def train():
identifier = id()
# directory 作成
os.mkdir(identifier)
# generator 単体で学習
G = generator()
G.compile(optimizer=Adam(lr=3e-04), loss='mae', metrics=['mae'])
encoder = Model(inputs=G.get_input_at(0),
outputs=G.get_layer(name='code').output)
# 学習する画像の準備
X = get_training_image(250000)
np.random.shuffle(X)
# training-set, test-set を分割
X_size = int(len(X))
train_size = int(X_size * 0.8)
test_size = X_size - train_size
X_train = X[0:train_size, :, :, :]
X_test = X[train_size:train_size + test_size, :, :, :]
epochs = 400000
batch_size = 16
num_batch = int(train_size / batch_size)
# TODO 調節
checkpoint = 10
for epoch in range(epochs):
start_time = time.time()
batch_x = 0
for _ in range(num_batch):
lb = batch_x
ub = min([batch_x + batch_size, train_size - 1])
G.train_on_batch(X_train[lb:ub, :, :, :], X_train[lb:ub, :, :, :])
batch_x += batch_size
# 訓練/テスト誤差の計算
train_score = G.evaluate(x=X_train, y=X_train, verbose=0)
test_score = G.evaluate(x=X_test, y=X_test, verbose=0)
now = datetime.datetime.now().strftime("%Y/%m/%d %H:%M:%S")
elapsed = time.time() - start_time
print("train epoch %06d: train loss %f: test loss %f: %.fs: %s" %
(epoch, train_score[0], test_score[0], elapsed, now))
if epoch % checkpoint == 0 and epoch > 0:
filename = "%s/input-predict-%06d.png" % (identifier, epoch)
Y_pred = G.predict(X_train[0:1, :, :, :])
scipy.misc.imsave(filename, np.concatenate([X_train[0],
Y_pred[0]]))
# モデルの保存
G.save("%s/G.h5" % (identifier))
def generator(self, x_init, reuse=False, scope="generator"):
if self.light:
with tf.variable_scope(scope, reuse=reuse):
G = generator_lite(x_init)
return G.fake
else:
with tf.variable_scope(scope, reuse=reuse):
G = generator(x_init)
return G.fake
def create_model(inputs, targets):
def create_discriminator(discrim_inputs, discrim_targets):
n_layers = 3
layers = []
# 2x [batch, height, width, in_channels] => [batch, height, width, in_channels * 2]
input = tf.concat([discrim_inputs, discrim_targets], axis=3)
# layer_1: [batch, 256, 256, in_channels * 2] => [batch, 128, 128, ndf]
with tf.variable_scope("layer_1"):
convolved = discrim_conv(input, a.ndf, stride=2)
rectified = lrelu(convolved, 0.2)
layers.append(rectified)
# layer_2: [batch, 128, 128, ndf] => [batch, 64, 64, ndf * 2]
# layer_3: [batch, 64, 64, ndf * 2] => [batch, 32, 32, ndf * 4]
# layer_4: [batch, 32, 32, ndf * 4] => [batch, 31, 31, ndf * 8]
for i in range(n_layers):
with tf.variable_scope("layer_%d" % (len(layers) + 1)):
out_channels = a.ndf * min(2**(i + 1), 8)
stride = 1 if i == n_layers - 1 else 2 # last layer here has stride 1
convolved = discrim_conv(layers[-1],
out_channels,
stride=stride)
normalized = batchnorm(convolved)
rectified = lrelu(normalized, 0.2)
layers.append(rectified)
# layer_5: [batch, 31, 31, ndf * 8] => [batch, 30, 30, 1]
with tf.variable_scope("layer_%d" % (len(layers) + 1)):
convolved = discrim_conv(rectified, out_channels=1, stride=1)
output = tf.sigmoid(convolved)
layers.append(output)
return layers[-1]
with tf.variable_scope("generator"):
out_channels = int(targets.get_shape()[-1])
outputs = generator(inputs)
# create two copies of discriminator, one for real pairs and one for fake pairs
# they share the same underlying variables
with tf.name_scope("real_discriminator"):
with tf.variable_scope("discriminator"):
# 2x [batch, height, width, channels] => [batch, 30, 30, 1]
predict_real = create_discriminator(inputs, targets)
with tf.name_scope("fake_discriminator"):
with tf.variable_scope("discriminator", reuse=True):
# 2x [batch, height, width, channels] => [batch, 30, 30, 1]
predict_fake = create_discriminator(inputs, outputs)
def predict():
cropsize = args.crop_size
img_string = tf.compat.v1.read_file(args.input_dir)
flash_input = tf.compat.v1.image.decode_image(img_string)
flash_input = tf.image.convert_image_dtype(flash_input, dtype=tf.float32)
begin = (int((args.img_h - cropsize) / 2), int(
(args.img_w - cropsize) / 2), 0)
flash_crop = tf.slice(flash_input, begin, [cropsize, cropsize, 3])
flash_crop = tf.reshape(tf.expand_dims(flash_crop**2.2, axis=0),
[1, cropsize, cropsize, 3])
latentcode = net.latentz_encoder(flash_crop, True)
predictions = deprocess(net.generator(latentcode, True))
wv, inten = net.generate_vl(cropsize * 2, cropsize * 2)
rerender = net.CTRender(predictions, wv, wv)
display_fetches = save_outputs(predictions, flash_crop, rerender)
return display_fetches
def main():
parser = argparse.ArgumentParser(description='waveone flavor codec')
parser.add_argument('--gmodel',
'-g',
help='Generator model (.h5)',
required=True)
parser.add_argument('--lrmodel',
'-l',
help='LinearRegression model (.pkl)',
required=True)
parser.add_argument('--mode',
'-m',
type=int,
default=0,
help='Codec mode [enc=0, dec=1]')
parser.add_argument('--input_path',
'-i',
help='Input image or Input code',
required=True)
parser.add_argument('--output_path',
'-o',
help='Output image or Output code',
required=True)
args = parser.parse_args()
mode = args.mode
gmodel = args.gmodel
lrmodel = args.lrmodel
input_path = args.input_path
output_path = args.output_path
G = generator()
G.load_weights(gmodel)
with open(lrmodel, "rb") as f:
LR = pickle.load(f)
if mode == 0:
encode(G, LR, input_path, output_path)
else:
decode(G, LR, input_path, output_path)
def test():
if not os.path.exists(args.test_output):
os.makedirs(args.test_output)
datalists = make_test_data_list(args.test_data_path)
test_image = tf.placeholder(
tf.float32,
shape=[1, args.image_size[0], args.image_size[1], 3],
name='test_image')
fake = generator(image=test_image, reuse=False, name='generator')
restore_var = [v for v in tf.global_variables() if 'generator' in v.name]
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
saver = tf.train.Saver(var_list=restore_var, max_to_keep=1)
checkpoint = tf.train.latest_checkpoint(args.snapshot_dir)
saver.restore(sess, checkpoint)
total_step = len(datalists)
for step in range(total_step):
test_image_name, test_img = TestImageReader(datalists, step,
args.image_size[1],
args.image_size[0])
batch_image = np.expand_dims(np.array(test_img).astype(np.float32),
axis=0)
feed_dict = {test_image: batch_image}
fake_value = sess.run(fake, feed_dict=feed_dict)
write_image = get_picture(test_img, fake_value)
write_image_name = args.test_output + "/" + test_image_name + ".jpg"
cv2.imwrite(write_image_name, write_image)
print('step {:d}'.format(step))
def main(FLAGS):
if not os.path.isdir(FLAGS.data_dir):
os.makedirs(FLAGS.data_dir)
download_data(FLAGS.data_dir, FLAGS.train_data, FLAGS.test_data)
net = GAN(real_size,
z_size,
learning_rate,
alpha=FLAGS.alpha,
beta1=FLAGS.beta1)
'''--------Load data--------'''
train_data = loadmat(FLAGS.data_dir + '/' + FLAGS.train_data)
test_data = loadmat(FLAGS.data_dir + '/' + FLAGS.test_data)
dataset = preprocessor.Dataset(train_data, test_data)
'''--------Build net--------'''
saver = tf.train.Saver()
sample_z = np.random.uniform(-1, 1, size=(72, z_size))
samples, losses = [], []
steps = 0
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for e in range(FLAGS.epochs):
for x, y in dataset.batches(FLAGS.batch_size):
steps += 1
# Sample random noise for G
batch_z = np.random.uniform(-1,
1,
size=(FLAGS.batch_size, z_size))
# Run optimizers
_ = sess.run(net.d_opt,
feed_dict={
net.input_real: x,
net.input_z: batch_z
})
_ = sess.run(net.g_opt,
feed_dict={
net.input_z: batch_z,
net.input_real: x
})
if steps % FLAGS.print_every == 0:
# At the end of each epoch, get the losses and print them out
train_loss_d = net.d_loss.eval({
net.input_z: batch_z,
net.input_real: x
})
train_loss_g = net.g_loss.eval({net.input_z: batch_z})
print('Epoch {}/{}...'.format(e + 1, FLAGS.epochs),
'Discriminator Loss: {:.4f}'.format(train_loss_d),
'Generator loss: {:.4f}'.format(train_loss_g))
# Save losses to view after traning
losses.append((train_loss_d, train_loss_g))
if steps % FLAGS.show_every == 0:
gen_samples = sess.run(generator(net.input_z,
3,
reuse=True,
training=False),
feed_dict={net.input_z: sample_z})
samples.append(gen_samples)
_ = utils.view_samples(-1,
samples,
6,
12,
figsize=(FLAGS.h_figsize,
FLAGS.v_figsize))
plt.show()
saver.save(sess, './checkpoints/generator.ckpt')
with open('samples.pkl', 'wb') as f:
pkl.dump(samples, f)
fig, ax = plt.subplot()
losses = np.array(losses)
plt.plot(losses.T[0], label='Discriminator', alpha=0.5)
plt.plot(losses.T[1], label='Generator', alpha=0.5)
plt.title('Training Losses')
plt.legend()
plt.show()
def train_gan():
identifier = id()
dir_path = "%s-gan-w" % identifier
os.mkdir(dir_path)
G = generator()
D = discriminator()
# TODO optimizer 調節
D.compile(optimizer=Adam(lr=3e-04),
loss='binary_crossentropy',
metrics=['accuracy'])
# build GAN
G_input = Input(shape=(128, 128, 3))
G_output = G(G_input)
raw_image_input = Input(shape=(128, 128, 3))
D_output = D([raw_image_input, G_output])
GAN = Model(inputs=[G_input, raw_image_input],
outputs=[G_output, D_output])
gan_loss = ['mae', 'binary_crossentropy']
gan_loss_weight = [1e2, 1]
D.trainable = False
# TODO optimizer 調節
GAN.compile(optimizer=Adam(lr=1e-06),
loss=gan_loss,
loss_weights=gan_loss_weight)
X = get_training_image(100000)
X_size = int(len(X) * 0.5)
train_size = int(X_size * 0.8)
test_size = X_size - train_size
X_train = X[0:train_size, :, :, :]
X_test = X[train_size:train_size + test_size, :, :, :]
epochs = 400000
batch_size = 16
num_batch = int(train_size / batch_size)
checkpoint = 1
for epoch in range(epochs):
start_time = time.time()
batch_x = 0
# D の学習を行うか判定
# acc が一定以上ならば学習しない
X_test_from_G = G.predict(X_test)
y_test_D = np.tile([1, 0], [test_size, 1])
loss_D = D.evaluate([X_test, X_test_from_G], y_test_D, verbose=0)
acc_D = loss_D[1]
trainable_D = True
if acc_D > 0.95:
trainable_D = False
for _ in range(num_batch):
lb = batch_x
ub = min([batch_x + batch_size, train_size - 1])
sz = ub - lb
# D の学習データを用意する
# G の生成データと生画像データのペアを用意しシャッフルして学習
X_from_G = G.predict(X_train[lb:ub, :, :, :])
X_as_Gen = np.zeros(shape=(sz, 128, 128, 3))
X_as_raw = np.zeros(shape=(sz, 128, 128, 3))
swap_labels = np.random.randint(2, size=sz)
y_train_D = np.zeros(shape=(sz, 2))
for i, bit in enumerate(swap_labels):
if bit == 0:
X_as_Gen[i] = X_from_G[i]
X_as_raw[i] = X_train[sz + i]
y_train_D[i] = np.asarray([1, 0])
else:
X_as_Gen[i] = X_train[sz + i]
X_as_raw[i] = X_from_G[i]
y_train_D[i] = np.asarray([0, 1])
# GAN の学習
y_train_GAN = np.tile([1, 0], [sz, 1])
GAN.train_on_batch(
[X_train[lb:ub, :, :, :], X_train[lb:ub, :, :, :]],
[X_from_G, y_train_GAN])
# D の学習
if trainable_D:
D.train_on_batch([X_as_raw, X_as_Gen], y_train_D)
# 状態更新
batch_x += batch_size
# logging
elapsed = time.time() - start_time
X_test_from_G = G.predict(X_test)
y_test_D = np.tile([1, 0], [test_size, 1])
loss_D = D.evaluate([X_test, X_test_from_G], y_test_D, verbose=0)
loss_GAN = GAN.evaluate([X_test, X_test], [X_test_from_G, y_test_D],
verbose=0)
now = datetime.datetime.now().strftime("%Y/%m/%d %H:%M:%S")
print(
"epoch %06d: D loss %.3f, acc %.3f: GAN loss %.3f, G loss %.3f D loss %.3f: trainable_D %d: %.3fs: %s"
% (epoch, loss_D[0], loss_D[1], loss_GAN[0], loss_GAN[1],
loss_GAN[2], trainable_D, elapsed, now))
# checkpoint
if epoch > 0 and (epoch % checkpoint == 0):
# 画像の保存
pred_image_path = "%s/G-predict-%06d.png" % (dir_path, epoch)
pred = G.predict(X_test)
tests = np.hstack(X_test[0:10])
preds = np.hstack(pred[0:10])
if epoch == checkpoint:
imsave("%s/tests-%06d.png" % (dir_path, epoch), tests)
imsave("%s/preds-%06d.png" % (dir_path, epoch), preds)
# モデルの保存
G.save("%s/G.h5" % dir_path)
D.save("%s/D.h5" % dir_path)
GAN.save("%s/GAN.h5" % dir_path)
def main():
kernel1 = gaussian_kernel(101, 101)
input_2b_sliced = concat_inputs(args.input_dir, kernel1)
examples = load_examples(input_2b_sliced, inputsize)
examples_flashes = examples.concats[:, :, :, 0:3]
examples_inputs = tf.map_fn(lambda x: tf.image.central_crop(x, 0.5),
elems=examples_flashes)
examples_inputs = tf.reshape(examples_inputs,
[BATCH_SIZE, inputsize, inputsize, 3])
wv = examples.concats[:, :, :, 3:6]
# inten = examples_concats[:,:,:,6:9]
initd = examples.concats[:, :, :, 9:12]
latentcode = net.latentz_encoder(examples_inputs)
predictions = deprocess(net.generator(latentcode))
net_rerender = net.CTRender(predictions, wv, wv)
prediffuse = predictions[:, :, :, 3:6]
dis_real = net.Discriminator_patch(examples_flashes, reuse=False)
dis_fake = net.Discriminator_patch(net_rerender, reuse=True)
dis_cost = net.patchGAN_d_loss(dis_fake, dis_real)
gen_fake = net.patchGAN_g_loss(dis_fake)
train_vars = tf.compat.v1.trainable_variables()
# encoder_vars = [v for v in train_vars if 'en_' in v.name]
decodernr_vars = [v for v in train_vars if 'denr_' in v.name]
decoderds_vars = [v for v in train_vars if 'deds_' in v.name]
discriminator_vars = [v for v in train_vars if 'd_' in v.name]
gnr_vars = decodernr_vars
gds_vars = decoderds_vars
diffuseloss = tf.reduce_mean(tf.abs(prediffuse - initd))
gnr_cost = ganscale * (gen_fake) + diffuseloss
gds_cost = ganscale * (gen_fake) + diffuseloss
gnr_optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=lr,
name='gnr_opt',
beta1=0.,
beta2=0.9).minimize(
gnr_cost,
var_list=gnr_vars)
gds_optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=lr,
name='gds_opt',
beta1=0.,
beta2=0.9).minimize(
gds_cost,
var_list=gds_vars)
d_optimizer = tf.compat.v1.train.AdamOptimizer(
learning_rate=lr, name='d_opt', beta1=0.,
beta2=0.9).minimize(dis_cost, var_list=discriminator_vars)
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.compat.v1.Session(config=config)
g_loss_summary = tf.summary.scalar("g_loss", gnr_cost)
d_loss_summary = tf.summary.scalar("d_loss", dis_cost)
train_summary = tf.compat.v1.summary.merge([
tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.SUMMARIES,
'g_loss'),
tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.SUMMARIES, 'd_loss')
])
train_writer = tf.compat.v1.summary.FileWriter(args.log_dir, sess.graph)
saver = tf.compat.v1.train.Saver(max_to_keep=10)
sess.run([
examples.iterator.initializer,
tf.compat.v1.global_variables_initializer()
])
for step in range(args.max_step):
for i in range(5):
_, g_loss = sess.run([gnr_optimizer, gnr_cost])
for i in range(1):
_, g_loss = sess.run([gds_optimizer, gds_cost])
_, d_loss = sess.run([d_optimizer, dis_cost])
if step % 500 == 0 or (step + 1) == args.max_step:
print('Step %d, g_loss = %.4f, d_loss = %.4f' %
(step, g_loss, d_loss))
summary_str = sess.run(train_summary)
train_writer.add_summary(summary_str, step)
if step % 1000 == 0 or (step + 1) == args.max_step:
checkpoint_path = os.path.join(args.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
display_fetches = save_outputs(predictions, examples_flashes,
net_rerender)
results = sess.run(display_fetches)
save_images(results, args.output_dir, step, "training_tiles")
prediction_fetches = predict()
predict_maps = sess.run(prediction_fetches)
save_images(predict_maps, args.output_dir, step, "predicted_maps")
clip_value = [-0.01,0.01]
tf.reset_default_graph()
image_record = data.readRecord('../data/train_clean.tfrecords')
train_from_checkpoint = True
checkpoint_dir = "../model/20190110-090118_ft_from_20190109-090711/"
stddev_scheme = [0] if train_from_checkpoint else [1e-3]#[ii*1e-5 for ii in range(50,0,-1)]+[0] #[0.01,0.009,...,0.001]
scheme_step = 2000
## define input
real_image = tf.placeholder(tf.float32, (batch_size,64,64,3))
inptG = tf.placeholder(tf.float32, (batch_size, noise_size))
gn_stddev = tf.placeholder(tf.float32, [])
training = tf.placeholder(tf.bool, [])
fake_image = net.generator(inptG, training)
real_image_batch = data.get_batch_image([image_record], batch_size)
noise_batch = data.get_batch_noise(noise_size, batch_size)
noise_batch_show = data.get_batch_noise(noise_size, top_k)
real_scores = net.discriminator(real_image, gn_stddev, training)
fake_scores = net.discriminator(fake_image, gn_stddev, training)
# topk_scores, topk_index = tf.nn.top_k(tf.reshape(fake_scores,[-1,]),top_k)
m_real_score = tf.reduce_mean(real_scores)
m_fake_score = tf.reduce_mean(fake_scores)
# define losses
d_loss = net.loss_fn_d(real_scores, fake_scores)
g_loss = net.loss_fn_g(fake_scores)
