_batch_size)
likelihood_data_loader = Gen_Data_loader(config_gen.gen_batch_size)
dis_data_loader = Dis_dataloader(config_dis.dis_batch_size)
generator = Generator(config=config_gen)
generator.build()
rollout_gen = rollout(config=config_gen)
target_params = pickle.load(open('save/target_params.pkl', 'rb'),
encoding='iso-8859-1')
target_lstm = TARGET_LSTM(config=config_gen,
params=target_params) # The oracle model 预测模型
discriminator = Discriminator(config=config_dis)
discriminator.build_discriminator()
pretrained_optimizer = tf.train.AdamOptimizer(
config_train.gen_learning_rate)
gradients, variables = zip(*pretrained_optimizer.compute_gradients(
generator.pretrained_loss, var_list=var_pretrained))
gradients, _ = tf.clip_by_global_norm(gradients, config_train.grad_clip)
gen_pre_update = pretrained_optimizer.apply_gradients(
zip(gradients, variables))
sess = tf.Session()
sess.run(tf.global_variables_initializer())
generate_samples(sess, target_lstm, config_train.batch_size,
config_train.generated_num, config_train.positive_file)
gen_data_loader.create_batches(config_train.positive_file)
log = open('save/experiment-log.txt', 'w')
print('Start pre-training generator....')
log.write('pre-training...\n')
for epoch in range(config_train.pretrained_epoch_num):
gen_data_loader.reset_pointer()
for it in range(gen_data_loader.num_batch):
batch = gen_data_loader.next_batch()
_, g_loss = sess.run(
[gen_pre_update, generator.pretrained_loss],
feed_dict={
generator.input_seqs_pre: batch,
generator.input_seqs_mask: np.ones_like(batch)
})
if epoch % config_train.test_per_epoch == 0:
generate_samples(sess, generator, config_train.batch_size,
config_train.generated_num,
config_train.eval_file)
likelihood_data_loader.create_batches(config_train.eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
print('pre-train ', epoch, ' test_loss ', test_loss)
buffer = 'epoch:\t' + str(epoch) + '\tnll:\t' + str(
test_loss) + '\n'
log.write(buffer)
print('Start pre-training discriminator...')
for t in range(config_train.dis_update_time_pre):
print("Times: " + str(t))
generate_samples(sess, generator, config_train.batch_size,
config_train.generated_num,
config_train.negative_file)
dis_data_loader.load_train_data(config_train.positive_file,
config_train.negative_file)
for _ in range(config_train.dis_update_time_pre):
dis_data_loader.reset_pointer()
for it in range(dis_data_loader.num_batch):
x_batch, y_batch = dis_data_loader.next_batch()
feed_dict = {
discriminator.input_x:
x_batch,
discriminator.input_y:
y_batch,
discriminator.dropout_keep_prob:
config_dis.dis_dropout_keep_prob # dropout
}
_ = sess.run(discriminator.train_op, feed_dict)
train_adv_opt = tf.train.AdamOptimizer(config_train.gen_learning_rate)
gradients, variables = zip(*train_adv_opt.compute_gradients(
generator.gen_loss_adv, var_list=var_pretrained))
gradients, _ = tf.clip_by_global_norm(gradients, config_train.grad_clip)
train_adv_update = train_adv_opt.apply_gradients(zip(gradients, variables))
# Initialize global variables of optimizer for adversarial training
uninitialized_var = [
e for e in tf.global_variables() if e not in tf.trainable_variables()
]
init_vars_uninit_op = tf.variables_initializer(uninitialized_var)
sess.run(init_vars_uninit_op)
for total_batch in range(config_train.total_batch):
for iter_gen in range(config_train.gen_update_time):
samples = sess.run(generator.sample_word_list_reshpae)
feed = {'pred_seq_rollout:0': samples}
reward_rollout = []
for iter_roll in range(config_train.rollout_num):
rollout_list = sess.run(rollout_gen.sample_rollout_step,
feed_dict=feed)
rollout_list_stack = np.vstack(rollout_list)
reward_rollout_seq = sess.run(
discriminator.ypred_for_auc,
feed_dict={
discriminator.input_x: rollout_list_stack,
discriminator.dropout_keep_prob: 1.0
})
reward_last_tok = sess.run(discriminator.ypred_for_auc,
feed_dict={
discriminator.input_x: samples,
discriminator.dropout_keep_prob:
1.0
})
reward_allseq = np.concatenate(
(reward_rollout_seq, reward_last_tok), axis=0)[:, 1]
reward_tmp = []
for r in range(config_gen.gen_batch_size):
reward_tmp.append(reward_allseq[range(
r,
config_gen.gen_batch_size * config_gen.sequence_length,
config_gen.gen_batch_size)])
reward_rollout.append(np.array(reward_tmp))
rewards = np.sum(reward_rollout,
axis=0) / config_train.rollout_num
_, gen_loss = sess.run(
[train_adv_update, generator.gen_loss_adv],
feed_dict={
generator.input_seqs_adv: samples,
generator.rewards: rewards
})
if total_batch % config_train.test_per_epoch == 0 or total_batch == config_train.total_batch - 1:
generate_samples(sess, generator, config_train.batch_size,
config_train.generated_num,
config_train.eval_file)
likelihood_data_loader.create_batches(config_train.eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'epoch:\t' + str(total_batch) + '\tnll:\t' + str(
test_loss) + '\n'
print('total_batch: ', total_batch, 'test_loss: ', test_loss)
log.write(buffer)
for _ in range(config_train.dis_update_time_adv):
generate_samples(sess, generator, config_train.batch_size,
config_train.generated_num,
config_train.negative_file)
dis_data_loader.load_train_data(config_train.positive_file,
config_train.negative_file)
for _ in range(config_train.dis_update_time_adv):
dis_data_loader.reset_pointer()
for it in range(dis_data_loader.num_batch):
x_batch, y_batch = dis_data_loader.next_batch()
feed = {
discriminator.input_x:
x_batch,
discriminator.input_y:
y_batch,
discriminator.dropout_keep_prob:
config_dis.dis_dropout_keep_prob
}
_ = sess.run(discriminator.train_op, feed)
log.close()
def main():
random.seed(SEED)
np.random.seed(SEED)
assert START_TOKEN == 0
vocab_w2idx, vocab_idx2w, len_vocab_w2idx = utils.load_vocab(vacab_file)
gen_data_loader = Gen_Data_loader(BATCH_SIZE, seq_len=SEQ_LENGTH)
dis_data_loader = Dis_dataloader(BATCH_SIZE, seq_len=SEQ_LENGTH)
generator = Generator(vocab_size, BATCH_SIZE, EMB_DIM, HIDDEN_DIM, SEQ_LENGTH, START_TOKEN)
discriminator = Discriminator(sequence_length=SEQ_LENGTH, num_classes=2, vocab_size=vocab_size, embedding_size=dis_embedding_dim,
filter_sizes=dis_filter_sizes, num_filters=dis_num_filters, l2_reg_lambda=dis_l2_reg_lambda)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
gen_data_loader.create_batches(positive_file)
print("gen_data_loader num_batches: ", gen_data_loader.num_batch)
# pre-train generator
print 'Start pre-training...'
for epoch in xrange(PRE_EPOCH_NUM):
loss = pre_train_epoch(sess, generator, gen_data_loader)
if epoch % 1 == 0:
utils.test_demo(sess, generator, 8, 8, eval_file, vocab_idx2w, epoch)
print 'pre-train epoch ', epoch, 'train_loss ', loss
print 'Start pre-training discriminator...'
# Train 3 epoch on the generated data and do this for 50 times
for _ in range(10):
utils.generate_samples(sess, generator, 64, generated_num, negative_file)
dis_data_loader.load_train_data(positive_file, negative_file)
for _ in range(3):
dis_data_loader.reset_pointer()
for it in xrange(dis_data_loader.num_batch):
x_batch, y_batch = dis_data_loader.next_batch()
feed = {
discriminator.input_x: x_batch,
discriminator.input_y: y_batch,
discriminator.dropout_keep_prob: dis_dropout_keep_prob
}
_ = sess.run(discriminator.train_op, feed)
rollout = ROLLOUT(generator, 0.8)
print '#########################################################################'
print 'Start Adversarial Training...'
for total_batch in range(TOTAL_BATCH):
# Train the generator for one step
for it in range(1):
samples = generator.generate(sess)
rewards = rollout.get_reward(sess, samples, 16, discriminator)
feed = {generator.x: samples, generator.rewards: rewards}
_ = sess.run(generator.g_updates, feed_dict=feed)
# Test
if total_batch % 5 == 0 or total_batch == TOTAL_BATCH - 1:
print 'adversarial-train epoch ', total_batch
utils.test_demo(sess, generator, 8, 8, eval_file, vocab_idx2w, total_batch)
# Update roll-out parameters
rollout.update_params()
# Train the discriminator
for _ in range(5):
dis_data_loader.load_train_data(positive_file, negative_file)
for _ in range(3):
dis_data_loader.reset_pointer()
for it in xrange(dis_data_loader.num_batch):
x_batch, y_batch = dis_data_loader.next_batch()
feed = {
discriminator.input_x: x_batch,
discriminator.input_y: y_batch,
discriminator.dropout_keep_prob: dis_dropout_keep_prob
}
_ = sess.run(discriminator.train_op, feed)
def main(unused_argv):
gen_data_loader = dataloader.Gen_data_loader(64)
dis_data_loader = dataloader.Dis_data_loader(64)
likelihood_data_loader = dataloader.Gen_data_loader(64)
gen_model = generator.Generator()
gen_model.build()
dis_model = discriminator.Discriminator()
dis_model.build()
rollout_model = rollout.rollout()
target_params = pickle.load(open("save/target_params.pkl", "rb"),
encoding="iso-8859-1")
target_lstm_model = target_lstm.TARGET_LSTM(params=target_params)
# Build optimizer op for pretraining
pretrained_optimizer = tf.train.AdamOptimizer(1e-2)
var_pretrained = [
v for v in tf.trainable_variables() if "teller" in v.name
]
gradients1, variables1 = zip(*pretrained_optimizer.compute_gradients(
gen_model.pretrained_loss, var_list=var_pretrained))
gradients1, _ = tf.clip_by_global_norm(gradients1, 5.0)
gen_pre_update = pretrained_optimizer.apply_gradients(
zip(gradients1, variables1))
# Build optimizer op for adversarial training
advtrained_optimizer = tf.train.AdamOptimizer(1e-2)
gradients2, variables2 = zip(*advtrained_optimizer.compute_gradients(
gen_model.adversarial_loss, var_list=var_pretrained))
gradients2, _ = tf.clip_by_global_norm(gradients2, 5.0)
gen_adv_update = advtrained_optimizer.apply_gradients(
zip(gradients2, variables2))
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
log = open("save/log.txt", "w", encoding="utf-8")
print("Start pre-training generator.....")
utils.generate_samples(sess, target_lstm_model, 64, 1000,
"save/real_data.txt")
gen_data_loader.create_batches("save/real_data.txt")
log.write("Start pre-training generator.....\n")
for epoch in range(20):
gen_data_loader.reset_pointer()
for it in range(gen_data_loader.num_batch):
batch = gen_data_loader.next_batch()
_, g_loss = sess.run(
[gen_pre_update, gen_model.pretrained_loss],
feed_dict={
gen_model.input_seqs_pre: batch,
gen_model.input_seqs_mask: np.ones_like(batch)
})
if epoch % 5 == 0:
utils.generate_samples(sess, gen_model, 64, 1000,
"save/eval_data.txt")
likelihood_data_loader.create_batches("save/eval_data.txt")
test_loss = utils.target_loss(sess, target_lstm_model,
likelihood_data_loader)
print("epoch:", epoch, " test_loss:", test_loss)
buffer = 'epoch:\t' + str(epoch) + '\tnll:\t' + str(
test_loss) + '\n'
log.write(buffer)
print("Start pre-training discriminator.....")
for t in range(2):
print("Times: " + str(t))
utils.generate_samples(sess, gen_model, 64, 1000,
"save/generate_sample.txt")
dis_data_loader.create_batches("save/real_data.txt",
"save/generate_sample.txt")
for epoch in range(20):
dis_data_loader.reset_pointer()
for it in range(dis_data_loader.num_batch):
x_batch, y_batch = dis_data_loader.next_batch()
sess.run(dis_model.train_op,
feed_dict={
dis_model.input_data: x_batch,
dis_model.input_label: y_batch,
dis_model.keep_prob: 0.5
})
print("Start adversarial training.....")
for total_batch in range(20):
for gen_step in range(1):
samples = sess.run(gen_model.sample_word_list_reshape)
reward_rollout = []
for rollout_num in range(16):
rollout_list = sess.run(
rollout_model.sample_rollout_step,
feed_dict={rollout_model.pre_seq: samples})
rollout_list_stack = np.vstack(rollout_list)
reward_rollout_seq = sess.run(dis_model.ypred_for_auc,
feed_dict={
dis_model.input_data:
rollout_list_stack,
dis_model.keep_prob: 1.0
})
reward_all_seq = reward_rollout_seq[:, 1]
reward_tmp = []
for r in range(64):
reward_tmp.append(reward_all_seq[range(r, 64 * 20, 64)])
reward_rollout.append(np.array(reward_tmp))
rewards = np.mean(reward_rollout, axis=0)
_, gen_loss = sess.run(
[gen_adv_update, gen_model.adversarial_loss],
feed_dict={
gen_model.input_seqs_adv: samples,
gen_model.rewards: rewards
})
if total_batch % 5 == 0 or total_batch == 199:
utils.generate_samples(sess, gen_model, 64, 1000,
"save/eval_data.txt")
likelihood_data_loader.create_batches("save/eval_data.txt")
test_loss = utils.target_loss(sess, target_lstm_model,
likelihood_data_loader)
print('total_batch:', total_batch, ' test_loss:', test_loss)
buffer = 'total_batch:\t' + str(total_batch) + '\tnll:\t' + str(
test_loss) + '\n'
log.write(buffer)
for dis_step in range(5):
utils.generate_samples(sess, gen_model, 64, 1000,
"save/generate_sample.txt")
dis_data_loader.create_batches("save/real_data.txt",
"save/generate_sample.txt")
for epoch in range(3):
dis_data_loader.reset_pointer()
for it in range(dis_data_loader.num_batch):
x_batch, y_batch = dis_data_loader.next_batch()
sess.run(dis_model.train_op,
feed_dict={
dis_model.input_data: x_batch,
dis_model.input_label: y_batch,
dis_model.keep_prob: 0.5
})
log.close()
sess.close()
def main(unused_argv):
config_train = training_config()
config_gen = generator_config()
config_dis = discriminator_config()
np.random.seed(config_train.seed)
assert config_train.start_token == 0
gen_data_loader = Gen_Data_loader(config_gen.gen_batch_size)
likelihood_data_loader = Gen_Data_loader(config_gen.gen_batch_size)
dis_data_loader = Dis_dataloader(config_dis.dis_batch_size)
generator = Generator(config=config_gen)
generator.build()
rollout_gen = rollout(config=config_gen)
#Build target LSTM
target_params = pickle.load(open('save/target_params.pkl','rb'),encoding='iso-8859-1')
target_lstm = TARGET_LSTM(config=config_gen, params=target_params) # The oracle model
# Build discriminator
discriminator = Discriminator(config=config_dis)
discriminator.build_discriminator()
# Build optimizer op for pretraining
pretrained_optimizer = tf.train.AdamOptimizer(config_train.gen_learning_rate)
var_pretrained = [v for v in tf.trainable_variables() if 'teller' in v.name]
gradients, variables = zip(
*pretrained_optimizer.compute_gradients(generator.pretrained_loss, var_list=var_pretrained))
gradients, _ = tf.clip_by_global_norm(gradients, config_train.grad_clip)
gen_pre_update = pretrained_optimizer.apply_gradients(zip(gradients, variables))
sess = tf.Session()
sess.run(tf.global_variables_initializer())
generate_samples(sess,target_lstm,config_train.batch_size,config_train.generated_num,config_train.positive_file)
gen_data_loader.create_batches(config_train.positive_file)
log = open('save/experiment-log.txt','w')
print('Start pre-training generator....')
log.write('pre-training...\n')
for epoch in range(config_train.pretrained_epoch_num):
gen_data_loader.reset_pointer()
for it in range(gen_data_loader.num_batch):
batch = gen_data_loader.next_batch()
_,g_loss = sess.run([gen_pre_update,generator.pretrained_loss],feed_dict={generator.input_seqs_pre:batch,
generator.input_seqs_mask:np.ones_like(batch)})
if epoch % config_train.test_per_epoch == 0:
#进行测试，通过Generator产生一批序列，
generate_samples(sess,generator,config_train.batch_size,config_train.generated_num,config_train.eval_file)
# 创建这批序列的data-loader
likelihood_data_loader.create_batches(config_train.eval_file)
# 使用oracle 计算 交叉熵损失nll
test_loss = target_loss(sess,target_lstm,likelihood_data_loader)
# 打印并写入日志
print('pre-train ',epoch, ' test_loss ',test_loss)
buffer = 'epoch:\t' + str(epoch) + '\tnll:\t' + str(test_loss) + '\n'
log.write(buffer)
print('Start pre-training discriminator...')
for t in range(config_train.dis_update_time_pre):
print("Times: " + str(t))
generate_samples(sess,generator,config_train.batch_size,config_train.generated_num,config_train.negative_file)
dis_data_loader.load_train_data(config_train.positive_file,config_train.negative_file)
for _ in range(config_train.dis_update_time_pre):
dis_data_loader.reset_pointer()
for it in range(dis_data_loader.num_batch):
x_batch,y_batch = dis_data_loader.next_batch()
feed_dict = {
discriminator.input_x : x_batch,
discriminator.input_y : y_batch,
discriminator.dropout_keep_prob : config_dis.dis_dropout_keep_prob
}
_ = sess.run(discriminator.train_op,feed_dict)
# Build optimizer op for adversarial training
train_adv_opt = tf.train.AdamOptimizer(config_train.gen_learning_rate)
gradients, variables = zip(*train_adv_opt.compute_gradients(generator.gen_loss_adv, var_list=var_pretrained))
gradients, _ = tf.clip_by_global_norm(gradients, config_train.grad_clip)
train_adv_update = train_adv_opt.apply_gradients(zip(gradients, variables))
# Initialize global variables of optimizer for adversarial training
uninitialized_var = [e for e in tf.global_variables() if e not in tf.trainable_variables()]
init_vars_uninit_op = tf.variables_initializer(uninitialized_var)
sess.run(init_vars_uninit_op)
# Start adversarial training
for total_batch in range(config_train.total_batch):
for iter_gen in range(config_train.gen_update_time):
samples = sess.run(generator.sample_word_list_reshpae)
feed = {'pred_seq_rollout:0':samples}
reward_rollout = []
for iter_roll in range(config_train.rollout_num):
rollout_list = sess.run(rollout_gen.sample_rollout_step,feed_dict=feed)
# np.vstack 它是垂直（按照行顺序）的把数组给堆叠起来。
rollout_list_stack = np.vstack(rollout_list)
reward_rollout_seq = sess.run(discriminator.ypred_for_auc,feed_dict={
discriminator.input_x:rollout_list_stack,discriminator.dropout_keep_prob:1.0
})
reward_last_tok = sess.run(discriminator.ypred_for_auc,feed_dict={
discriminator.input_x:samples,discriminator.dropout_keep_prob:1.0
})
reward_allseq = np.concatenate((reward_rollout_seq,reward_last_tok),axis=0)[:,1]
reward_tmp = []
for r in range(config_gen.gen_batch_size):
reward_tmp.append(reward_allseq[range(r,config_gen.gen_batch_size * config_gen.sequence_length,config_gen.gen_batch_size)])
reward_rollout.append(np.array(reward_tmp))
rewards = np.sum(reward_rollout,axis = 0) / config_train.rollout_num
_,gen_loss = sess.run([train_adv_update,generator.gen_loss_adv],feed_dict={generator.input_seqs_adv:samples,
generator.rewards:rewards})
if total_batch % config_train.test_per_epoch == 0 or total_batch == config_train.total_batch - 1:
generate_samples(sess, generator, config_train.batch_size, config_train.generated_num, config_train.eval_file)
likelihood_data_loader.create_batches(config_train.eval_file)
test_loss = target_loss(sess, target_lstm, likelihood_data_loader)
buffer = 'epoch:\t' + str(total_batch) + '\tnll:\t' + str(test_loss) + '\n'
print ('total_batch: ', total_batch, 'test_loss: ', test_loss)
log.write(buffer)
for _ in range(config_train.dis_update_time_adv):
generate_samples(sess,generator,config_train.batch_size,config_train.generated_num,config_train.negative_file)
dis_data_loader.load_train_data(config_train.positive_file,config_train.negative_file)
for _ in range(config_train.dis_update_time_adv):
dis_data_loader.reset_pointer()
for it in range(dis_data_loader.num_batch):
x_batch,y_batch = dis_data_loader.next_batch()
feed = {
discriminator.input_x:x_batch,
discriminator.input_y:y_batch,
discriminator.dropout_keep_prob:config_dis.dis_dropout_keep_prob
}
_ = sess.run(discriminator.train_op,feed)
log.close()
def epoch(self, epoch_index):
gen_input = Variable(self.sample_fn(self.samples_count)).cuda()
out = generate_samples(self.trainer.G, gen_input)
for proc in self.output_postprocessors:
proc(out, self.trainer.cur_nimg // 1000)
def train(model):
# Checks what kind of model it is training
if model.model_type == "infoGAN":
is_infogan = True
else:
is_infogan = False
# Makes sure we have a dir to save the model and training info
if not os.path.exists(model.save_dir):
os.makedirs(model.save_dir)
# Creates artificial labels that just indicates to the loss object if prediction of D should be 0 or 1
if model.gpu_mode:
y_real_ = Variable(torch.ones(model.batch_size,
1).cuda(model.gpu_id)) # all ones
y_fake_ = Variable(
torch.zeros(model.batch_size, 1).cuda(model.gpu_id)) # all zeros
else:
y_real_ = Variable(torch.ones(model.batch_size, 1))
y_fake_ = Variable(torch.zeros(model.batch_size, 1))
model.D.train() # sets discriminator in train mode
# TRAINING LOOP
start_time = time.time()
print('[*] TRAINING STARTS')
for epoch in range(model.epoch):
model.G.train() # sets generator in train mode
epoch_start_time = time.time()
# For each minibatch returned by the data_loader
for step, (x_, _) in enumerate(model.data_loader):
if step == model.data_loader.dataset.__len__() // model.batch_size:
break
# Creates a minibatch of latent vectors
z_ = torch.rand((model.batch_size, model.z_dim))
# Creates a minibatch of discrete and continuous codes
c_disc_ = torch.from_numpy(
np.random.multinomial(
1,
model.c_disc_dim * [float(1.0 / model.c_disc_dim)],
size=[model.batch_size])).type(torch.FloatTensor)
for i in range(model.n_disc_code - 1):
c_disc_ = torch.cat([
c_disc_,
torch.from_numpy(
np.random.multinomial(
1,
model.c_disc_dim * [float(1.0 / model.c_disc_dim)],
size=[model.batch_size])).type(torch.FloatTensor)
],
dim=1)
c_cont_ = torch.from_numpy(
np.random.uniform(-1,
1,
size=(model.batch_size,
model.c_cont_dim))).type(
torch.FloatTensor)
# Convert to Variables (sends to GPU if needed)
if model.gpu_mode:
x_ = Variable(x_.cuda(model.gpu_id))
z_ = Variable(z_.cuda(model.gpu_id))
c_disc_ = Variable(c_disc_.cuda(model.gpu_id))
c_cont_ = Variable(c_cont_.cuda(model.gpu_id))
else:
x_ = Variable(x_)
z_ = Variable(z_)
c_disc_ = Variable(c_disc_)
c_cont_ = Variable(c_cont_)
# update D network
model.D_optimizer.zero_grad()
D_real, _, _ = model.D(x_, model.dataset)
D_real_loss = model.BCE_loss(D_real, y_real_)
G_ = model.G(z_, c_cont_, c_disc_, model.dataset)
D_fake, _, _ = model.D(G_, model.dataset)
D_fake_loss = model.BCE_loss(D_fake, y_fake_)
D_loss = D_real_loss + D_fake_loss
model.train_history['D_loss'].append(D_loss.data[0])
D_loss.backward(retain_graph=is_infogan)
model.D_optimizer.step()
# update G network
model.G_optimizer.zero_grad()
G_ = model.G(z_, c_cont_, c_disc_, model.dataset)
D_fake, D_cont, D_disc = model.D(G_, model.dataset)
G_loss = model.BCE_loss(D_fake, y_real_)
model.train_history['G_loss'].append(G_loss.data[0])
G_loss.backward(retain_graph=is_infogan)
model.G_optimizer.step()
# information loss
if is_infogan:
disc_loss = 0
for i, ce_loss in enumerate(model.CE_losses):
i0 = i * model.c_disc_dim
i1 = (i + 1) * model.c_disc_dim
disc_loss += ce_loss(D_disc[:, i0:i1],
torch.max(c_disc_[:, i0:i1], 1)[1])
cont_loss = model.MSE_loss(D_cont, c_cont_)
info_loss = disc_loss + cont_loss
model.train_history['info_loss'].append(info_loss.data[0])
info_loss.backward()
model.info_optimizer.step()
# Prints training info every 100 steps
if ((step + 1) % 100) == 0:
if is_infogan:
print(
"Epoch: [{:2d}] [{:4d}/{:4d}] D_loss: {:.8f}, G_loss: {:.8f}, info_loss: {:.8f}"
.format((epoch + 1), (step + 1),
model.data_loader.dataset.__len__() //
model.batch_size, D_loss.data[0],
G_loss.data[0], info_loss.data[0]))
else:
print(
"Epoch: [{:2d}] [{:4d}/{:4d}] D_loss: {:.8f}, G_loss: {:.8f}"
.format((epoch + 1), (step + 1),
model.data_loader.dataset.__len__() //
model.batch_size, D_loss.data[0],
G_loss.data[0]))
model.train_history['per_epoch_time'].append(time.time() -
epoch_start_time)
# Saves samples
utils.generate_samples(
model, os.path.join(model.save_dir, "epoch{}.png".format(epoch)))
model.train_history['total_time'].append(time.time() - start_time)
print("Avg one epoch time: %.2f, total %d epochs time: %.2f" %
(np.mean(model.train_history['per_epoch_time']), model.epoch,
model.train_history['total_time'][0]))
print("[*] TRAINING FINISHED")
# Saves the model
model.save()
# Saves the plot of losses for G and D
utils.save_loss_plot(model.train_history,
filename=os.path.join(model.save_dir, "curves.png"),
infogan=is_infogan)