def train(self):
if not os.path.exists(os.path.join("data", self.dataset_name)):
print "No GAN training files found. Training aborted. =("
return
dataset_files = glob.glob("data/" + self.dataset_name + "/*.png")
dataset_files.sort(key=ops.alphanum_key)
dataset_files = np.array(dataset_files)
dataset_encodings = np.load("data/" + self.dataset_name +
"/encondings.npy")
n_files = dataset_files.shape[0]
training_step = 0
test_idxs = np.random.permutation(range(n_files))[0:self.batch_size]
test_imgs = ops.load_imgbatch(dataset_files[test_idxs], color=False)
test_encs = dataset_encodings[test_idxs, :]
self.session.run(tf.initialize_all_variables())
self.load()
for epoch in xrange(self.n_iterations):
rand_idxs = np.random.permutation(range(n_files))
n_batches = n_files // self.batch_size
for batch_i in xrange(n_batches):
idxs_i = rand_idxs[batch_i * self.batch_size:(batch_i + 1) *
self.batch_size]
imgs_batch = ops.load_imgbatch(dataset_files[idxs_i],
color=False)
self.session.run(self.optimizer,
feed_dict={
self.images: imgs_batch,
self.encodings:
dataset_encodings[idxs_i, :],
self.train_flag: True
})
training_step += 1
current_loss = self.session.run(self.loss,
feed_dict={
self.images: test_imgs,
self.encodings: test_encs,
self.train_flag: False
})
print "Epoch {}/{}, Batch {}/{}, Loss {}".format(
epoch + 1, self.n_iterations, batch_i + 1, n_batches,
current_loss)
# Save checkpoint
if training_step % 1000 == 0:
if not os.path.exists("checkpoint"):
print "Checkpoint folder not found. reating one..."
os.makedirs("checkpoint")
print "Done."
create_folder('checkpoint/InversePrGAN{}'.format(
self.dataset_name))
self.saver.save(
self.session,
'checkpoint/InversePrGAN{}/model.ckpt'.format(
self.dataset_name),
global_step=training_step)
def train(self):
if not os.path.exists(os.path.join("data", "train")):
print "No training files found. Training aborted. =("
return
dataset_files = glob.glob("data/train/*.png")
dataset_files.sort(key=ops.alphanum_key)
dataset_files = np.array(dataset_files)
dataset_params = np.load("train_params.npy")
n_files = dataset_params.shape[0]
testset_idxs = np.random.choice(range(n_files), self.batch_size)
test_imgs = ops.load_imgbatch(dataset_files[testset_idxs])
training_step = 0
self.session.run(tf.initialize_all_variables())
for epoch in xrange(self.n_iterations):
rand_idxs = np.random.permutation(range(n_files))
n_batches = n_files // self.batch_size
for batch_i in xrange(n_batches):
idxs_i = rand_idxs[batch_i * self.batch_size: (batch_i + 1) * self.batch_size]
imgs_batch = ops.load_imgbatch(dataset_files[idxs_i])
self.session.run(self.optimizer, feed_dict={self.img_params: dataset_params[idxs_i, :],
self.final_image: imgs_batch})
training_step += 1
current_loss = self.session.run(self.loss, feed_dict={self.img_params: dataset_params[testset_idxs, :],
self.final_image: test_imgs})
print "Epoch {}/{}, Batch {}/{}, Loss {}".format(epoch + 1, self.n_iterations,
batch_i + 1, n_batches, current_loss)
# Save checkpoint
if training_step % 1000 == 0:
if not os.path.exists("checkpoint"):
print "Checkpoint folder not found. Creating one..."
os.makedirs("checkpoint")
print "Done."
self.saver.save(self.session, 'checkpoint/model.ckpt', global_step=training_step)
def test(self, path):
self.load()
imgs_path = glob.glob(path)
imgs_path.sort(key=ops.alphanum_key)
imgs_path = np.array(imgs_path)
imgs_batch = ops.load_imgbatch(imgs_path[range(8 * 337, 8 * 337 + 64)],
color=False)
encs = self.z.eval(session=self.session,
feed_dict={
self.images: imgs_batch,
self.train_flag: False
})
np.save("inverse_encs.npy", encs)
def test(self):
test_files = glob.glob("data/test/*.png")
test_files.sort(key=ops.alphanum_key)
test_files = np.array(test_files)
test_params = np.load("test_params.npy")
n_files = test_params.shape[0]
test_idxs = np.random.choice(range(n_files), self.batch_size)
test_imgs = ops.load_imgbatch(test_files[test_idxs])
ops.save_images(test_imgs, [8, 8], 'ground_truth.png')
result_imgs = self.forward_batch(test_params[test_idxs, :])
ops.save_images(result_imgs, [8, 8], 'test_results.png')
def train(self):
if not os.path.exists(os.path.join("data", self.dataset_name)):
print "No GAN training files found. Training aborted. =("
return
dataset_files = glob.glob("data/" + self.dataset_name + "/*.png")
dataset_files = np.array(dataset_files)
n_files = dataset_files.shape[0]
sample_z = np.random.uniform(-1, 1, [self.batch_size, self.z_size])
training_step = 0
self.session.run(tf.initialize_all_variables())
self.load()
for epoch in xrange(self.n_iterations):
rand_idxs = np.random.permutation(range(n_files))
n_batches = n_files // self.batch_size
for batch_i in xrange(n_batches):
idxs_i = rand_idxs[batch_i * self.batch_size:(batch_i + 1) *
self.batch_size]
imgs_batch = ops.load_imgbatch(dataset_files[idxs_i],
color=False)
#imgs_batch = ops.load_voxelbatch(dataset_files[idxs_i])
batch_z = np.random.uniform(-1, 1,
[self.batch_size, self.z_size])
dloss_fake = self.D_fake.eval(session=self.session,
feed_dict={
self.z: batch_z,
self.train_flag: False
})
dloss_real = self.D_real.eval(session=self.session,
feed_dict={
self.images: imgs_batch,
self.train_flag: False
})
gloss = self.G_loss.eval(session=self.session,
feed_dict={
self.z: batch_z,
self.images: imgs_batch,
self.train_flag: False
})
train_discriminator = True
margin = 0.8
dacc_real = np.mean(dloss_real)
dacc_fake = np.mean(np.ones_like(dloss_fake) - dloss_fake)
dacc = (dacc_real + dacc_fake) * 0.5
#print np.mean(dloss_real)
#print np.mean(dloss_fake)
if dacc > margin:
train_discriminator = False
#if dloss_fake > 1.0-margin or dloss_real > 1.0-margin:
# train_generator = False
#if train_discriminator is False and train_generator is False:
# train_generator = train_discriminator = True
print "EPOCH[{}], BATCH[{}/{}]".format(epoch, batch_i,
n_batches)
print "Discriminator avg acc: {}".format(dacc)
print "Discriminator real mean: {}".format(np.mean(dloss_real))
print "Discriminator fake mean: {}".format(np.mean(dloss_fake))
print "Generator Loss:{}".format(gloss)
# Update discriminator
if train_discriminator:
print "***Discriminator trained.***"
self.session.run(self.D_optim,
feed_dict={
self.images: imgs_batch,
self.z: batch_z,
self.train_flag: True
})
# Update generator
#if dacc > 0.9:
# self.session.run(self.G_optim_classic, feed_dict={self.z: batch_z})
#if dacc > margin + 1.0:
self.session.run(self.G_optim_classic,
feed_dict={
self.z: batch_z,
self.images: imgs_batch,
self.train_flag: True
})
#self.session.run(self.G_optim, feed_dict={self.z: batch_z, self.images: imgs_batch, self.train_flag: True})
if batch_i % 50 == 0:
rendered_images = self.G.eval(session=self.session,
feed_dict={
self.z: sample_z,
self.images: imgs_batch,
self.train_flag: False
})
rendered_images = np.array(rendered_images)
voxels = self.voxels.eval(session=self.session,
feed_dict={
self.z: sample_z,
self.images: imgs_batch,
self.train_flag: False
})
voxels = np.array(voxels)
create_folder("results/{}".format(self.dataset_name))
ops.save_images(
rendered_images, [8, 8], "results/{}/{}.png".format(
self.dataset_name, epoch * n_batches + batch_i))
ops.save_images(imgs_batch, [8, 8], "sanity_chairs.png")
ops.save_voxels(voxels,
"results/{}".format(self.dataset_name))
print "Saving checkpoint..."
create_folder('checkpoint/{}'.format(self.dataset_name))
self.saver.save(self.session,
'checkpoint/{}/model.ckpt'.format(
self.dataset_name),
global_step=training_step)
print "***CHECKPOINT SAVED***"
training_step += 1
self.history["generator"].append(gloss)
self.history["discriminator_real"].append(dloss_real)
self.history["discriminator_fake"].append(dloss_fake)
np.save(os.path.join(self.logpath, "generator.npy"),
np.array(self.history["generator"]))
np.save(os.path.join(self.logpath, "discriminator_real.npy"),
np.array(self.history["discriminator_real"]))
np.save(os.path.join(self.logpath, "discriminator_fake.npy"),
np.array(self.history["discriminator_fake"]))
def train(self):
if not os.path.exists(os.path.join("data", "gan")):
print "No GAN training files found. Training aborted. =("
return
dataset_files = glob.glob("data/gan/*.png")
dataset_files = np.array(dataset_files)
n_files = dataset_files.shape[0]
sample_z = np.random.uniform(-1, 1, [self.batch_size, self.z_size])
self.session.run(tf.initialize_all_variables())
self.rendernet.load('checkpoint')
for epoch in xrange(self.n_iterations):
rand_idxs = np.random.permutation(range(n_files))
n_batches = n_files // self.batch_size
for batch_i in xrange(n_batches):
idxs_i = rand_idxs[batch_i * self.batch_size:(batch_i + 1) *
self.batch_size]
imgs_batch = ops.load_imgbatch(dataset_files[idxs_i])
batch_z = np.random.uniform(-1, 1,
[self.batch_size, self.z_size])
dloss_fake = self.D_loss_fake.eval(session=self.session,
feed_dict={self.z: batch_z})
dloss_real = self.D_loss_real.eval(
session=self.session, feed_dict={self.images: imgs_batch})
gloss = self.G_loss.eval(session=self.session,
feed_dict={self.z: batch_z})
train_discriminator = True
train_generator = True
margin = 0.3
if dloss_fake < margin or dloss_real < margin:
train_discriminator = False
if dloss_fake > 1.0 - margin or dloss_real > 1.0 - margin:
train_generator = False
if train_discriminator is False and train_generator is False:
train_generator = train_discriminator = True
# Update discriminator
if train_discriminator:
self.session.run(self.D_optim,
feed_dict={
self.images: imgs_batch,
self.z: batch_z
})
# Update generator
if train_generator:
for i in xrange(5):
self.session.run(self.G_optim,
feed_dict={self.z: batch_z})
if batch_i % 10 == 0:
rendered_images = self.G.eval(session=self.session,
feed_dict={self.z: sample_z})
rendered_images = np.array(rendered_images)
ops.save_images(
rendered_images, [8, 8],
"results/gancubes{}.png".format(epoch * n_batches +
batch_i))
print "EPOCH[{}], BATCH[{}/{}]".format(epoch, batch_i,
n_batches)
print "Discriminator Loss - Real:{} / Fake:{} - Total:{}".format(
dloss_real, dloss_fake, dloss_real + dloss_fake)
print "Generator Loss:{}".format(gloss)