def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim', type=int, default=100,
help='Noise dimension')
parser.add_argument('--t_dim', type=int, default=256,
help='Text feature dimension')
parser.add_argument('--batch_size', type=int, default=64,
help='Batch Size')
parser.add_argument('--image_size', type=int, default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim', type=int, default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim', type=int, default=64,
help='Number of conv in the first layer discr.')
parser.add_argument('--gfc_dim', type=int, default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length', type=int, default=2400,
help='Caption Vector Length')
parser.add_argument('--data_dir', type=str, default="Data",
help='Data Directory')
parser.add_argument('--learning_rate', type=float, default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1', type=float, default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs', type=int, default=600,
help='Max number of epochs')
parser.add_argument('--save_every', type=int, default=30,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model', type=str, default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set', type=str, default="flowers",
help='Dat set: MS-COCO, flowers')
args = parser.parse_args()
model_options = {
'z_dim' : args.z_dim,
't_dim' : args.t_dim,
'batch_size' : args.batch_size,
'image_size' : args.image_size,
'gf_dim' : args.gf_dim,
'df_dim' : args.df_dim,
'gfc_dim' : args.gfc_dim,
'caption_vector_length' : args.caption_vector_length
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
d_optim = tf.train.AdamOptimizer(args.learning_rate, beta1 = args.beta1).minimize(loss['d_loss'], var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate, beta1 = args.beta1).minimize(loss['g_loss'], var_list=variables['g_vars'])
sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
print "load?"
saver = tf.train.Saver(max_to_keep=None)
if args.resume_model:
print "loading"
saver.restore(sess, args.resume_model)
print "done loading model"
print "load training data"
loaded_data, test_data = load_training_data(args.data_dir, args.data_set)
print "done loading training data"
print(args.data_dir)
print(args.data_set)
logger.info("Starting")
for i in range(1, args.epochs + 1):
print "epoch", i
batch_no = 0
num_batches = int(len(loaded_data['image_list']) / args.batch_size)
print len(loaded_data['image_list'])
total_d_loss = 0
total_g_loss = 0
while batch_no*args.batch_size < loaded_data['data_length']:
#while batch_no < 2:
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(batch_no, args.batch_size,
args.image_size, args.z_dim, args.caption_vector_length, 'train', args.data_dir, args.data_set, loaded_data)
# DISCR UPDATE
check_ts = [ checks['d_loss1'] , checks['d_loss2'], checks['d_loss3']]
_, d_loss, gen, d1, d2, d3 = sess.run([d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
#print "d1", d1
#print "d2", d2
#print "d3", d3
#print "D", d_loss
# GEN UPDATE
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
#print "LOSSES", d_loss, g_loss, batch_no, i, len(loaded_data['image_list'])/ args.batch_size
total_d_loss += d_loss
total_g_loss += g_loss
print "epoch {}/{} batch {}/{} - d_loss {} g_loss: {}".format(i, args.epochs, batch_no+1, num_batches, d_loss, g_loss)
#print "d_loss: {}".format(d_loss)
#print "g_loss: {}".format(g_loss)
batch_no += 1
if (batch_no % args.save_every) == 0:
print "Saving Images, Model"
save_for_vis(args.data_dir, real_images, gen, image_files)
save_path = saver.save(sess, "Data/Models/latest_model_{}_temp.ckpt".format(args.data_set))
if i%10 == 0:
save_path = saver.save(sess, "Data/Models/model_after_{}_epoch_{}.ckpt".format(args.data_set, i))
total_d_loss /= num_batches
total_g_loss /= num_batches
logger.info('epoch {} d_loss {} g_loss {}'.format(i, total_d_loss, total_g_loss))
if i % 10 == 0:
print "testing"
batch_no = 0
num_batches = int(len(test_data['image_list']) / args.batch_size)
tp = 0
fp = 0
while batch_no*args.batch_size < test_data['data_length']:
real_images, caption_vectors, z_noise = get_test_batch(batch_no, args.batch_size, args.image_size, args.z_dim,
args.caption_vector_length, args.data_dir, args.data_set, test_data)
true_pos = tf.reduce_mean(tf.round(checks['disc_real_image']))
false_pos = tf.reduce_mean(tf.round(checks['disc_fake_image']))
batch_tp, batch_fp = sess.run([true_pos, false_pos],
feed_dict = {
input_tensors['t_real_image']: real_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
}
)
print "Test batch {}/{}".format(batch_no+1, num_batches), "tp:", batch_tp, "fp:", batch_fp
tp += batch_tp
fp += batch_fp
batch_no += 1
tp /= num_batches
fp /= num_batches
print "Test Metrics:", tp, fp
logger.info('epoch {} true_positive {} false_positive {}'.format(i, tp, fp))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=2400,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="Data",
help='Data Directory')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=1000,
help='Max number of epochs')
parser.add_argument(
'--save_every',
type=int,
default=60,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model',
type=str,
default="Data/Models/model_after_epoch_628.ckpt",
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set',
type=str,
default="flowers",
help='Dat set: MS-COCO, flowers')
parser.add_argument('--data_set2',
type=str,
default="birds",
help='Dat set: MS-COCO, flowers')
args = parser.parse_args()
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
d_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['d_loss'],
var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['g_loss'],
var_list=variables['g_vars'])
sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
saver = tf.train.Saver()
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set)
loaded_data2 = load_training_data(args.data_dir, args.data_set2)
for i in range(629, args.epochs):
batch_no_flowers = 0
batch_no_birds = 0
while batch_no_flowers * args.batch_size < loaded_data[
'data_length'] and batch_no_birds * args.batch_size < loaded_data2[
'data_length']:
coin = random.uniform(0, 1)
if batch_no_flowers * args.batch_size > loaded_data['data_length']:
coin = 0
if batch_no_birds * args.batch_size > loaded_data2['data_length']:
coin = 1
if batch_no_flowers * args.batch_size < loaded_data[
'data_length'] and coin > 0.5:
print("Flower batch {}".format(batch_no_flowers))
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no_flowers, args.batch_size, args.image_size,
args.z_dim, args.caption_vector_length, 'train',
args.data_dir, args.data_set, loaded_data)
# DISCR UPDATE
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
_, d_loss, gen, d1, d2, d3 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
# GEN UPDATE
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("LOSSES", d_loss, g_loss, batch_no_flowers, i,
len(loaded_data['image_list']) / args.batch_size)
batch_no_flowers += 1
if (batch_no_flowers % args.save_every) == 0:
print("Saving Images, Model")
save_for_vis(args.data_dir, real_images, gen, image_files,
args.data_set)
# save_path = saver.save(sess, "Data/Models/latest_model_{}_temp.ckpt".format(args.data_set))
if batch_no_birds * args.batch_size < loaded_data2[
'data_length'] and coin <= 0.5:
print("Birds batch {}".format(batch_no_birds))
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no_birds, args.batch_size, args.image_size,
args.z_dim, args.caption_vector_length, 'train',
args.data_dir, args.data_set2, loaded_data2)
# DISCR UPDATE
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
_, d_loss, gen, d1, d2, d3 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
# GEN UPDATE
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("LOSSES", d_loss, g_loss, batch_no_birds, i,
len(loaded_data['image_list']) / args.batch_size)
batch_no_birds += 1
if (batch_no_birds % args.save_every) == 0:
print("Saving Images, Model")
save_for_vis(args.data_dir, real_images, gen, image_files,
args.data_set2)
# save_path = saver.save(sess, "Data/Models/latest_model_{}_temp.ckpt".format(args.data_set2))
if i % 2 == 0:
save_path = saver.save(
sess, "Data/Models/model_after_epoch_{}.ckpt".format(i))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim', type=int, default=100,
help='Noise dimension')
parser.add_argument('--t_dim', type=int, default=256,
help='Text feature dimension')
parser.add_argument('--batch_size', type=int, default=64,
help='Batch Size')
parser.add_argument('--image_size', type=int, default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim', type=int, default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim', type=int, default=64,
help='Number of conv in the first layer discr.')
parser.add_argument('--gfc_dim', type=int, default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length', type=int, default=2400,
help='Caption Vector Length')
parser.add_argument('--data_dir', type=str, default="Data",
help='Data Directory')
parser.add_argument('--learning_rate', type=float, default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1', type=float, default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs', type=int, default=35,
help='Max number of epochs')
parser.add_argument('--save_every', type=int, default=30,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model', type=str, default="Data/Models/model_after_faces_epoch_95.ckpt",
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set', type=str, default="faces",
help='Dat set: MS-COCO, flowers')
args = parser.parse_args()
model_options = {
'z_dim' : args.z_dim,
't_dim' : args.t_dim,
'batch_size' : args.batch_size,
'image_size' : args.image_size,
'gf_dim' : args.gf_dim,
'df_dim' : args.df_dim,
'gfc_dim' : args.gfc_dim,
'caption_vector_length' : args.caption_vector_length
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
with tf.variable_scope(tf.get_variable_scope(), reuse=False):
d_optim = tf.train.AdamOptimizer(args.learning_rate, beta1 = args.beta1).minimize(loss['d_loss'], var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate, beta1 = args.beta1).minimize(loss['g_loss'], var_list=variables['g_vars'])
sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
saver = tf.train.Saver()
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set)
for i in range(args.epochs):
batch_no = 0
list_losses_d=[]
list_losses_g = []
list_batches=[]
while batch_no*args.batch_size < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(batch_no, args.batch_size,
args.image_size, args.z_dim, args.caption_vector_length, 'train', args.data_dir, args.data_set, loaded_data)
# DISCR UPDATE
check_ts = [ checks['d_loss1'] , checks['d_loss2'], checks['d_loss3']]
_, d_loss, gen, d1, d2, d3 = sess.run([d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
print "d1", d1
print "d2", d2
print "d3", d3
print "D", d_loss
# GEN UPDATE
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
list_losses_d.append(d_loss)
list_losses_g.append(g_loss)
list_batches.append(batch_no)
print "LOSSES", d_loss, g_loss, batch_no, i, len(loaded_data['image_list'])/ args.batch_size
batch_no += 1
if (batch_no % args.save_every) == 0:
print "Saving Images, Model"
save_for_vis(args.data_dir, real_images, gen, image_files)
save_path = saver.save(sess, "Data/Models/latest_model_{}_temp.ckpt".format(args.data_set))
if i%5 == 0:
save_path = saver.save(sess, "Data/Models/model_after_{}_epoch_{}.ckpt".format(args.data_set, i))
with open("Data/plots/losses_discriminator_epoch_{}.txt".format(i), 'w') as f:
for s in list_losses_d:
f.write(str(s) + '\n')
with open("Data/plots/losses_generator_epoch_{}.txt".format(i), 'w') as f:
for s in list_losses_g:
f.write(str(s) + '\n')
datasets_root_dir = 'datasets'
loaded_data = load_training_data(datasets_root_dir, 'flowers', 512, 312)
model_options = {
'z_dim': 100,
't_dim': 256,
'batch_size': 64,
'image_size': 128,
'gf_dim': 64,
'df_dim': 64,
'caption_vector_length': 512,
'n_classes': 312
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
sessdefault = tf.compat.v1.InteractiveSession()
sessflowers = tf.compat.v1.InteractiveSession()
sessbirds = tf.compat.v1.InteractiveSession()
sessanime = tf.compat.v1.InteractiveSession()
tf.compat.v1.initialize_all_variables().run()
saver = tf.compat.v1.train.Saver(max_to_keep=10000)
logging.info('Trying to resume model from ' + str(tf.train.latest_checkpoint('./checkpoints/')))
if tf.train.latest_checkpoint('./checkpoints/default/') is not None:
saver.restore(sessdefault, tf.train.latest_checkpoint('./checkpoints/default/'))
logging.info('Successfully loaded model from ./checkpoints/default/')
else:
def main():
print('lol')
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim', type=int, default=100,
help='Noise dimension')
parser.add_argument('--t_dim', type=int, default=256,
help='Text feature dimension')
parser.add_argument('--batch_size', type=int, default=64,
help='Batch Size')
parser.add_argument('--image_size', type=int, default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim', type=int, default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim', type=int, default=64,
help='Number of conv in the first layer discr.')
parser.add_argument('--gfc_dim', type=int, default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length', type=int, default=4800,
help='Caption Vector Length')
parser.add_argument('--data_dir', type=str, default="/media/ssd_working_space/osaid/Data",
help='Data Directory')
parser.add_argument('--learning_rate', type=float, default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1', type=float, default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs', type=int, default=200,
help='Max number of epochs')
parser.add_argument('--save_every', type=int, default=30,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model', type=str, default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set', type=str, default="face",
help='Dat set: MS-COCO, flowers')
args = parser.parse_args()
model_options = {
'z_dim' : args.z_dim,
't_dim' : args.t_dim,
'batch_size' : args.batch_size,
'image_size' : args.image_size,
'gf_dim' : args.gf_dim,
'df_dim' : args.df_dim,
'gfc_dim' : args.gfc_dim,
'caption_vector_length' : args.caption_vector_length
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
d_optim = tf.train.AdamOptimizer(args.learning_rate, beta1 = args.beta1).minimize(loss['d_loss'], var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate, beta1 = args.beta1).minimize(loss['g_loss'], var_list=variables['g_vars'])
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
saver = tf.train.Saver()
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set)
for i in range(args.epochs):
batch_no = 0
while batch_no*args.batch_size < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(batch_no, args.batch_size,
args.image_size, args.z_dim, args.caption_vector_length, 'train', args.data_dir, args.data_set, loaded_data)
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
_, g_loss, fake_img = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
# DISCR UPDATE
if(batch_no%4==0):
check_ts = [ checks['d_loss1'] , checks['d_loss2'], checks['d_loss3']]
_, d_loss, gen, d1, d2, d3 = sess.run([d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict = {
input_tensors['t_real_image'] : fake_img,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
else:
check_ts = [ checks['d_loss1'] , checks['d_loss2'], checks['d_loss3']]
_, d_loss, gen, d1, d2, d3 = sess.run([d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
# print('here')
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
print("LOSSES", d_loss, g_loss, batch_no, i, len(loaded_data['image_list'])/ args.batch_size)
batch_no += 1
if (batch_no % args.save_every) == 0:
print("Saving Images, Model")
save_for_vis(i,batch_no,args.data_dir, real_images, gen, image_files)
save_path = saver.save(sess, "/media/ssd_working_space/osaid/Data/Models/latest_model_{}_temp.ckpt".format(args.data_set))
if i%5 == 0:
save_path = saver.save(sess, "/media/ssd_working_space/osaid/Data/Models/model_after_{}_epoch_{}.ckpt".format(args.data_set, i))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
"""
=====================================================================================================
yokuwakarann nannyanenn 256tte caption vector tono tigai #toha
=====================================================================================================
"""
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=2400,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="Data",
help='Data Directory')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=1,
help='Max number of epochs')
parser.add_argument(
'--save_every',
type=int,
default=100,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model',
type=str,
default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set',
type=str,
default="flowers",
help='Dat set: ImageNet, MS-COCO, flowers')
args = parser.parse_args()
# reference model.py self.options
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
_save_path = "Data/Models/" + str(args.epochs) + "_" + str(
args.caption_vector_length) + "CapVecDims" + "_"
print("== finish checking GAN_model_options")
input_tensors, variables, loss, outputs, checks = gan.build_model()
print("== finish building GAN_model")
d_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['d_loss'],
var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['g_loss'],
var_list=variables['g_vars'])
print("== finish initing the optimizer")
sess = tf.InteractiveSession()
print("== Now runing...")
tf.initialize_all_variables().run()
saver = tf.train.Saver()
if args.resume_model:
print("== Now loading the saved model...")
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set)
print(len(loaded_data))
print("== Finish loading the training data")
for i in range(args.epochs):
batch_no = 0
while batch_no * args.batch_size < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no, args.batch_size, args.image_size, args.z_dim,
args.caption_vector_length, 'train', args.data_dir,
args.data_set, loaded_data)
# DISCR UPDATE
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
_, d_loss, gen, d1, d2, d3 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
# GEN UPDATE
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("LOSSES (D/G)", d_loss, "/", g_loss, "batch_no:", batch_no,
"/",
len(loaded_data['image_list']) / args.batch_size, "Ittr:", i)
batch_no += 1
if (batch_no % args.save_every) == 0:
print("Saving Model")
save_for_vis(args.data_dir, real_images, gen, image_files)
save_path = saver.save(
sess, "Data/Models/" + str(args.data_set) +
"/latest_model_{}_temp.ckpt".format(args.data_set))
if i % 5 == 0:
save_path = saver.save(
sess, "Data/Models/" + str(args.data_set) +
"/model_after_{}_epoch_{}.ckpt".format(args.data_set, i))
#_save_path="Data/Models/"+args.epochs+"_"+args.caption_vector_length+"Captiondims"+"_"
if args.data_set == "flowers":
_save_model_fullpath = _save_path + "model.ckpt"
elif args.data_set == "ImageNet":
_save_model_fullpath = _save_path + "ImageNet_model.ckpt"
saver.save(sess, _save_model_fullpath)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=2400,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="Data",
help='Data Directory')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=600,
help='Max number of epochs')
parser.add_argument(
'--save_every',
type=int,
default=30,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model',
type=str,
default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set',
type=str,
default="flowers",
help='Dat set: MS-COCO, flowers')
parser.add_argument('--model', type=str, default="gan", help='Type of GAN')
parser.add_argument('--lam',
type=float,
default=0.25,
help='lambda of O-GAN')
args = parser.parse_args()
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length,
'lam': args.lam
}
if args.model == 'o-gan':
gan = o_gan_model.O_GAN(model_options)
else:
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
tf.summary.scalar('g_loss', loss['g_loss'])
tf.summary.scalar('d_loss', loss['d_loss'])
with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
d_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['d_loss'],
var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['g_loss'],
var_list=variables['g_vars'])
sess = tf.InteractiveSession()
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter("path")
writer.add_graph(sess.graph)
tf.initialize_all_variables().run()
saver = tf.train.Saver()
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set)
for i in range(args.epochs):
batch_no = 0
while batch_no * args.batch_size < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no, args.batch_size, args.image_size, args.z_dim,
args.caption_vector_length, 'train', args.data_dir,
args.data_set, loaded_data)
# DISCR UPDATE
if args.model == 'gan':
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
_, d_loss, gen, d1, d2, d3 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
else:
check_ts = [
checks['mean_disc_real_image'],
checks['mean_wrong_real_image'],
checks['mean_disc_fake_image'], checks['z1_corr'],
checks['z2_corr']
]
_, d_loss, gen, md_real_image, mw_real_image, md_fake_image, z1, z2 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("md_real_image", md_real_image[0])
print("mw_real_image", mw_real_image[0])
print("md_fake_image", md_fake_image[0])
# print("z1", z1)
# print("z2", z2)
print("D", d_loss)
# GEN UPDATE
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen, result = sess.run(
[g_optim, loss['g_loss'], outputs['generator'], merged],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
writer.add_summary(result,
i * loaded_data['data_length'] + batch_no)
print("LOSSES", d_loss, g_loss, batch_no, i,
len(loaded_data['image_list']) / args.batch_size)
batch_no += 1
if (batch_no % args.save_every) == 0:
print("Saving Images, Model")
save_for_vis(args.data_dir, args.model, real_images, gen,
image_files)
save_path = saver.save(
sess, "Data/Models/{}/latest_model_{}_temp.ckpt".format(
args.model, args.data_set))
if i % 5 == 0:
save_path = saver.save(
sess, "Data/Models/{}/model_after_{}_epoch_{}.ckpt".format(
args.model, args.data_set, i))
writer.close()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=2400,
help='Caption Vector Length')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--gen_updates',
type=int,
default=10,
help='Generator updates per discriminator update')
parser.add_argument('--epochs',
type=int,
default=200,
help='Max number of epochs')
parser.add_argument('--resume_model',
type=str,
default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--resume_epoch',
type=int,
default=0,
help='Number of epochs already trained')
parser.add_argument('--image_dir',
type=str,
default="Data/mscoco_raw/processed",
help='Directory of image')
parser.add_argument('--experiment',
type=str,
default="default",
help='Experiment to save to and load captions for')
parser.add_argument('--style_gan',
type=bool,
default=False,
help='adds style loss to generator loss')
parser.add_argument('--transfer',
action='store_true',
help='does transfer learning')
parser.add_argument('--split',
type=str,
default="train",
help='use val for validation set, train for train\
mostly a debug flag')
parser.add_argument('--extra_32',
action='store_true',
help='extra conv layer when the image is at size 32')
parser.add_argument('--extra_64',
action='store_true',
help='extra conv layer when the image is at size 64')
parser.add_argument('--vgg',
action='store_true',
help='use vgg like layout')
args = parser.parse_args()
if args.vgg and (args.extra_32 or args.extra_64):
raise Exception(
"Cannot perform both vgg and extra_x mods at the same time")
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length,
'extra_32': args.extra_32,
'extra_64': args.extra_64,
'vgg': args.vgg,
'style_gan': args.style_gan
}
tbdir = "Data/Experiments/{}/".format(args.experiment)
tbpath = os.path.join(tbdir, "tensorboard")
if not os.path.isdir(tbdir):
os.makedirs(tbdir)
if not os.path.isdir(tbpath):
os.makedirs(tbpath)
tbwriter = tf.summary.FileWriter(tbpath)
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
d_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['d_loss'],
var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['g_loss'],
var_list=variables['g_vars'])
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
checkpointer = tf.train.Saver()
perm_saver = tf.train.Saver(max_to_keep=None)
if args.transfer:
transfer_learning(sess)
if args.resume_model:
checkpointer.restore(sess, args.resume_model)
loaded_data = load_training_data(args.split, args.experiment)
for i in range(args.resume_epoch, args.epochs + 1):
batch_no = 0
gen_images = None
random.shuffle(loaded_data['image_list'])
print(loaded_data['data_length'])
while batch_no * args.batch_size < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no, args.batch_size, args.image_size, args.z_dim,
args.caption_vector_length, args.image_dir, loaded_data)
# DISCR UPDATE
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
_, d_loss, gen, d1, d2, d3 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
g_loss = None
for _ in range(args.gen_updates):
# GEN UPDATE
_, g_loss, gen_images = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
summary = tf.Summary(value=[
tf.Summary.Value(tag="d_loss", simple_value=d_loss),
tf.Summary.Value(tag="d_loss1", simple_value=d1),
tf.Summary.Value(tag="d_loss2", simple_value=d2),
tf.Summary.Value(tag="d_loss3", simple_value=d3),
tf.Summary.Value(tag="g_loss", simple_value=g_loss)
])
global_step = i * loaded_data[
'data_length'] / args.batch_size + batch_no
tbwriter.add_summary(summary, global_step)
print("Epoch", i, "LOSSES", d_loss, g_loss, batch_no, i,
loaded_data['data_length'] / args.batch_size)
batch_no += 1
checkpointer.save(
sess,
"Data/Experiments/{}/model/checkpoint.ckpt".format(
args.experiment),
global_step=i)
if i > 0 and (i % 100) == 0:
print("Saving Images, Model")
save_for_vis(args.experiment, gen_images)
perm_saver.save(
sess, "Data/Experiments/{}/model/after_{}_epochs.ckpt".format(
args.experiment, i))
#!/usr/bin/env python3
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=23,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="hw3_data",
help='Data Directory')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.7,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=800,
help='Max number of epochs')
parser.add_argument(
'--save_every',
type=int,
default=100,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model',
type=str,
default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set',
type=str,
default="bonus",
help='Dat set: faces,bonus')
parser.add_argument(
'--vector',
type=int,
default=3,
help=
'method to encode captions, options: 1. uni-skip, 2. bi-skip, 3. combine-skip, 4. one-hot, 5. glove_50 , 6. glove_100 , 7. glove_200 , 8. glove_300'
)
parser.add_argument(
'--update_rate',
type=str,
default='1_2',
help='update rate between discrimminator and generator')
parser.add_argument('--gan_type',
type=int,
default=0,
help='GAN type: 0->DCGAN, 1->WGAN, 2->LSGAN, 3->BSGAN')
args = parser.parse_args()
#check if the caption vector length is correct:
if args.vector == 1:
args.caption_vector_length = 2400
if args.vector == 2:
args.caption_vector_length = 2400
if args.vector == 3:
args.caption_vector_length = 4800
if args.vector == 4:
args.caption_vector_length = 23
if args.vector == 5:
args.caption_vector_length = 100
if args.vector == 6:
args.caption_vector_length = 200
if args.vector == 7:
args.caption_vector_length = 400
if args.vector == 8:
args.caption_vector_length = 600
print(args.caption_vector_length)
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length,
'gan_type': args.gan_type
}
#GAN model
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
with tf.variable_scope(tf.get_variable_scope(), reuse=False):
if args.gan_type == 1: # WGAN
d_optim = tf.train.RMSPropOptimizer(
args.learning_rate,
beta1=args.beta1).minimize(loss['d_loss'],
var_list=variables['d_vars'])
g_optim = tf.train.RMSPropOptimizer(
args.learning_rate,
beta1=args.beta1).minimize(loss['g_loss'],
var_list=variables['g_vars'])
else:
d_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['d_loss'],
var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['g_loss'],
var_list=variables['g_vars'])
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 1.0
sess = tf.InteractiveSession(config=config)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(max_to_keep=None)
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set, args.vector)
for i in range(args.epochs):
batch_no = 0
while batch_no * args.batch_size < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no, args.batch_size, args.image_size, args.z_dim,
args.caption_vector_length, 'train', args.data_dir,
args.data_set, args.vector, loaded_data)
# DISCR UPDATE ( 5 times for WGAN )
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
if args.gan_type == 1: #WGAN
_, d_loss, gen, d1, d2, d3 = sess.run(
[
d_optim, loss['d_loss'], outputs['generator'],
clip_updates['clip_updates1'],
clip_updates['clip_updates2'],
clip_updates['clip_updates3']
] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
else:
_, d_loss, gen, d1, d2, d3 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
if args.gan_type == 1: #WAGN
_, d_loss, gen, d1, d2, d3 = sess.run(
[
d_optim, loss['d_loss'], outputs['generator'],
clip_updates['clip_updates1'],
clip_updates['clip_updates2'],
clip_updates['clip_updates3']
] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
_, d_loss, gen, d1, d2, d3 = sess.run(
[
d_optim, loss['d_loss'], outputs['generator'],
clip_updates['clip_updates1'],
clip_updates['clip_updates2'],
clip_updates['clip_updates3']
] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
_, d_loss, gen, d1, d2, d3 = sess.run(
[
d_optim, loss['d_loss'], outputs['generator'],
clip_updates['clip_updates1'],
clip_updates['clip_updates2'],
clip_updates['clip_updates3']
] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
_, d_loss, gen, d1, d2, d3 = sess.run(
[
d_optim, loss['d_loss'], outputs['generator'],
clip_updates['clip_updates1'],
clip_updates['clip_updates2'],
clip_updates['clip_updates3']
] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
# GEN UPDATE
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
if args.update_rate == '1_2':
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("LOSSES", d_loss, g_loss, batch_no, i,
len(loaded_data['image_list']) / args.batch_size)
batch_no += 1
if (batch_no % args.save_every) == 0:
#print("Saving Images, Model")
save_for_vis(args.data_dir, real_images, gen, image_files,
args.vector, args.update_rate)
save_path = saver.save(
sess,
join(
args.data_dir, vector_name[args.vector - 1],
args.update_rate,
"Models/latest_model_{}_temp.ckpt".format(
args.data_set)))
if i % 40 == 0:
save_path = saver.save(
sess,
join(
args.data_dir, vector_name[args.vector - 1],
args.update_rate,
"Models/model_after_{}_epoch_{}.ckpt".format(
args.data_set, i)))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise Dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=2400,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="./",
help='Data Directory')
parser.add_argument('--model_path',
type=str,
default='Data/Models/latest_faces_model.ckpt',
help='Trained Model Path')
parser.add_argument('--n_images',
type=int,
default=5,
help='Number of Images per Caption')
parser.add_argument('--caption_thought_vectors',
type=str,
default='Data/sample_caption_vectors.hdf5',
help='Caption Thought Vector File')
args = parser.parse_args()
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.n_images,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
_, _, _, _, _ = gan.build_model()
sess = tf.InteractiveSession()
saver = tf.train.Saver()
saver.restore(sess, args.model_path)
input_tensors, outputs = gan.build_generator()
h = h5py.File(args.caption_thought_vectors)
caption_vectors = np.array(h['vectors'])
caption_image_dic = {}
for cn, caption_vector in enumerate(caption_vectors):
caption_images = []
z_n1 = np.random.uniform(-0.5, 0.5, [1, args.z_dim])
z_n2 = np.random.uniform(-10, 10, [args.n_images - 1, args.z_dim])
z_noise = np.append(z_n1, z_n2, axis=0)
caption = [caption_vector[0:args.caption_vector_length]
] * args.n_images
[gen_image] = sess.run(
[outputs['generator']],
feed_dict={
input_tensors['t_real_caption']: caption,
input_tensors['t_z']: z_noise,
})
# print gen_image.shape
caption_images = [
gen_image[i, :, :, :] for i in range(0, args.n_images)
]
caption_image_dic[cn] = caption_images
print "Generated", cn, "captions."
make_sure_path_exists(join(args.data_dir, 'samples'))
for f in os.listdir(join(args.data_dir, 'samples')):
if os.path.isfile(f):
os.unlink(join(args.data_dir, 'samples/' + f))
for cn in range(0, len(caption_vectors)):
caption_images = []
for i, im in enumerate(caption_image_dic[cn]):
im_name = "sample{}_{}.jpg".format(cn, i)
scipy.misc.imsave(
join(args.data_dir, 'samples/{}'.format(im_name)), im)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument(
'--t_dim',
type=int,
default=1024, #default=256,
help='Text feature dimension')
parser.add_argument(
'--batch_size',
type=int,
default=64, #LEECHANGE default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=4,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=4800,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="Data",
help='Data Directory')
parser.add_argument(
'--learning_rate',
type=float,
default=1e-5, #LEECHANGE default=0.0002,
help='Learning Rate')
parser.add_argument(
'--beta1',
type=float,
default=.5, #LEECHANGE default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=6000,
help='Max number of epochs')
parser.add_argument(
'--save_every',
type=int,
default=30,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model',
type=str,
default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set',
type=str,
default="flowers",
help='Dat set: MS-COCO, flowers')
args = parser.parse_args()
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
#d_optim = tf.train.AdamOptimizer(args.learning_rate, beta1 = args.beta1, beta2 = .9).minimize(loss['d_loss'], var_list=variables['d_vars'])
#g_optim = tf.train.AdamOptimizer(args.learning_rate, beta1 = args.beta1, beta2 = .9).minimize(loss['g_loss'], var_list=variables['g_vars'])
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
gvs = optimizer.compute_gradients(loss['d_loss'],
var_list=variables['d_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
d_optim = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['d_loss_full'],
var_list=variables['d_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
d_optim_full = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['d_loss_mid'],
var_list=variables['d_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
d_optim_mid = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['d_loss_small'],
var_list=variables['d_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
d_optim_small = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['d_loss_small_full'],
var_list=variables['d_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
d_optim_small_full = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['g_loss'],
var_list=variables['g_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
g_optim = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['g_loss_full'],
var_list=variables['g_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
g_optim_full = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['g_loss_mid'],
var_list=variables['g_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
g_optim_mid = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['g_loss_small'],
var_list=variables['g_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
g_optim_small = optimizer.apply_gradients(capped_gvs)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=.9)
gvs = optimizer.compute_gradients(loss['g_loss_small_mid'],
var_list=variables['g_vars'])
capped_gvs = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gvs
if grad is not None]
g_optim_small_mid = optimizer.apply_gradients(capped_gvs)
sess = tf.InteractiveSession()
if prince:
sess.run(tf.global_variables_initializer())
else:
tf.initialize_all_variables().run()
saver = tf.train.Saver()
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set)
init = 1
d_avg_full, d_avg_mid, d_avg_sml = 0, 0, 0
lb = 0 #-.3
disc_break = -.3
for i in range(args.epochs):
img_idx = 0
batch_no = 0
while batch_no * args.batch_size < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no, args.batch_size, args.image_size, args.z_dim,
args.caption_vector_length, 'train', args.data_dir,
args.data_set, loaded_data)
# DISCR UPDATE
check_ts = [
checks['d_loss_full'], checks['d_loss_mid'],
checks['d_loss_small'], checks['g_loss_full'],
checks['g_loss_mid'], checks['g_loss_small']
]
'''
sess.run([g_optim],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
input_tensors['l2reg']: 0
})
sess.run([g_optim_full],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
input_tensors['l2reg']: 0
})
running = []
if d_avg_full < lb:
running += [g_optim_full]
if d_avg_mid < lb:
running += [g_optim_mid]
if d_avg_sml < lb:
running += [g_optim_small]
if d_avg_full + d_avg_mid + d_avg_sml > 12:
running += [d_optim]
sess.run(running,
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
input_tensors['l2reg']: 0
})
elif len(running) > 0:
sess.run(running,
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
input_tensors['l2reg']: 0
})
'''
if disc_break > 0:
disc_break -= 1
sess.run(
[g_optim],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
input_tensors['l2reg']: 0
})
else:
_, _, d_loss, gen, gen_small, gen_mid, real_small, real_mid, g_loss, df, dm, ds, gf, gm, gs = sess.run(
[
g_optim, d_optim, loss['d_loss'], outputs['generator'],
outputs['generator_small_image'],
outputs['generator_mid_image'],
outputs['real_small_image'], outputs['real_mid_image'],
loss['g_loss']
] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
input_tensors['l2reg']: 0
})
df /= 10.
dm /= 10.
ds /= 10.
gf *= 10.
gm *= 10.
gs *= 10.
d_avg_full = d_avg_full * .8 + .2 * (df - gf)
d_avg_mid = d_avg_mid * .8 + .2 * (dm - gm)
d_avg_sml = d_avg_sml * .8 + .2 * (ds - gs)
if d_avg_full + d_avg_mid + d_avg_sml < 0:
disc_break = np.abs(
int(d_avg_full + d_avg_mid + d_avg_sml)) * 3 + 1
if 1: #batch_no % 2 == 0:
idx = np.random.randint(1, 10)
img_full = gen[idx, :, :, :]
img_small = gen_small[idx, :, :, :]
r_small = real_small[idx, :, :, :]
img_mid = gen_mid[idx, :, :, :]
r_mid = real_mid[idx, :, :, :]
scipy.misc.imsave(
'output_images/' + str(i) + '_img_idx:' +
str(batch_no) + 'full.jpg', img_full)
scipy.misc.imsave(
'output_images/' + str(i) + '_img_idx:' +
str(batch_no) + 'tiny.jpg', img_small)
scipy.misc.imsave(
'output_images/' + str(i) + '_img_idx:' +
str(batch_no) + 'mid.jpg', img_mid)
scipy.misc.imsave(
'output_images/' + str(i) + '_img_idx:' +
str(batch_no) + 'a_r_small.jpg', r_small)
scipy.misc.imsave(
'output_images/' + str(i) + '_img_idx:' +
str(batch_no) + 'a_r_mid.jpg', r_mid)
img_idx += 1
'''
sess.run([g_optim_full],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
input_tensors['l2reg']: 0
})
'''
print 'd_loss_full', df
print 'd_loss_mid', dm
print 'd_loss_small', ds
print 'g_loss_full', gf
print 'g_loss_mid', gm
print 'g_loss_small', gs
print "D", d_loss
print 'd_avg_full', d_avg_full
print 'd_avg_mid', d_avg_mid
print 'd_avg_sml', d_avg_sml
print "LOSSES", d_loss, g_loss, batch_no, i, len(
loaded_data['image_list']) / args.batch_size
batch_no += 1
if (batch_no % args.save_every) == 0:
print "Saving Images, Model"
#Lee commented the following line out because it crashed. No idea what it was trying to do.
#save_for_vis(args.data_dir, real_images, gen, image_files)
save_path = saver.save(
sess, "Data/Models/latest_model_1_{}_temp.ckpt".format(
args.data_set))
if i % 5 == 0:
save_path = saver.save(
sess, "Data/Models/model_after_{}_epoch_{}.ckpt".format(
args.data_set, i))
def main():
z_dim = 100
t_dim = 256
image_size = 64
model_type = 'gan'
n_images = 5
# model Parameters
gf_dim = 64
df_dim = 64
# fully connected
gfc_dim = 1024
caption_vector_length = 600
MODEL_DIR = 'gan-models'
MODEL_NAME = 'final_model.ckpt'
OUTPUT_DIR = 'samples'
caption_vectors_name = 'test_caption_vectors.hdf5'
model_options = {
'z_dim': z_dim,
't_dim': t_dim,
'batch_size': n_images,
'image_size': image_size,
'gf_dim': gf_dim,
'df_dim': df_dim,
'gfc_dim': gfc_dim,
'caption_vector_length': caption_vector_length
}
gan = model.GAN(model_options)
_, _, _, _, _ = gan.build_model()
sess = tf.Session()
saver = tf.train.Saver()
saver.restore(sess, join(MODEL_DIR, model_type, 'Models', MODEL_NAME))
input_tensors, outputs = gan.build_generator()
h = h5py.File(join(MODEL_DIR, model_type, caption_vectors_name))
caption_image_dic = {}
for i, key in enumerate(h):
np.random.seed(1005)
caption_images = []
z_noise = np.random.uniform(0, 0, [n_images, z_dim])
caption = np.array(
[h[key][0, :caption_vector_length] for i in range(n_images)])
[gen_image] =\
sess.run([outputs['generator']],
feed_dict = {input_tensors['t_real_caption'] : caption,
input_tensors['t_z'] : z_noise} )
caption_image_dic[key] =\
[gen_image[i, :, :, :] for i in range(0, n_images)]
if not os.path.exists(OUTPUT_DIR):
os.makedirs(OUTPUT_DIR)
for key in h:
for i, im in enumerate(caption_image_dic[key]):
scipy.misc.imsave(
join(OUTPUT_DIR, 'sample_' + key + '_' + str(i + 1) + '.jpg'),
im)
import datasetloader
import cifarloader
run_config = tf.ConfigProto()
run_config.gpu_options.allow_growth = True
# run optimization
with tf.Session(config=run_config) as sess:
# load options
opt = options.option_initialize()
# load dataset
if opt.dataseyt == cifar:
data_loader = cifarloader.cifarDataLoader(
opt) #for image generation using cifar
else:
data_loader = datasetloader.datasetloader(
opt) #for image generation using celebA or imagenet
Gan = model.GAN(opt, sess)
batch_iters = data_loader.batch_iters()
print('batch_iters:', batch_iters)
if opt.train == False:
Gan.sampling()
else:
epochs = (opt.iters - Gan.counter) // batch_iters
for epoch in range(epochs):
for idx in range(0, batch_iters):
batch_images, batch_z = data_loader.Loaddata(idx)
Gan.optimization(batch_images, batch_z)
print(Gan.mean)
print(max(Gan.mean))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise Dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=2400,
help='Caption Vector Length')
parser.add_argument('--n_images',
type=int,
default=4,
help='Number of Images per Caption')
parser.add_argument('--split',
type=str,
default='gen',
help='train/val/test/gen')
parser.add_argument('--experiment',
type=str,
default="default",
help='Experiment of dataset')
parser.add_argument(
'--epoch',
type=int,
default=None,
help='Epoch of the trained model to load. Defaults to latest checkpoint'
)
parser.add_argument('--transfer',
action='store_true',
help='does transfer learning')
parser.add_argument('--extra_32',
action='store_true',
help='extra conv layer when the image is at size 32')
parser.add_argument('--extra_64',
action='store_true',
help='extra conv layer when the image is at size 64')
parser.add_argument('--vgg',
action='store_true',
help='use vgg like layout')
args = parser.parse_args()
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.n_images,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length,
'extra_32': args.extra_32,
'extra_64': args.extra_64,
'vgg': args.vgg
}
data_dir = os.path.join("Data", "Experiments",
"{}".format(args.experiment))
caption_thought_vectors = os.path.join(
data_dir, '{}_captions.hdf5'.format(args.split))
save_dir = os.path.join(data_dir, "{}_samples".format(args.split))
model_path = os.path.join(data_dir, "model")
checkpoint = tf.train.latest_checkpoint(model_path)
if args.epoch is not None:
checkpoint = os.path.join(model_path,
"after_{}_epochs.ckpt".format(args.epoch))
save_dir = os.path.join(save_dir, "epoch_{}".format(args.epoch))
else:
save_dir = os.path.join(save_dir, "latest".format(args.epoch))
gan = model.GAN(model_options)
_, _, _, _, _ = gan.build_model()
sess = tf.InteractiveSession()
saver = tf.train.Saver()
saver.restore(sess, checkpoint)
input_tensors, outputs = gan.build_generator()
h = h5py.File(caption_thought_vectors)
caption_vectors = {}
if args.split == 'gen':
caption_vectors["generated"] = np.array(h['vectors'])
else:
class_name = list(h.keys())[0]
for img_file, vector in h[class_name].items():
img_id = os.path.splitext(img_file)[0]
caption_vectors[img_id] = np.array(vector)
generated_images = {}
for img_id, vectors in caption_vectors.items():
caption_image_dic = {}
for cn, caption_vector in enumerate(vectors):
caption_images = []
z_noise = 1 * np.random.uniform(-1, 1, [args.n_images, args.z_dim])
caption = [caption_vector[0:args.caption_vector_length]
] * args.n_images
[gen_image] = sess.run(
[outputs['generator']],
feed_dict={
input_tensors['t_real_caption']: caption,
input_tensors['t_z']: z_noise,
})
caption_images = [
gen_image[i, :, :, :] for i in range(0, args.n_images)
]
caption_image_dic[cn] = caption_images
print("Generated {} images for {}".format(cn, img_id))
generated_images[img_id] = caption_image_dic
if os.path.isdir(save_dir):
shutil.rmtree(save_dir)
os.makedirs(save_dir, exist_ok=True)
for img_id, caption_image_dic in generated_images.items():
for cn, images in caption_image_dic.items():
for i, im in enumerate(images):
scipy.misc.imsave(
join(
save_dir,
'{}_{}_image_{}_{}.jpg'.format(img_id,
args.image_size, cn,
chr(ord('A') + i))), im)
def main():
# argunment parser variable.
parser = argparse.ArgumentParser()
# add the following as possible arguments that can be passed while running the file.
# [
# d_dim (Noise dimension)
# t_dim (Text feature dimension)
# batch_size (No of images used in training during iterations)
# gf_dim (neurons in generators first layer)
# df_dim (neurons in discriminators first layer)
# data_dir (Path to data directory)
# learning rate
# beta1 (momentum value for adam update)
# epochs (Number of epochs) **10 epoch take around 6-7 hours**
# save_every (number of iterations over which the model is saved)
# resume_model (to resume the training of a model from file)
# data_set (which data set to train on)
# ]
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=2400,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="Data",
help='Data Directory')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=600,
help='Max number of epochs')
parser.add_argument(
'--save_every',
type=int,
default=30,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model',
type=str,
default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set',
type=str,
default="flowers",
help='Which data set?')
args = parser.parse_args()
# Dict defining the model properties depending upon the command line arguments.
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
# Call the model.GAN function from the model file and pass the above dictionary to create a model based on those properties.
gan = model.GAN(model_options)
# "gan" is the handle to that model for rest of the code
# Unpacking the values sent by build_model() function
input_tensors, variables, loss, outputs, checks = gan.build_model()
# Based on loss recieved from gan.build_model() use adam optimizer to minimize the loss
with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
d_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['d_loss'],
var_list=variables['d_vars'])
with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
g_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['g_loss'],
var_list=variables['g_vars'])
# Initialize all variables
sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
# If resuming a trained model for further training
saver = tf.train.Saver()
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set)
j = 0 #To keep track of iterations
# For "args.epochs" number of epochs---
for i in range(args.epochs):
batch_no = 0
print("Batch size: ", args.batch_size)
print("loaded_data['data_length']: ",
loaded_data['data_length']) #6000
while batch_no * args.batch_size < loaded_data['data_length']:
print("batch_no:", batch_no + 1, "iteration_no:", j + 1, "epoch:",
i + 1)
# Create a training batch which is fed into the dicriminator in the current batch.
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no, args.batch_size, args.image_size, args.z_dim,
args.caption_vector_length, 'train', args.data_dir,
args.data_set, loaded_data)
# DISCR UPDATE
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
# Feed in input from the training batch using feed_dict to the placeholders
_, d_loss, gen, d1, d2, d3 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# Print the discriminator losses
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
# GEN UPDATE
# Feed in input from the training batch using feed_dict to the placeholders
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# Print final loss of current batch
print("LOSSES", d_loss, g_loss, batch_no, i,
len(loaded_data['image_list']) / args.batch_size, "\n")
batch_no += 1
j += 1
# Regularly save the network
if (batch_no % args.save_every) == 0:
print("Saving Images, Model", "\n\n")
save_for_vis(args.data_dir, real_images, gen, image_files)
save_path = saver.save(
sess, "Data/Models/latest_model_{}_temp.ckpt".format(
args.data_set))
if i % 5 == 0:
save_path = saver.save(
sess, "Data/Models/model_after_{}_epoch_{}.ckpt".format(
args.data_set, i))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise Dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=768,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="Data",
help='Data Directory')
parser.add_argument(
'--model_path',
type=str,
default='Data/Models/latest_bert_model_flowers_temp.ckpt',
help='Trained Model Path')
parser.add_argument('--n_images',
type=int,
default=5,
help='Number of Images per Caption')
parser.add_argument('--caption_thought_vectors',
type=str,
default='Data/submission_bert_caption_vectors.hdf5',
help='Caption Thought Vector File')
# conda install -c conda-forge tensorflow==1.15
args = parser.parse_args()
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.n_images,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
_, _, _, _, _ = gan.build_model()
sess = tf.InteractiveSession()
saver = tf.train.Saver()
saver.restore(sess, args.model_path)
input_tensors, outputs = gan.build_generator()
h = h5py.File(args.caption_thought_vectors)
caption_vectors = np.array(h['vectors'])
caption_image_dic = {}
for cn, caption_vector in enumerate(caption_vectors):
caption_images = []
z_noise = np.random.uniform(-1, 1, [args.n_images, args.z_dim])
caption = [caption_vector[0:args.caption_vector_length]
] * args.n_images
[gen_image] = sess.run(
[outputs['generator']],
feed_dict={
input_tensors['t_real_caption']: caption,
input_tensors['t_z']: z_noise,
})
caption_images = [
gen_image[i, :, :, :] for i in range(0, args.n_images)
]
caption_image_dic[cn] = caption_images
print("Generated", cn)
for f in os.listdir(join(args.data_dir, 'val_samples')):
if os.path.isfile(f):
os.unlink(join(args.data_dir, 'val_samples/' + f))
for cn in range(0, len(caption_vectors)):
caption_images = []
for i, im in enumerate(caption_image_dic[cn]):
# im_name = "caption_{}_{}.jpg".format(cn, i)
# scipy.misc.imsave( join(args.data_dir, 'val_samples/{}'.format(im_name)) , im)
caption_images.append(im)
caption_images.append(np.zeros((64, 5, 3)))
combined_image = np.concatenate(caption_images[0:-1], axis=1)
imageio.imwrite(
join(args.data_dir,
'val_samples/combined_image_{}.jpg'.format(cn)),
combined_image)
def map_fn(index=None, flags=None):
torch.set_default_tensor_type('torch.FloatTensor')
torch.manual_seed(1234)
train_data = dataset.DATA(config.TRAIN_DIR)
if config.MULTI_CORE:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_data,
num_replicas=xm.xrt_world_size(),
rank=xm.get_ordinal(),
shuffle=True)
else:
train_sampler = torch.utils.data.RandomSampler(train_data)
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=flags['batch_size']
if config.MULTI_CORE else config.BATCH_SIZE,
sampler=train_sampler,
num_workers=flags['num_workers'] if config.MULTI_CORE else 4,
drop_last=True,
pin_memory=True)
if config.MULTI_CORE:
DEVICE = xm.xla_device()
else:
DEVICE = config.DEVICE
netG = model.colorization_model().double()
netD = model.discriminator_model().double()
VGG_modelF = torchvision.models.vgg16(pretrained=True).double()
VGG_modelF.requires_grad_(False)
netG = netG.to(DEVICE)
netD = netD.to(DEVICE)
VGG_modelF = VGG_modelF.to(DEVICE)
optD = torch.optim.Adam(netD.parameters(), lr=2e-4, betas=(0.5, 0.999))
optG = torch.optim.Adam(netG.parameters(), lr=2e-4, betas=(0.5, 0.999))
## Trains
train_start = time.time()
losses = {
'G_losses': [],
'D_losses': [],
'EPOCH_G_losses': [],
'EPOCH_D_losses': [],
'G_losses_eval': []
}
netG, optG, netD, optD, epoch_checkpoint = utils.load_checkpoint(
config.CHECKPOINT_DIR, netG, optG, netD, optD, DEVICE)
netGAN = model.GAN(netG, netD)
for epoch in range(
epoch_checkpoint, flags['num_epochs'] +
1 if config.MULTI_CORE else config.NUM_EPOCHS + 1):
print('\n')
print('#' * 8, f'EPOCH-{epoch}', '#' * 8)
losses['EPOCH_G_losses'] = []
losses['EPOCH_D_losses'] = []
if config.MULTI_CORE:
para_train_loader = pl.ParallelLoader(
train_loader, [DEVICE]).per_device_loader(DEVICE)
engine.train(para_train_loader,
netGAN,
netD,
VGG_modelF,
optG,
optD,
device=DEVICE,
losses=losses)
elapsed_train_time = time.time() - train_start
print("Process", index, "finished training. Train time was:",
elapsed_train_time)
else:
engine.train(train_loader,
netGAN,
netD,
VGG_modelF,
optG,
optD,
device=DEVICE,
losses=losses)
#########################CHECKPOINTING#################################
utils.create_checkpoint(epoch,
netG,
optG,
netD,
optD,
max_checkpoint=config.KEEP_CKPT,
save_path=config.CHECKPOINT_DIR)
########################################################################
utils.plot_some(train_data, netG, DEVICE, epoch)
gc.collect()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--rnn_hidden',
type=int,
default=200,
help='Number of nodes in the rnn hidden layer')
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--word_dim',
type=int,
default=256,
help='Word embedding matrix dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--batch_size', type=int, default=2, help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=32,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=20,
help='Caption Vector Length')
parser.add_argument('--data_dir',
type=str,
default="../data",
help='Data Directory')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=600,
help='Max number of epochs')
parser.add_argument(
'--save_every',
type=int,
default=30,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model',
type=str,
default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set',
type=str,
default="flowers",
help='Dat set: MS-COCO, flowers')
args = parser.parse_args()
model_options = {
'rnn_hidden': args.rnn_hidden,
'word_dim': args.word_dim,
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
with tf.variable_scope(tf.get_variable_scope()) as scope:
input_tensors, variables, loss, outputs, checks = gan.build_model()
# with tf.variable_scope("Adam", reuse=False):
# d_optim = tf.train.AdamOptimizer().minimize(loss['d_loss'], var_list=variables['d_vars'], name='Adam_d')
# g_optim = tf.train.AdamOptimizer().minimize(loss['g_loss'], var_list=variables['g_vars'], name='Adam_g')
d_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['d_loss'],
var_list=variables['d_vars'],
name='Adam_d')
g_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(
loss['g_loss'],
var_list=variables['g_vars'],
name='Adam_g')
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
saver = tf.train.Saver()
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_dir, args.data_set,
args.caption_vector_length,
args.image_size)
for i in range(args.epochs):
print("epoch" + str(i))
batch_no = 0
index4shuffle = [i for i in range(len(loaded_data))]
random.shuffle(index4shuffle)
while (batch_no + 1) * args.batch_size < len(loaded_data):
real_images, wrong_images, caption_vectors, z_noise, image_files = get_training_batch(
batch_no, args.batch_size, args.image_size, args.z_dim,
args.caption_vector_length, 'train', args.data_dir,
args.data_set,
index4shuffle[batch_no * args.batch_size:(batch_no + 1) *
args.batch_size], loaded_data)
print(caption_vectors)
# DISCR UPDATE
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
_, d_loss, gen, d1, d2, d3 = sess.run(
[d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
print("d1", d1)
print("d2", d2)
print("d3", d3)
print("D", d_loss)
# GEN UPDATE
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run(
[g_optim, loss['g_loss'], outputs['generator']],
feed_dict={
input_tensors['t_real_image']: real_images,
input_tensors['t_wrong_image']: wrong_images,
input_tensors['t_real_caption']: caption_vectors,
input_tensors['t_z']: z_noise,
})
# print("LOSSES")
print("d_loss=" + str(d_loss) + ", g_loss=" + str(g_loss))
batch_no += 1
if (batch_no % args.save_every) == 0:
print("Saving Images, Model")
save_for_vis(args.data_dir, real_images, gen, image_files)
save_path = saver.save(
sess, "../models/latest_model_{}_temp.ckpt".format(
args.data_set))
if i % 5 == 0:
save_path = saver.save(
sess, "../models/model_after_{}_epoch_{}.ckpt".format(
args.data_set, i))
def main():
model_options = {
'z_dim' : 100,
't_dim' : 256,
'batch_size' : 64,
'image_size' : 64,
'gf_dim' : 64,
'df_dim' : 64,
'gfc_dim' : 1024,
'caption_vector_length' : 4800
}
beta1 = 0.5
lr = 2e-4
z_dim = 100
t_dim = 256
batch_size = 64
image_size = 64
gfc_dim = 1024
caption_vector_length = 4800
epochs = 600
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
checkpoint_dir = "./DCModel"
d_optim = tf.train.AdamOptimizer(lr*0.5, beta1 = beta1).minimize(loss['d_loss'], var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(lr, beta1 = beta1).minimize(loss['g_loss'], var_list=variables['g_vars'])
#sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state("Data/Models/")
saver.restore(sess, ckpt.model_checkpoint_path)
print("Model restored.")
loaded_data = load_training_data()
start_time = time.time()
for i in range(epochs):
#batch_no = 0
for batch_no in range(loaded_data['data_length'] // batch_size):
#while batch_no* < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise = get_training_batch(batch_no, batch_size, z_dim, loaded_data)
# DISCR UPDATE
check_ts = [ checks['d_loss1'] , checks['d_loss2'], checks['d_loss3']]
_, d_loss, gen, d1, d2, d3 = sess.run([d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise
})
# GEN UPDATE
# GEN UPDATE TWICE, to make sure d_loss does not go to 0
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
_, g_loss, gen = sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
#batch_no += 1
#if batch_no == 171:
# continue
if batch_no == 100:
print "Saving Images, Model"
#print "d1", d1
#print "d2", d2
#print "d3", d3
#print "D", d_loss
print "Epoch: [%2d], [%4d/%4d] d_loss: %.8f, g_loss: %.8f, time: %4.4f" %(i, batch_no, len(loaded_data['image_list'])/ batch_size, d_loss, g_loss,
time.time() - start_time )
save_for_vis(real_images, gen)
saver.save(sess, checkpoint_dir)
if i%3 == 0:
save_path = saver.save(sess, "Data/Models/model_after_epoch_{}.ckpt".format(i))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim', type=int, default=100,
help='Noise dimension')
parser.add_argument('--t_dim', type=int, default=256,
help='Text feature dimension')
parser.add_argument('--batch_size', type=int, default=64,
help='Batch Size')
parser.add_argument('--image_size', type=int, default=128,
help='Image Size a, a x a')
parser.add_argument('--gf_dim', type=int, default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim', type=int, default=64,
help='Number of conv in the first layer discr.')
parser.add_argument('--caption_vector_length', type=int, default=4800,
help='Caption Vector Length')
parser.add_argument('--n_classes', type=int, default=102,
help='Number of classes/class labels')
parser.add_argument('--data_dir', type=str, default="Data",
help='Data Directory')
parser.add_argument('--learning_rate', type=float, default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1', type=float, default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--images_per_caption', type=int, default=30,
help='The number of images that you want to generate '
'per text description')
parser.add_argument('--data_set', type=str, default="flowers",
help='Dat set: MS-COCO, flowers')
parser.add_argument('--checkpoints_dir', type=str, default="/tmp",
help='Path to the checkpoints directory')
args = parser.parse_args()
datasets_root_dir = join(args.data_dir, 'datasets')
loaded_data = load_training_data(datasets_root_dir, args.data_set,
args.caption_vector_length,
args.n_classes)
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'caption_vector_length': args.caption_vector_length,
'n_classes': loaded_data['n_classes']
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
#saver = tf.compat.v1.train.Saver(max_to_keep=1)
print('Trying to resume model from ' +
str(tf.train.latest_checkpoint(args.checkpoints_dir)))
if tf.train.latest_checkpoint(args.checkpoints_dir) is not None:
#saver.restore(sess, tf.train.latest_checkpoint(args.checkpoints_dir))
print('Successfully loaded model from ')
else:
print('Could not load checkpoints. Please provide a valid path to'
' your checkpoints directory')
exit()
print('Starting to generate images from text descriptions.')
for sel_i, text_cap in enumerate(loaded_data['text_caps']['features']):
print('Text idx: {}\nRaw Text: {}\n'.format(sel_i, text_cap))
captions_1, image_files_1, image_caps_1, image_ids_1,\
image_caps_ids_1 = get_caption_batch(loaded_data, datasets_root_dir,
dataset=args.data_set, batch_size=args.batch_size)
captions_1[args.batch_size-1, :] = text_cap
for z_i in range(args.images_per_caption):
z_noise = np.random.uniform(-1, 1, [args.batch_size, args.z_dim])
val_feed = {
input_tensors['t_real_caption'].name: captions_1,
input_tensors['t_z'].name: z_noise,
input_tensors['t_training'].name: True
}
val_gen = sess.run(
[outputs['generator']],
feed_dict=val_feed)
dump_dir = os.path.join(args.data_dir,
'images_generated_from_text')
save_distributed_image_batch(dump_dir, val_gen, sel_i, z_i,
args.batch_size)
print('Finished generating images from text description')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for fully connected layer')
parser.add_argument('--caption_vector_length',
'-cvl',
type=int,
default=600,
help='Caption Vector Length')
parser.add_argument('--method_dir',
'-md',
type=str,
default='',
help='method directory')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=600,
help='Max number of epochs')
parser.add_argument(
'--save_every',
type=int,
default=30,
help='Save Model/Samples every x iterations over batches')
parser.add_argument('--resume_model',
type=str,
default=None,
help='Pre-Trained Model Path, to resume from')
parser.add_argument('--data_set',
type=str,
default='faces',
help='data set: faces')
parser.add_argument('--imgs_dir',
type=str,
default='imgs',
help='images directory')
parser.add_argument('--caption_vectors',
type=str,
default='caption_vectors.hdf5',
help='encoded training caption')
parser.add_argument('--dis_updates',
'-du',
type=int,
default=1,
help='discriminator update per round')
parser.add_argument('--gen_updates',
'-gu',
type=int,
default=2,
help='generator update per round')
args = parser.parse_args()
if args.method_dir == '':
print('need to specify method_dir!')
exit(1)
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
with tf.variable_scope(tf.get_variable_scope(), reuse=False):
d_optim =\
tf.train.AdamOptimizer(\
args.learning_rate, beta1 = args.beta1
).minimize(loss['d_loss'], var_list=variables['d_vars'])
g_optim =\
tf.train.AdamOptimizer(\
args.learning_rate, beta1 = args.beta1
).minimize(loss['g_loss'], var_list=variables['g_vars'])
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver(max_to_keep=None)
if args.resume_model:
saver.restore(sess, args.resume_model)
loaded_data = load_training_data(args.data_set, args.method_dir,
args.imgs_dir, args.caption_vectors)
for i in range(1, args.epochs + 1):
batch_no = 0
while batch_no * args.batch_size < loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files =\
get_training_batch(batch_no, args.batch_size, args.image_size,
args.z_dim, args.caption_vector_length, 'train',
args.method_dir, args.imgs_dir, args.data_set,
loaded_data)
# DISCR UPDATE
for j in range(args.dis_updates):
check_ts = [
checks['d_loss1'], checks['d_loss2'], checks['d_loss3']
]
_, d_loss, gen, d1, d2, d3 =\
sess.run([d_optim, loss['d_loss'], outputs['generator']] + check_ts,
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
print('d1 = {:5f} d2 = {:5f} d3 = {:5f} '
'D = {:5f}'.format(d1, d2, d3, d_loss))
# GEN UPDATE
for j in range(args.gen_updates):
_, g_loss, gen =\
sess.run([g_optim, loss['g_loss'], outputs['generator']],
feed_dict = {
input_tensors['t_real_image'] : real_images,
input_tensors['t_wrong_image'] : wrong_images,
input_tensors['t_real_caption'] : caption_vectors,
input_tensors['t_z'] : z_noise,
})
print('d_loss = {:5f} g_loss = {:5f} batch_no = {} '
'epochs = {}'.format(d_loss, g_loss, batch_no, i))
print('-' * 60)
batch_no += 1
if (batch_no % args.save_every) == 0:
save_for_vis(args.data_set, args.method_dir, real_images, gen,
image_files)
save_path =\
saver.save(sess, join(args.data_set, args.method_dir, 'Models',
'latest_model_'
'{}_temp.ckpt'.format(args.data_set)))
if i % 50 == 0:
save_path =\
saver.save(sess, join(args.data_set,
args.method_dir, 'Models', 'model_after_'
'{}_epoch_{}.ckpt'.format(args.data_set, i)))
def train(args):
height, width, channel = 28, 28, 1
batch_size = args.batch_size
z_size = args.nd # 噪声维数
real_img = tf.placeholder(tf.float32, [batch_size, height, width, channel],
name='img')
z = tf.placeholder(tf.float32, [batch_size, z_size], name='z')
label = tf.placeholder(tf.float32, [batch_size, 10], name='label') # 0~9
gan = model.GAN(height, width, channel)
gan.set_batch_size(batch_size)
fake_img = gan.generator(z, label)
real_result = gan.discriminator(real_img, label, reuse=False)
fake_result = gan.discriminator(fake_img, label, reuse=True)
real = tf.reduce_sum(label * real_result, 1)
fake = tf.reduce_sum(label * fake_result, 1)
d_loss = -tf.reduce_mean(tf.log(real) + tf.log(1. - fake))
g_loss = -tf.reduce_mean(tf.log(fake))
t_vars = tf.trainable_variables()
d_vars = [var for var in t_vars if 'dis' in var.name]
g_vars = [var for var in t_vars if 'gen' in var.name]
d_optimizer = tf.train.AdamOptimizer(learning_rate=0.0002, beta1=0.5) \
.minimize(d_loss, var_list=d_vars)
g_optimizer = tf.train.AdamOptimizer(learning_rate=0.0002, beta1=0.5) \
.minimize(g_loss, var_list=g_vars)
data = DataProvider()
train_num = data.get_train_num()
batch_num = int(train_num / args.batch_size)
saver = tf.train.Saver(max_to_keep=1)
model_dir = args.model_dir
if (not os.path.exists(model_dir)):
os.mkdir(model_dir)
accuracy_real = tf.reduce_mean(
tf.cast(tf.equal(tf.argmax(label, 1), tf.argmax(real_result, 1)),
'float'))
accuracy_fake = tf.reduce_mean(
tf.cast(tf.equal(tf.argmax(label, 1), tf.argmax(fake_result, 1)),
'float'))
with tf.Session() as sess:
counter = 0
sess.run(tf.global_variables_initializer())
for epoch in range(args.epoch):
for batch in range(batch_num):
counter += 1
train_data, label_data = data.next_batch(batch_size)
batch_z = np.random.normal(0, 1, [batch_size, z_size]).astype(
np.float_)
sess.run(d_optimizer,
feed_dict={
real_img: train_data,
z: batch_z,
label: label_data
})
sess.run(g_optimizer,
feed_dict={
z: batch_z,
label: label_data
})
if (counter % 20 == 0):
dloss, gloss, ac_real, ac_fake = sess.run(
[d_loss, g_loss, accuracy_real, accuracy_fake],
feed_dict={
real_img: train_data,
z: batch_z,
label: label_data
})
print('iter:', counter, 'd_loss:', dloss, 'g_loss:', gloss,
'ac_real:', ac_real, 'ac_fake:', ac_fake)
if (counter % 200 == 0):
saver.save(sess, os.path.join(model_dir, 'model'))
def init_random(shape):
return np.random.random_sample(shape)
def next_batch(x, y, batch_size=BATCH_SIZE):
i = 0
while (i < len(x)):
yield x[i:i + batch_size], y[i:i + batch_size]
i = i + batch_size
if __name__ == '__main__':
tmp_buff = open('tmp.out', 'a')
gan = model.GAN()
images, labels = load_data.load_SVHN()
label_ = tf.placeholder(tf.float32, [None, 2])
# sess = tf.Session()
train_step = tf.train.AdamOptimizer(1e-4).minimize(
gan.gan.loss(logit=gan.gan.h_fc8, label=label_))
correct_prediction = tf.equal(tf.argmax(label_, 1),
tf.argmax(gan.gan.h_fc8, 1))
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
for step in range(EPOCH_SIZE):
if step % 10 == 0:
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim', type=int, default=100,
help='Noise dimension')
parser.add_argument('--t_dim', type=int, default=256,
help='Text feature dimension')
parser.add_argument('--batch_size', type=int, default=64,
help='Batch Size')
parser.add_argument('--image_size', type=int, default=128,
help='Image Size a, a x a')
parser.add_argument('--gf_dim', type=int, default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim', type=int, default=64,
help='Number of conv in the first layer discr.')
parser.add_argument('--caption_vector_length', type=int, default=4800,
help='Caption Vector Length')
parser.add_argument('--n_classes', type=int, default=102,
help='Number of classes/class labels')
parser.add_argument('--data_dir', type=str, default="Data",
help='Data Directory')
parser.add_argument('--learning_rate', type=float, default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1', type=float, default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--data_set', type=str, default="flowers",
help='Dat set: flowers')
parser.add_argument('--output_dir', type=str,
default="Data/synthetic_dataset",
help='The directory in which this dataset will be '
'created')
parser.add_argument('--checkpoints_dir', type=str, default="/tmp",
help='Path to the checkpoints directory')
parser.add_argument('--n_interp', type=int, default=100,
help='The difference between each interpolation. '
'Should ideally be a multiple of 10')
parser.add_argument('--n_images', type=int, default=500,
help='Number of images to randomply sample for '
'generating interpolation results')
args = parser.parse_args()
datasets_root_dir = join(args.data_dir, 'datasets')
loaded_data = load_training_data(datasets_root_dir, args.data_set,
args.caption_vector_length, args.n_classes)
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'caption_vector_length': args.caption_vector_length,
'n_classes': loaded_data['n_classes']
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
saver = tf.train.Saver(max_to_keep=10000)
print('resuming model from checkpoint' +
str(tf.train.latest_checkpoint(args.checkpoints_dir)))
if tf.train.latest_checkpoint(args.checkpoints_dir) is not None:
saver.restore(sess, tf.train.latest_checkpoint(args.checkpoints_dir))
print('Successfully loaded model')
else:
print('Could not load checkpoints')
exit()
random.shuffle(loaded_data['image_list'])
selected_images = loaded_data['image_list'][:args.n_images]
cap_id = [0]#[np.random.randint(0, 4) for cap_i in range(len(selected_images))]
print('Generating Images by interpolating z')
bar = progressbar.ProgressBar(redirect_stdout=True,
max_value=args.n_images)
for sel_i, (sel_img, sel_cap) in enumerate(zip(selected_images, cap_id)):
captions, image_files, image_caps, image_ids, image_caps_ids = \
get_images_z_intr(sel_img, sel_cap, loaded_data,
datasets_root_dir, args.data_set, args.batch_size)
z_noise_1 = np.full((args.batch_size, args.z_dim), -1.0)
z_noise_2 = np.full((args.batch_size, args.z_dim), 1.0)
intr_z_list = get_interp_vec(z_noise_1, z_noise_2, args.z_dim,
args.n_interp, args.batch_size)
for z_i, z_noise in enumerate(intr_z_list):
val_feed = {
input_tensors['t_real_caption'].name: captions,
input_tensors['t_z'].name: z_noise,
input_tensors['t_training'].name: True
}
val_gen = sess.run([outputs['generator']], feed_dict=val_feed)
save_distributed_image_batch(args.output_dir, val_gen, sel_i, z_i,
sel_img, sel_cap, args.batch_size)
bar.update(sel_i)
bar.finish()
print('Finished generating interpolated images')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--batch_size',
type=int,
default=64,
help='Batch Size')
parser.add_argument('--image_size',
type=int,
default=128,
help='Image Size a, a x a')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument('--caption_vector_length',
type=int,
default=4800,
help='Caption Vector Length')
parser.add_argument('--n_classes',
type=int,
default=102,
help='Number of classes/class labels')
parser.add_argument('--data_dir',
type=str,
default="Data",
help='Data Directory')
parser.add_argument('--learning_rate',
type=float,
default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1',
type=float,
default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs',
type=int,
default=200,
help='Max number of epochs')
parser.add_argument('--data_set',
type=str,
default="flowers",
help='Dat set: flowers')
parser.add_argument('--output_dir',
type=str,
default="Data/ds",
help='The directory in which this dataset will be '
'created')
parser.add_argument('--checkpoints_dir',
type=str,
default="/tmp",
help='Path to the checkpoints directory')
args = parser.parse_args()
model_stage_1_ds_tr, model_stage_1_ds_val, datasets_root_dir = \
prepare_dirs(args)
loaded_data = load_training_data(datasets_root_dir, args.data_set,
args.caption_vector_length,
args.n_classes)
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'caption_vector_length': args.caption_vector_length,
'n_classes': loaded_data['n_classes']
}
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
sess = tf.InteractiveSession()
tf.initialize_all_variables().run()
saver = tf.train.Saver(max_to_keep=10000)
print('resuming model from checkpoint' +
str(tf.train.latest_checkpoint(args.checkpoints_dir)))
if tf.train.latest_checkpoint(args.checkpoints_dir) is not None:
saver.restore(sess, tf.train.latest_checkpoint(args.checkpoints_dir))
print('Successfully loaded model from ')
else:
print('Could not load checkpoints')
exit()
print('Generating images for the captions in the training set at ' +
model_stage_1_ds_tr)
for i in range(args.epochs):
batch_no = 0
while batch_no * args.batch_size + args.batch_size < \
loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files, \
real_classes, wrong_classes, image_caps, image_ids, \
image_caps_ids = get_training_batch(batch_no, args.batch_size,
args.image_size, args.z_dim, datasets_root_dir,
args.data_set, loaded_data)
feed = {
input_tensors['t_real_image'].name: real_images,
input_tensors['t_wrong_image'].name: wrong_images,
input_tensors['t_real_caption'].name: caption_vectors,
input_tensors['t_z'].name: z_noise,
input_tensors['t_real_classes'].name: real_classes,
input_tensors['t_wrong_classes'].name: wrong_classes,
input_tensors['t_training'].name: True
}
g_loss, gen = sess.run([loss['g_loss'], outputs['generator']],
feed_dict=feed)
print("LOSSES", g_loss, batch_no, i,
len(loaded_data['image_list']) / args.batch_size)
batch_no += 1
save_distributed_image_batch(model_stage_1_ds_tr, gen, image_caps,
image_ids, image_caps_ids)
print('Finished generating images for the training set captions.\n\n')
print('Generating images for the captions in the validation set at ' +
model_stage_1_ds_val)
for i in range(args.epochs):
batch_no = 0
while batch_no * args.batch_size + args.batch_size < \
loaded_data['val_data_len']:
val_captions, val_image_files, val_image_caps, val_image_ids, \
val_image_caps_ids, val_z_noise = get_val_caps_batch(batch_no,
args.batch_size, args.z_dim, loaded_data, args.data_set,
datasets_root_dir)
val_feed = {
input_tensors['t_real_caption'].name: val_captions,
input_tensors['t_z'].name: val_z_noise,
input_tensors['t_training'].name: True
}
val_gen, val_attn_spn = sess.run(
[outputs['generator'], checks['attn_span']],
feed_dict=val_feed)
print("LOSSES", batch_no, i,
len(loaded_data['val_img_list']) / args.batch_size)
batch_no += 1
save_distributed_image_batch(model_stage_1_ds_val, val_gen,
val_image_caps, val_image_ids,
val_image_caps_ids, val_attn_spn)
print('Finished generating images for the validation set captions.\n\n')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise Dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--image_size',
type=int,
default=64,
help='Image Size')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
type=int,
default=4800,
help='Caption Vector Length')
#parser.add_argument('--data_dir', type=str, default="Data",
# help='Data Directory') old code
parser.add_argument('--data_dir',
type=str,
default="/media/ssd_working_space/osaid/Data",
help='Data Directory')
#parser.add_argument('--model_path', type=str, default='Data/Models/latest_model_flowers_temp.ckpt',help='Trained Model Path') old code
parser.add_argument(
'--model_path',
type=str,
default=
'/media/ssd_working_space/osaid/Data/Models/latest_model_face_temp.ckpt',
help='Trained Model Path')
parser.add_argument('--n_images',
type=int,
default=1,
help='Number of Images per Caption')
#parser.add_argument('--caption_thought_vectors', type=str, default='Data/sample_caption_vectors.hdf5', help='Caption Thought Vector File') old code
parser.add_argument(
'--caption_thought_vectors',
type=str,
default=
'/media/ssd_working_space/osaid/Data/sample_caption_vectors.hdf5',
help='Caption Thought Vector File')
args = parser.parse_args()
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.n_images,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
_, _, _, _, _ = gan.build_model()
sess = tf.InteractiveSession()
saver = tf.train.Saver()
saver.restore(sess, args.model_path)
input_tensors, outputs = gan.build_generator()
# h = h5py.File('/media/ssd_working_space/osaid/Data/flower_tv.hdf5')
# image_name = [l for l in h]
# caption_vectors = np.array([h[l][0] for l in h])
test_encoding = pickle.load(
open('/media/ssd_working_space/osaid/celebA/train/encoding', 'rb'))
image_name = [key for key in test_encoding.keys()]
caption_vectors = [
test_encoding[key].reshape(-1) for key in test_encoding.keys()
]
caption_image_dic = {}
for cn, caption_vector in enumerate(caption_vectors):
caption_images = []
z_noise = np.random.uniform(-1, 1, [args.n_images, args.z_dim])
caption = [caption_vector[0:args.caption_vector_length]
] * args.n_images
[gen_image] = sess.run(
[outputs['generator']],
feed_dict={
input_tensors['t_real_caption']: caption,
input_tensors['t_z']: z_noise,
})
caption_images = [
gen_image[i, :, :, :] for i in range(0, args.n_images)
]
caption_image_dic[cn] = caption_images
print("Generated", cn)
#for f in os.listdir( join(args.data_dir, 'val_samples')):
#if os.path.isfile(f):
#os.unlink(join(args.data_dir, 'val_samples/' + f))
for cn in range(0, len(caption_vectors)):
caption_images = []
# for i, im in enumerate( caption_image_dic[ cn ] ):
# im_name = "caption_{}_{}.jpg".format(cn, i)
# scipy.misc.imsave( join(args.data_dir, 'val_samples/{}'.format(im_name)) , im)
# caption_images.append( im )
# caption_images.append( np.zeros((64, 5, 3)) )
# combined_image = np.concatenate( caption_images[0:-1], axis = 1 )
# if cn < 6000:
# #scipy.misc.imsave( join(args.data_dir, 'sr_train_lr/image_{}.jpg'.format(cn)) , caption_image_dic[ cn ][0])
# scipy.misc.imsave( '/media/ssd_working_space/osaid/Data/sr_train_lr/'+image_name[cn] , caption_image_dic[ cn ][0])
# else:
#scipy.misc.imsave( join(args.data_dir, 'sr_test_lr/image_{}.jpg'.format(cn)) , caption_image_dic[ cn ][0])
scipy.misc.imsave(
'/media/ssd_working_space/osaid/Data/infered_train_images/' +
image_name[cn], caption_image_dic[cn][0])
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim', type=int, default=100,
help='Noise dimension')
parser.add_argument('--t_dim', type=int, default=256,
help='Text feature dimension')
parser.add_argument('--batch_size', type=int, default=64,
help='Batch Size')
parser.add_argument('--image_size', type=int, default=128,
help='Image Size a, a x a')
parser.add_argument('--gf_dim', type=int, default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim', type=int, default=64,
help='Number of conv in the first layer discr.')
parser.add_argument('--caption_vector_length', type=int, default=4800,
help='Caption Vector Length')
parser.add_argument('--n_classes', type = int, default = 102,
help = 'Number of classes/class labels')
parser.add_argument('--data_dir', type=str, default="Data",
help='Data Directory')
parser.add_argument('--learning_rate', type=float, default=0.0002,
help='Learning Rate')
parser.add_argument('--beta1', type=float, default=0.5,
help='Momentum for Adam Update')
parser.add_argument('--epochs', type=int, default=200,
help='Max number of epochs')
parser.add_argument('--save_every', type=int, default=30,
help='Save Model/Samples every x iterations over '
'batches')
parser.add_argument('--resume_model', type=bool, default=False,
help='Pre-Trained Model load or not')
parser.add_argument('--data_set', type=str, default="flowers",
help='Dat set: MS-COCO, flowers')
parser.add_argument('--model_name', type=str, default="TAC_GAN",
help='model_1 or model_2')
parser.add_argument('--train', type = bool, default = True,
help = 'True while training and otherwise')
args = parser.parse_args()
model_dir, model_chkpnts_dir, model_samples_dir, model_val_samples_dir,\
model_summaries_dir = initialize_directories(args)
datasets_root_dir = join(args.data_dir, 'datasets')
loaded_data = load_training_data(datasets_root_dir, args.data_set,
args.caption_vector_length,
args.n_classes)
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.batch_size,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'caption_vector_length': args.caption_vector_length,
'n_classes': loaded_data['n_classes']
}
# Initialize and build the GAN model
gan = model.GAN(model_options)
input_tensors, variables, loss, outputs, checks = gan.build_model()
d_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(loss['d_loss'],
var_list=variables['d_vars'])
g_optim = tf.train.AdamOptimizer(args.learning_rate,
beta1=args.beta1).minimize(loss['g_loss'],
var_list=variables['g_vars'])
global_step_tensor = tf.Variable(1, trainable=False, name='global_step')
merged = tf.summary.merge_all()
sess = tf.InteractiveSession()
summary_writer = tf.summary.FileWriter(model_summaries_dir, sess.graph)
tf.global_variables_initializer().run()
saver = tf.train.Saver(max_to_keep=10000)
if args.resume_model:
print('Trying to resume training from a previous checkpoint' +
str(tf.train.latest_checkpoint(model_chkpnts_dir)))
if tf.train.latest_checkpoint(model_chkpnts_dir) is not None:
saver.restore(sess, tf.train.latest_checkpoint(model_chkpnts_dir))
print('Successfully loaded model. Resuming training.')
else:
print('Could not load checkpoints. Training a new model')
global_step = global_step_tensor.eval()
gs_assign_op = global_step_tensor.assign(global_step)
for i in range(args.epochs):
batch_no = 0
while batch_no * args.batch_size + args.batch_size < \
loaded_data['data_length']:
real_images, wrong_images, caption_vectors, z_noise, image_files, \
real_classes, wrong_classes, image_caps, image_ids = \
get_training_batch(batch_no, args.batch_size,
args.image_size, args.z_dim,
'train', datasets_root_dir,
args.data_set, loaded_data)
# DISCR UPDATE
check_ts = [checks['d_loss1'], checks['d_loss2'],
checks['d_loss3'], checks['d_loss1_1'],
checks['d_loss2_1']]
feed = {
input_tensors['t_real_image'].name : real_images,
input_tensors['t_wrong_image'].name : wrong_images,
input_tensors['t_real_caption'].name : caption_vectors,
input_tensors['t_z'].name : z_noise,
input_tensors['t_real_classes'].name : real_classes,
input_tensors['t_wrong_classes'].name : wrong_classes,
input_tensors['t_training'].name : args.train
}
_, d_loss, gen, d1, d2, d3, d4, d5= sess.run([d_optim,
loss['d_loss'],outputs['generator']] + check_ts,
feed_dict=feed)
print("D total loss: {}\n"
"D loss-1 [Real/Fake loss for real images] : {} \n"
"D loss-2 [Real/Fake loss for wrong images]: {} \n"
"D loss-3 [Real/Fake loss for fake images]: {} \n"
"D loss-4 [Aux Classifier loss for real images]: {} \n"
"D loss-5 [Aux Classifier loss for wrong images]: {}"
" ".format(d_loss, d1, d2, d3, d4, d5))
# GEN UPDATE
_, g_loss, gen = sess.run([g_optim, loss['g_loss'],
outputs['generator']], feed_dict=feed)
# GEN UPDATE TWICE
_, summary, g_loss, gen, g1, g2 = sess.run([g_optim, merged,
loss['g_loss'], outputs['generator'], checks['g_loss_1'],
checks['g_loss_2']], feed_dict=feed)
summary_writer.add_summary(summary, global_step)
print("\n\nLOSSES\nDiscriminator Loss: {}\nGenerator Loss: {"
"}\nBatch Number: {}\nEpoch: {},\nTotal Batches per "
"epoch: {}\n".format( d_loss, g_loss, batch_no, i,
int(len(loaded_data['image_list']) / args.batch_size)))
print("\nG loss-1 [Real/Fake loss for fake images] : {} \n"
"G loss-2 [Aux Classifier loss for fake images]: {} \n"
" ".format(g1, g2))
global_step += 1
sess.run(gs_assign_op)
batch_no += 1
if (i % args.save_every) == 0 and i != 0:
# if (batch_no % args.save_every) == 0 and batch_no != 0:
print("Saving Images and the Model\n\n")
save_for_vis(model_samples_dir, real_images, gen, image_files,
image_caps, image_ids)
save_path = saver.save(sess,
join(model_chkpnts_dir,
"latest_model_{}_temp.ckpt".format(
args.data_set)))
# Getting a batch for validation
val_captions, val_image_files, val_image_caps, val_image_ids = \
get_val_caps_batch(args.batch_size, loaded_data,
args.data_set, datasets_root_dir)
shutil.rmtree(model_val_samples_dir)
os.makedirs(model_val_samples_dir)
for val_viz_cnt in range(0, 4):
val_z_noise = np.random.uniform(-1, 1, [args.batch_size,
args.z_dim])
val_feed = {
input_tensors['t_real_caption'].name : val_captions,
input_tensors['t_z'].name : val_z_noise,
input_tensors['t_training'].name : True
}
val_gen = sess.run([outputs['generator']],
feed_dict=val_feed)
save_for_viz_val(model_val_samples_dir, val_gen,
val_image_files, val_image_caps,
val_image_ids, args.image_size,
val_viz_cnt)
# Save the model after save_every epoch
if i % args.save_every == 0 and i != 0:
epoch_dir = os.path.join(model_chkpnts_dir, str(i))
print("saving epoch %s" % epoch_dir)
if not os.path.exists(epoch_dir):
os.makedirs(epoch_dir)
checkpoint_path = os.path.join(epoch_dir, "model_after_{}_epoch_{}.ckpt".format(args.data_set, i))
print("save model to %s" % checkpoint_path)
save_path = saver.save(sess, checkpoint_path)
print("saved to %s" % save_path)
val_captions, val_image_files, val_image_caps, val_image_ids = \
get_val_caps_batch(args.batch_size, loaded_data,
args.data_set, datasets_root_dir)
shutil.rmtree(model_val_samples_dir)
os.makedirs(model_val_samples_dir)
for val_viz_cnt in range(0, 10):
val_z_noise = np.random.uniform(-1, 1, [args.batch_size,
args.z_dim])
val_feed = {
input_tensors['t_real_caption'].name : val_captions,
input_tensors['t_z'].name : val_z_noise,
input_tensors['t_training'].name : True
}
val_gen = sess.run([outputs['generator']], feed_dict=val_feed)
save_for_viz_val(model_val_samples_dir, val_gen,
val_image_files, val_image_caps,
val_image_ids, args.image_size,
val_viz_cnt)
# create required directory
required_dirs = ["param", "result", "mnist"]
hoge.make_dir(required_dirs)
mnist_data = MNIST('./mnist/',
train=True,
download=True,
transform=transforms.ToTensor())
dataloader = DataLoader(mnist_data, batch_size=mini_batch_num, shuffle=True)
print("\n")
# train restart
if interrupt_flag:
f = open("./param/tmp.pickle", mode="rb")
init_epoch = pickle.load(f)
model = model.GAN(dataloader, interrupting=True)
else:
init_epoch = 1
model = model.GAN(dataloader)
del dataloader
for epoch in range(init_epoch, epochs + 1):
print("Epoch[%d/%d]:" % (epoch, epochs))
model.study(epoch)
model.evaluate()
model.save_tmp_weight(epoch)
model.eval_pic(epoch)
model.output(epoch)
#model.output()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--z_dim',
type=int,
default=100,
help='Noise Dimension')
parser.add_argument('--t_dim',
type=int,
default=256,
help='Text feature dimension')
parser.add_argument('--image_size',
'-is',
type=int,
default=64,
help='Image Size')
parser.add_argument('--gf_dim',
type=int,
default=64,
help='Number of conv in the first layer gen.')
parser.add_argument('--df_dim',
type=int,
default=64,
help='Number of conv in the first layer discr.')
parser.add_argument(
'--gfc_dim',
type=int,
default=1024,
help='Dimension of gen untis for for fully connected layer 1024')
parser.add_argument('--caption_vector_length',
'-cvl',
type=int,
default=2400,
help='Caption Vector Length')
parser.add_argument('--data_set',
'-ds',
type=str,
default='faces',
help='data directory')
parser.add_argument('--method_dir',
'-md',
type=str,
default='',
help='method directory')
parser.add_argument('--model_path',
'-mp',
type=str,
default='latest_model_faces_temp.ckpt',
help='Trained Model Path')
parser.add_argument('--n_images',
'-ni',
type=int,
default=5,
help='Number of Images per Caption')
parser.add_argument('--caption_vectors',
'-cv',
type=str,
default='test_caption_vectors.hdf5',
help='Caption Thought Vector File')
parser.add_argument('--out_dir',
'-od',
type=str,
default='samples',
help='output directory')
args = parser.parse_args()
model_options = {
'z_dim': args.z_dim,
't_dim': args.t_dim,
'batch_size': args.n_images,
'image_size': args.image_size,
'gf_dim': args.gf_dim,
'df_dim': args.df_dim,
'gfc_dim': args.gfc_dim,
'caption_vector_length': args.caption_vector_length
}
gan = model.GAN(model_options)
_, _, _, _, _ = gan.build_model()
sess = tf.Session()
saver = tf.train.Saver()
saver.restore(
sess, join(args.data_set, args.method_dir, 'Models', args.model_path))
input_tensors, outputs = gan.build_generator()
h = h5py.File(join(args.data_set, args.method_dir, args.caption_vectors))
caption_image_dic = {}
for i, key in enumerate(h):
caption_images = []
z_noise = np.random.uniform(-1, 1, [args.n_images, args.z_dim])
caption = np.array([
h[key][0, :args.caption_vector_length]
for i in range(args.n_images)
])
[gen_image] =\
sess.run([outputs['generator']],
feed_dict = {input_tensors['t_real_caption'] : caption,
input_tensors['t_z'] : z_noise} )
caption_image_dic[key] =\
[gen_image[i, :, :, :] for i in range(0, args.n_images)]
if not os.path.exists(args.out_dir):
os.makedirs(args.out_dir)
for key in h:
for i, im in enumerate(caption_image_dic[key]):
scipy.misc.imsave(
join(args.out_dir, 'sample_' + key + '_' + str(i) + '.jpg'),
im)