def dirtt():
T = tb.utils.TensorDict(dict(
sess = tf.Session(config=tb.growth_config()),
src_x = placeholder((None, 32, 32, 3)),
src_y = placeholder((None, args.Y)),
trg_x = placeholder((None, 32, 32, 3)),
trg_y = placeholder((None, args.Y)),
test_x = placeholder((None, 32, 32, 3)),
test_y = placeholder((None, args.Y)),
))
# Supervised and conditional entropy minimization
src_e = nn.classifier(T.src_x, phase=True, enc_phase=1, trim=args.trim)
trg_e = nn.classifier(T.trg_x, phase=True, enc_phase=1, trim=args.trim, reuse=True, internal_update=True)
src_p = nn.classifier(src_e, phase=True, enc_phase=0, trim=args.trim)
trg_p = nn.classifier(trg_e, phase=True, enc_phase=0, trim=args.trim, reuse=True, internal_update=True)
loss_src_class = tf.reduce_mean(softmax_xent(labels=T.src_y, logits=src_p))
loss_trg_cent = tf.reduce_mean(softmax_xent_two(labels=trg_p, logits=trg_p))
# Domain confusion
if args.dw > 0 and args.dirt == 0:
real_logit = nn.feature_discriminator(src_e, phase=True)
fake_logit = nn.feature_discriminator(trg_e, phase=True, reuse=True)
loss_disc = 0.5 * tf.reduce_mean(
sigmoid_xent(labels=tf.ones_like(real_logit), logits=real_logit) +
sigmoid_xent(labels=tf.zeros_like(fake_logit), logits=fake_logit))
loss_domain = 0.5 * tf.reduce_mean(
sigmoid_xent(labels=tf.zeros_like(real_logit), logits=real_logit) +
sigmoid_xent(labels=tf.ones_like(fake_logit), logits=fake_logit))
else:
loss_disc = constant(0)
loss_domain = constant(0)
# Virtual adversarial training (turn off src in non-VADA phase)
loss_src_vat = vat_loss(T.src_x, src_p, nn.classifier) if args.sw > 0 and args.dirt == 0 else constant(0)
loss_trg_vat = vat_loss(T.trg_x, trg_p, nn.classifier) if args.tw > 0 else constant(0)
# Evaluation (EMA)
ema = tf.train.ExponentialMovingAverage(decay=0.998)
var_class = tf.get_collection('trainable_variables', 'class/')
ema_op = ema.apply(var_class)
ema_p = nn.classifier(T.test_x, phase=False, reuse=True, getter=tb.tfutils.get_getter(ema))
# Teacher model (a back-up of EMA model)
teacher_p = nn.classifier(T.test_x, phase=False, scope='teacher')
var_main = tf.get_collection('variables', 'class/(?!.*ExponentialMovingAverage:0)')
var_teacher = tf.get_collection('variables', 'teacher/(?!.*ExponentialMovingAverage:0)')
teacher_assign_ops = []
for t, m in zip(var_teacher, var_main):
ave = ema.average(m)
ave = ave if ave else m
teacher_assign_ops += [tf.assign(t, ave)]
update_teacher = tf.group(*teacher_assign_ops)
teacher = tb.function(T.sess, [T.test_x], tf.nn.softmax(teacher_p))
# Accuracies
src_acc = basic_accuracy(T.src_y, src_p)
trg_acc = basic_accuracy(T.trg_y, trg_p)
ema_acc = basic_accuracy(T.test_y, ema_p)
fn_ema_acc = tb.function(T.sess, [T.test_x, T.test_y], ema_acc)
# Optimizer
dw = constant(args.dw) if args.dirt == 0 else constant(0)
cw = constant(1) if args.dirt == 0 else constant(args.bw)
sw = constant(args.sw) if args.dirt == 0 else constant(0)
tw = constant(args.tw)
loss_main = (dw * loss_domain +
cw * loss_src_class +
sw * loss_src_vat +
tw * loss_trg_cent +
tw * loss_trg_vat)
var_main = tf.get_collection('trainable_variables', 'class')
train_main = tf.train.AdamOptimizer(args.lr, 0.5).minimize(loss_main, var_list=var_main)
train_main = tf.group(train_main, ema_op)
if args.dw > 0 and args.dirt == 0:
var_disc = tf.get_collection('trainable_variables', 'disc')
train_disc = tf.train.AdamOptimizer(args.lr, 0.5).minimize(loss_disc, var_list=var_disc)
else:
train_disc = constant(0)
# Summarizations
summary_disc = [tf.summary.scalar('domain/loss_disc', loss_disc),]
summary_main = [tf.summary.scalar('domain/loss_domain', loss_domain),
tf.summary.scalar('class/loss_src_class', loss_src_class),
tf.summary.scalar('class/loss_trg_cent', loss_trg_cent),
tf.summary.scalar('lipschitz/loss_trg_vat', loss_trg_vat),
tf.summary.scalar('lipschitz/loss_src_vat', loss_src_vat),
tf.summary.scalar('hyper/dw', dw),
tf.summary.scalar('hyper/cw', cw),
tf.summary.scalar('hyper/sw', sw),
tf.summary.scalar('hyper/tw', tw),
tf.summary.scalar('acc/src_acc', src_acc),
tf.summary.scalar('acc/trg_acc', trg_acc)]
# Merge summaries
summary_disc = tf.summary.merge(summary_disc)
summary_main = tf.summary.merge(summary_main)
# Saved ops
c = tf.constant
T.ops_print = [c('disc'), loss_disc,
c('domain'), loss_domain,
c('class'), loss_src_class,
c('cent'), loss_trg_cent,
c('trg_vat'), loss_trg_vat,
c('src_vat'), loss_src_vat,
c('src'), src_acc,
c('trg'), trg_acc]
T.ops_disc = [summary_disc, train_disc]
T.ops_main = [summary_main, train_main]
T.fn_ema_acc = fn_ema_acc
T.teacher = teacher
T.update_teacher = update_teacher
return T
def model(FLAGS, gpu_config):
"""
:param FLAGS: Contains the experiment info
:return: (TensorDict) the model
"""
print(colored("Model initialization started", "blue"))
nn = network(FLAGS)
sz = FLAGS.sz
ch = FLAGS.ch
bs = FLAGS.bs
sbs = FLAGS.sbs
alpha = constant(FLAGS.alpha)
beta = constant(FLAGS.beta)
theta = constant(FLAGS.theta)
delta = constant(FLAGS.delta)
T = tb.utils.TensorDict(
dict(sess=tf.Session(config=tb.growth_config()),
x=placeholder((bs, sz, sz, ch)),
z=placeholder((bs, FLAGS.nz)),
pos=placeholder((bs * FLAGS.jcb, FLAGS.nz)),
iorth=placeholder((bs, FLAGS.jcb, FLAGS.jcb)),
lrD=placeholder(None),
lrG=placeholder(None),
seq_in=placeholder((10, sbs, sz, sz, ch)),
seq_out=placeholder((10, sbs, sz, sz, ch)),
val_seq_in=placeholder((10, 10, sz, sz, ch)),
val_seq_out=placeholder((10, 10, sz, sz, ch)),
test_seq_in=placeholder((10, 10, sz, sz, ch)),
lr=placeholder(None)))
# Compute G(x, z) and G(x, 0)
fake_x = nn.generator(T.x, T.z, phase=True)
# T.fake_x0 = fake_x0 = nn.generator(T.x, tf.zeros_like(T.z), phase=True)
fake_x0 = nn.generator(T.x, tf.zeros_like(T.z), phase=True)
# Compute discriminator logits
real_logit = nn.discriminator(T.x, phase=True)
fake_logit = nn.discriminator(fake_x, phase=True)
fake0_logit = nn.discriminator(fake_x0, phase=True)
# Adversarial generator
loss_disc = tf.reduce_mean(
sigmoid_xent(labels=tf.ones_like(real_logit), logits=real_logit) +
sigmoid_xent(labels=tf.zeros_like(fake_logit), logits=fake_logit) +
theta *
sigmoid_xent(labels=tf.zeros_like(fake0_logit), logits=fake0_logit))
loss_fake = tf.reduce_mean(
sigmoid_xent(labels=tf.ones_like(fake_logit), logits=fake_logit) +
theta *
sigmoid_xent(labels=tf.ones_like(fake0_logit), logits=fake0_logit))
# Locality
loss_local = tf.reduce_mean(abs_diff(labels=T.x, predictions=fake_x0))
# Orthogonality
pos = T.pos * delta
tiled_real_x = tf.tile(T.x, [FLAGS.jcb, 1, 1, 1])
pos_fake_x = nn.generator(tiled_real_x, pos, phase=True)
neg_fake_x = nn.generator(tiled_real_x, -pos, phase=True)
jx = (pos_fake_x - neg_fake_x) / (2 * delta)
jx = tf.reshape(jx, [bs, FLAGS.jcb, -1])
jx_t = tf.transpose(jx, [0, 2, 1])
loss_orth = tf.reduce_mean(abs_diff(tf.matmul(jx, jx_t), T.iorth))
loss_gen = loss_fake + alpha * loss_local + beta * loss_orth
# Optimizer
var_disc = tf.get_collection('trainable_variables', 'lgan/dsc')
train_disc = tf.train.AdamOptimizer(T.lrD, 0.5).minimize(loss_disc,
var_list=var_disc)
if FLAGS.clip:
clip_disc = [
p.assign(tf.clip_by_value(p, -0.01, 0.01)) for p in var_disc
]
var_gen = tf.get_collection('trainable_variables', 'lgan/gen')
train_gen = tf.train.AdamOptimizer(T.lrG, 0.5).minimize(loss_gen,
var_list=var_gen)
# Summarizations
summary_disc = [
tf.summary.scalar('disc/loss_disc', loss_disc),
]
summary_gen = [
tf.summary.scalar('gen/loss_gen', loss_gen),
tf.summary.scalar('gen/loss_fake', loss_fake),
tf.summary.scalar('gen/loss_local', loss_local),
tf.summary.scalar('gen/loss_orth', loss_orth),
tf.summary.scalar('hyper/alpha', alpha),
tf.summary.scalar('hyper/beta', beta),
tf.summary.scalar('hyper/theta', theta),
tf.summary.scalar('hyper/delta', delta),
tf.summary.scalar('hyper/lrD', T.lrD),
tf.summary.scalar('hyper/lrG', T.lrG),
tf.summary.scalar('hyper/var', FLAGS.var)
]
summary_image = [
tf.summary.image('image/x', T.x),
tf.summary.image('image/fake_x', fake_x),
tf.summary.image('image/fake_x0', fake_x0)
]
# Merge summaries
summary_disc = tf.summary.merge(summary_disc)
summary_gen = tf.summary.merge(summary_gen)
summary_image = tf.summary.merge(summary_image)
# Saved ops
c = tf.constant
T.ops_print = [
c('disc'), loss_disc,
c('gen'), loss_gen,
c('fake'), loss_fake,
c('local'), loss_local,
c('orth'), loss_orth
]
# T.ops_disc = [summary_disc, train_disc]
if FLAGS.clip:
T.ops_disc = [summary_disc, train_disc, clip_disc]
else:
T.ops_disc = [summary_disc, train_disc]
T.ops_gen = [summary_gen, train_gen]
T.ops_image = summary_image
if FLAGS.phase:
# LSTM initialization
seq_in = tf.reshape(T.seq_in, [-1, sz, sz, ch])
seq_out = tf.reshape(T.seq_out, [-1, sz, sz, ch])
val_seq_in = tf.reshape(T.val_seq_in, [-1, sz, sz, ch])
test_seq_in = tf.reshape(T.test_seq_in, [-1, sz, sz, ch])
enc_in = nn.generator(seq_in,
tf.zeros((10 * sbs, FLAGS.nz)),
phase=True,
enc=True)
enc_out = nn.generator(seq_out,
tf.zeros((10 * sbs, FLAGS.nz)),
phase=True,
enc=True)
val_enc_in = nn.generator(val_seq_in,
tf.zeros((10 * 10, FLAGS.nz)),
phase=True,
enc=True)
test_enc_in = nn.generator(test_seq_in,
tf.zeros((10 * 10, FLAGS.nz)),
phase=True,
enc=True)
enc_in = tf.stop_gradient(enc_in)
enc_out = tf.stop_gradient(enc_out)
val_enc_in = tf.stop_gradient(val_enc_in)
test_enc_in = tf.stop_gradient(test_enc_in)
enc_in = tf.squeeze(enc_in)
enc_out = tf.squeeze(enc_out)
val_enc_in = tf.squeeze(val_enc_in)
test_enc_in = tf.squeeze(test_enc_in)
enc_in = tf.reshape(enc_in, [-1, sbs, 3 * FLAGS.nz])
enc_out = tf.reshape(enc_out, [-1, sbs, 3 * FLAGS.nz])
val_enc_in = tf.reshape(val_enc_in, [-1, 10, 3 * FLAGS.nz])
test_enc_in = tf.reshape(test_enc_in, [-1, 10, 3 * FLAGS.nz])
with tf.variable_scope('lstm/in'):
in_cell = tf.contrib.cudnn_rnn.CudnnLSTM(FLAGS.nhl,
FLAGS.nhw,
dropout=0.5)
_, in_states = in_cell(enc_in, initial_state=None, training=True)
_, val_in_states = in_cell(val_enc_in,
initial_state=None,
training=False)
_, test_in_states = in_cell(test_enc_in,
initial_state=None,
training=False)
with tf.variable_scope('lstm/out'):
out_cell = tf.contrib.cudnn_rnn.CudnnLSTM(FLAGS.nhl,
FLAGS.nhw,
dropout=0.5)
outputs, _ = out_cell(tf.zeros_like(enc_out),
initial_state=in_states,
training=True)
val_outputs, _ = out_cell(tf.zeros_like(val_enc_in),
initial_state=val_in_states,
training=False)
test_outputs, _ = out_cell(tf.zeros_like(test_enc_in),
initial_state=test_in_states,
training=False)
enc_out_pred = tf.layers.dense(outputs,
3 * FLAGS.nz,
activation=None,
name='lstm_dense',
reuse=tf.AUTO_REUSE)
val_enc_out_pred = tf.layers.dense(val_outputs,
3 * FLAGS.nz,
activation=None,
name='lstm_dense',
reuse=tf.AUTO_REUSE)
test_enc_out_pred = tf.layers.dense(test_outputs,
3 * FLAGS.nz,
activation=None,
name='lstm_dense',
reuse=tf.AUTO_REUSE)
enc_out_pred_reshape = tf.reshape(enc_out_pred, [-1, 3 * FLAGS.nz])
enc_out_pred_reshape = tf.expand_dims(
tf.expand_dims(enc_out_pred_reshape, 1), 1)
val_enc_out_pred_reshape = tf.reshape(val_enc_out_pred,
[-1, 3 * FLAGS.nz])
val_enc_out_pred_reshape = tf.expand_dims(
tf.expand_dims(val_enc_out_pred_reshape, 1), 1)
test_enc_out_pred_reshape = tf.reshape(test_enc_out_pred,
[-1, 3 * FLAGS.nz])
test_enc_out_pred_reshape = tf.expand_dims(
tf.expand_dims(test_enc_out_pred_reshape, 1), 1)
seq_out_pred = nn.generator(enc_out_pred_reshape,
tf.zeros((10 * sbs, FLAGS.nz)),
phase=True,
dec=True)
seq_out_pred = tf.reshape(seq_out_pred, [10, sbs, sz, sz, ch])
val_seq_out_pred = nn.generator(val_enc_out_pred_reshape,
tf.zeros((10 * 10, FLAGS.nz)),
phase=True,
dec=True)
val_seq_out_pred = tf.reshape(val_seq_out_pred, [10, 10, sz, sz, ch])
test_seq_out_pred = nn.generator(test_enc_out_pred_reshape,
tf.zeros((10 * 10, FLAGS.nz)),
phase=True,
dec=True)
T.test_seq_out_pred = tf.reshape(test_seq_out_pred,
[10, 10, sz, sz, ch])
T.val_mae = tf.reduce_mean(
abs_diff(labels=T.val_seq_out, predictions=val_seq_out_pred))
loss_lstm = tf.reduce_mean(
abs_diff(labels=enc_out, predictions=enc_out_pred))
var_lstm = tf.get_collection('trainable_variables', 'lstm')
# train_lstm = tf.train.AdamOptimizer(FLAGS.lr, 0.5).minimize(loss_lstm, var_list=var_lstm)
train_lstm = tf.train.AdamOptimizer(T.lr,
0.5).minimize(loss_lstm,
var_list=var_lstm)
summary_lstm = [tf.summary.scalar('lstm/loss_lstm', loss_lstm)]
summary_lstm_image = [
tf.summary.image('lstm/seq_out', T.seq_out[:, 0, :, :, :]),
tf.summary.image('lstm/seq_out_pred', seq_out_pred[:, 0, :, :, :])
]
summary_lstm = tf.summary.merge(summary_lstm)
summary_lstm_image = tf.summary.merge(summary_lstm_image)
T.ops_lstm_print = [c('loss_lstm'), loss_lstm]
T.ops_lstm = [summary_lstm, train_lstm]
T.ops_lstm_image = summary_lstm_image
# T.test1 = seq_out_pred
print(colored("Model initialization ended", "blue"))
return T
def dirtt():
T = tb.utils.TensorDict(dict(
sess = tf.Session(config=tb.growth_config()),
src_x = placeholder((None, 500, 60, 1)),
src_y = placeholder((None, args.Y)),
trg_x = placeholder((None, 500, 60, 1)),
trg_y = placeholder((None, args.Y)),
test_x = placeholder((None, 500, 60, 1)),
test_y = placeholder((None, args.Y)),
))
# Supervised and conditional entropy minimization
src_e = nn.classifier(T.src_x, phase=True, enc_phase=1, trim=args.trim)
trg_e = nn.classifier(T.trg_x, phase=True, enc_phase=1, trim=args.trim, reuse=True, internal_update=True)
src_p = nn.classifier(src_e, phase=True, enc_phase=0, trim=args.trim)
trg_p = nn.classifier(trg_e, phase=True, enc_phase=0, trim=args.trim, reuse=True, internal_update=True)
loss_src_class = tf.reduce_mean(softmax_xent(labels=T.src_y, logits=src_p))
loss_trg_cent = tf.reduce_mean(softmax_xent_two(labels=trg_p, logits=trg_p))
# Domain confusion
if args.dw > 0 and args.dirt == 0:
real_logit = nn.feature_discriminator(src_e, phase=True)
fake_logit = nn.feature_discriminator(trg_e, phase=True, reuse=True)
loss_disc = 0.5 * tf.reduce_mean(
sigmoid_xent(labels=tf.ones_like(real_logit), logits=real_logit) +
sigmoid_xent(labels=tf.zeros_like(fake_logit), logits=fake_logit))
loss_domain = 0.5 * tf.reduce_mean(
sigmoid_xent(labels=tf.zeros_like(real_logit), logits=real_logit) +
sigmoid_xent(labels=tf.ones_like(fake_logit), logits=fake_logit))
else:
loss_disc = constant(0)
loss_domain = constant(0)
# Virtual adversarial training (turn off src in non-VADA phase)
loss_src_vat = vat_loss(T.src_x, src_p, nn.classifier) if args.sw > 0 and args.dirt == 0 else constant(0)
loss_trg_vat = vat_loss(T.trg_x, trg_p, nn.classifier) if args.tw > 0 else constant(0)
# Evaluation (EMA)
ema = tf.train.ExponentialMovingAverage(decay=0.998)
var_class = tf.get_collection('trainable_variables', 'class/')
ema_op = ema.apply(var_class)
ema_p = nn.classifier(T.test_x, phase=False, reuse=True, getter=tb.tfutils.get_getter(ema))
# Teacher model (a back-up of EMA model)
teacher_p = nn.classifier(T.test_x, phase=False, scope='teacher')
var_main = tf.get_collection('variables', 'class/(?!.*ExponentialMovingAverage:0)')
var_teacher = tf.get_collection('variables', 'teacher/(?!.*ExponentialMovingAverage:0)')
teacher_assign_ops = []
for t, m in zip(var_teacher, var_main):
ave = ema.average(m)
ave = ave if ave else m
teacher_assign_ops += [tf.assign(t, ave)]
update_teacher = tf.group(*teacher_assign_ops)
teacher = tb.function(T.sess, [T.test_x], tf.nn.softmax(teacher_p))
# Accuracies
src_acc = basic_accuracy(T.src_y, src_p)
trg_acc = basic_accuracy(T.trg_y, trg_p)
ema_acc = basic_accuracy(T.test_y, ema_p)
fn_ema_acc = tb.function(T.sess, [T.test_x, T.test_y], ema_acc)
# Optimizer
dw = constant(args.dw) if args.dirt == 0 else constant(0)
cw = constant(1) if args.dirt == 0 else constant(args.bw)
sw = constant(args.sw) if args.dirt == 0 else constant(0)
tw = constant(args.tw)
loss_main = (dw * loss_domain +
cw * loss_src_class +
sw * loss_src_vat +
tw * loss_trg_cent +
tw * loss_trg_vat)
var_main = tf.get_collection('trainable_variables', 'class')
train_main = tf.train.AdamOptimizer(args.lr, 0.5).minimize(loss_main, var_list=var_main)
train_main = tf.group(train_main, ema_op)
if args.dw > 0 and args.dirt == 0:
var_disc = tf.get_collection('trainable_variables', 'disc')
train_disc = tf.train.AdamOptimizer(args.lr, 0.5).minimize(loss_disc, var_list=var_disc)
else:
train_disc = constant(0)
# Summarizations
summary_disc = [tf.summary.scalar('domain/loss_disc', loss_disc),]
summary_main = [tf.summary.scalar('domain/loss_domain', loss_domain),
tf.summary.scalar('class/loss_src_class', loss_src_class),
tf.summary.scalar('class/loss_trg_cent', loss_trg_cent),
tf.summary.scalar('lipschitz/loss_trg_vat', loss_trg_vat),
tf.summary.scalar('lipschitz/loss_src_vat', loss_src_vat),
tf.summary.scalar('hyper/dw', dw),
tf.summary.scalar('hyper/cw', cw),
tf.summary.scalar('hyper/sw', sw),
tf.summary.scalar('hyper/tw', tw),
tf.summary.scalar('acc/src_acc', src_acc),
tf.summary.scalar('acc/trg_acc', trg_acc)]
# Merge summaries
summary_disc = tf.summary.merge(summary_disc)
summary_main = tf.summary.merge(summary_main)
# Saved ops
c = tf.constant
T.ops_print = [c('disc'), loss_disc,
c('domain'), loss_domain,
c('class'), loss_src_class,
c('cent'), loss_trg_cent,
c('trg_vat'), loss_trg_vat,
c('src_vat'), loss_src_vat,
c('src'), src_acc,
c('trg'), trg_acc]
T.ops_disc = [summary_disc, train_disc]
T.ops_main = [summary_main, train_main]
T.fn_ema_acc = fn_ema_acc
T.teacher = teacher
T.update_teacher = update_teacher
return T
def model(FLAGS, gpu_config):
"""
:param FLAGS: Contains the experiment info
:return: (TensorDict) the model
"""
print(colored("Model initialization started", "blue"))
nn = network(FLAGS)
sz = FLAGS.sz
ch = FLAGS.ch
bs = FLAGS.bs
sbs = FLAGS.sbs
T = tb.utils.TensorDict(dict(
sess=tf.Session(config=tb.growth_config()),
x=placeholder((bs, sz, sz, ch)),
lrD=placeholder(None),
lrG=placeholder(None),
seq_in=placeholder((10, sbs, sz, sz, ch)),
seq_out=placeholder((10, sbs, sz, sz, ch)),
val_seq_in=placeholder((10, 10, sz, sz, ch)),
val_seq_out=placeholder((10, 10, sz, sz, ch)),
test_seq_in=placeholder((10, 10, sz, sz, ch)),
lr=placeholder(None)
))
recon_x = nn.generator(T.x, phase=True)
# Compute discriminator logits
real_logit = nn.discriminator(T.x, phase=True)
fake_logit = nn.discriminator(recon_x, phase=True)
# Adversarial generator
loss_disc = tf.reduce_mean(
sigmoid_xent(labels=tf.ones_like(real_logit), logits=real_logit) +
sigmoid_xent(labels=tf.zeros_like(fake_logit), logits=fake_logit))
loss_fake = tf.reduce_mean(
sigmoid_xent(labels=tf.ones_like(fake_logit), logits=fake_logit))
loss_local = tf.reduce_mean(abs_diff(labels=T.x, predictions=recon_x))
loss_gen = loss_fake + FLAGS.alpha * loss_local
var_gen = tf.get_collection('trainable_variables', 'lgan/gen')
train_gen = tf.train.AdamOptimizer(T.lrG, 0.5).minimize(loss_gen, var_list=var_gen)
var_disc = tf.get_collection('trainable_variables', 'lgan/dsc')
train_disc = tf.train.AdamOptimizer(T.lrD, 0.5).minimize(loss_disc, var_list=var_disc)
# Summarizations
summary_disc = [tf.summary.scalar('disc/loss_disc', loss_disc)]
summary_gen = [tf.summary.scalar('gen/loss_gen', loss_gen),
tf.summary.scalar('gen/loss_local', loss_local),
tf.summary.scalar('gen/loss_fake', loss_fake),
tf.summary.scalar('hyper/lrD', T.lrD),
tf.summary.scalar('hyper/lrG', T.lrG)]
summary_image = [tf.summary.image('image/x', T.x),
tf.summary.image('image/recon_x', recon_x)]
# Merge summaries
summary_disc = tf.summary.merge(summary_disc)
summary_gen = tf.summary.merge(summary_gen)
summary_image = tf.summary.merge(summary_image)
# Saved ops
c = tf.constant
T.ops_print = [c('disc'), loss_disc,
c('gen'), loss_gen,
c('local'), loss_local,
c('fake'), loss_fake]
T.ops_disc = [summary_disc, train_disc]
T.ops_gen = [summary_gen, train_gen]
T.ops_image = summary_image
if FLAGS.phase:
# LSTM initialization
seq_in = tf.reshape(T.seq_in, [-1, sz, sz, ch])
seq_out = tf.reshape(T.seq_out, [-1, sz, sz, ch])
val_seq_in = tf.reshape(T.val_seq_in, [-1, sz, sz, ch])
test_seq_in = tf.reshape(T.test_seq_in, [-1, sz, sz, ch])
enc_in = nn.generator(seq_in, phase=True, enc=True)
enc_out = nn.generator(seq_out, phase=True, enc=True)
val_enc_in = nn.generator(val_seq_in, phase=True, enc=True)
test_enc_in = nn.generator(test_seq_in, phase=True, enc=True)
enc_in = tf.stop_gradient(enc_in)
enc_out = tf.stop_gradient(enc_out)
val_enc_in = tf.stop_gradient(val_enc_in)
test_enc_in = tf.stop_gradient(test_enc_in)
enc_in = tf.squeeze(enc_in)
enc_out = tf.squeeze(enc_out)
val_enc_in = tf.squeeze(val_enc_in)
test_enc_in = tf.squeeze(test_enc_in)
enc_in = tf.reshape(enc_in, [-1, sbs, FLAGS.nz])
enc_out = tf.reshape(enc_out, [-1, sbs, FLAGS.nz])
val_enc_in = tf.reshape(val_enc_in, [-1, 10, FLAGS.nz])
test_enc_in = tf.reshape(test_enc_in, [-1, 10, FLAGS.nz])
with tf.variable_scope('lstm/in'):
in_cell = tf.contrib.cudnn_rnn.CudnnLSTM(FLAGS.nhl, FLAGS.nhw, dropout=0.5)
_, in_states = in_cell(enc_in, initial_state=None, training=True)
_, val_in_states = in_cell(val_enc_in, initial_state=None, training=False)
_, test_in_states = in_cell(test_enc_in, initial_state=None, training=False)
with tf.variable_scope('lstm/out'):
out_cell = tf.contrib.cudnn_rnn.CudnnLSTM(FLAGS.nhl, FLAGS.nhw, dropout=0.5)
outputs, _ = out_cell(tf.zeros_like(enc_out), initial_state=in_states, training=True)
val_outputs, _ = out_cell(tf.zeros_like(val_enc_in), initial_state=val_in_states, training=False)
test_outputs, _ = out_cell(tf.zeros_like(test_enc_in), initial_state=test_in_states, training=False)
enc_out_pred = tf.layers.dense(outputs, FLAGS.nz, activation=None, name='lstm_dense', reuse=tf.AUTO_REUSE)
val_enc_out_pred = tf.layers.dense(val_outputs, FLAGS.nz, activation=None, name='lstm_dense',
reuse=tf.AUTO_REUSE)
test_enc_out_pred = tf.layers.dense(test_outputs, FLAGS.nz, activation=None, name='lstm_dense',
reuse=tf.AUTO_REUSE)
enc_out_pred_reshape = tf.reshape(enc_out_pred, [-1, FLAGS.nz])
enc_out_pred_reshape = tf.expand_dims(tf.expand_dims(enc_out_pred_reshape, 1), 1)
val_enc_out_pred_reshape = tf.reshape(val_enc_out_pred, [-1, FLAGS.nz])
val_enc_out_pred_reshape = tf.expand_dims(tf.expand_dims(val_enc_out_pred_reshape, 1), 1)
test_enc_out_pred_reshape = tf.reshape(test_enc_out_pred, [-1, FLAGS.nz])
test_enc_out_pred_reshape = tf.expand_dims(tf.expand_dims(test_enc_out_pred_reshape, 1), 1)
seq_out_pred = nn.generator(enc_out_pred_reshape, phase=True, dec=True)
seq_out_pred = tf.reshape(seq_out_pred, [10, sbs, sz, sz, ch])
val_seq_out_pred = nn.generator(val_enc_out_pred_reshape, phase=True, dec=True)
val_seq_out_pred = tf.reshape(val_seq_out_pred, [10, 10, sz, sz, ch])
test_seq_out_pred = nn.generator(test_enc_out_pred_reshape, phase=True, dec=True)
T.test_seq_out_pred = tf.reshape(test_seq_out_pred, [10, 10, sz, sz, ch])
T.val_mae = tf.reduce_mean(abs_diff(labels=T.val_seq_out, predictions=val_seq_out_pred))
loss_lstm = tf.reduce_mean(abs_diff(labels=enc_out, predictions=enc_out_pred))
var_lstm = tf.get_collection('trainable_variables', 'lstm')
# train_lstm = tf.train.AdamOptimizer(FLAGS.lr, 0.5).minimize(loss_lstm, var_list=var_lstm)
train_lstm = tf.train.AdamOptimizer(T.lr, 0.5).minimize(loss_lstm, var_list=var_lstm)
summary_lstm = [tf.summary.scalar('lstm/loss_lstm', loss_lstm)]
summary_lstm_image = [tf.summary.image('lstm/seq_out', T.seq_out[:, 0, :, :, :]),
tf.summary.image('lstm/seq_out_pred', seq_out_pred[:, 0, :, :, :])]
summary_lstm = tf.summary.merge(summary_lstm)
summary_lstm_image = tf.summary.merge(summary_lstm_image)
T.ops_lstm_print = [c('loss_lstm'), loss_lstm]
T.ops_lstm = [summary_lstm, train_lstm]
T.ops_lstm_image = summary_lstm_image
print(colored("Model initialization ended", "blue"))
return T
def gada():
T = tb.utils.TensorDict(dict(
sess = tf.Session(config=tb.growth_config()),
src_x = placeholder((None, 32, 32, 3)),
src_y = placeholder((None, args.Y)),
trg_x = placeholder((None, 32, 32, 3)),
trg_y = placeholder((None, args.Y)),
trg_z = placeholder((None, 100)),
test_x = placeholder((None, 32, 32, 3)),
test_y = placeholder((None, args.Y)),
))
# Supervised and conditional entropy minimization
src_e = nn.classifier(T.src_x, phase=True, enc_phase=1, enc_trim=args.etrim)
src_g = nn.classifier(src_e, phase=True, gen_trim=args.gtrim, gen_phase=1, enc_trim=args.etrim)
src_p = nn.classifier(src_g, phase=True, gen_trim=args.gtrim)
trg_e = nn.classifier(T.trg_x, phase=True, enc_phase=1, enc_trim=args.etrim, reuse=True, internal_update=True)
trg_g = nn.classifier(trg_e, phase=True, gen_trim=args.gtrim, gen_phase=1, enc_trim=args.etrim, reuse=True, internal_update=True)
trg_p = nn.classifier(trg_g, phase=True, gen_trim=args.gtrim, reuse=True, internal_update=True)
loss_src_class = tf.reduce_mean(softmax_xent(labels=T.src_y, logits=src_p))
loss_trg_cent = tf.reduce_mean(softmax_xent_two(labels=trg_p, logits=trg_p)) if args.tw > 0 else constant(0)
# Domain confusion
if args.dw > 0 and args.dirt == 0:
real_logit = nn.real_feature_discriminator(src_e, phase=True)
fake_logit = nn.real_feature_discriminator(trg_e, phase=True, reuse=True)
loss_disc = 0.5 * tf.reduce_mean(
sigmoid_xent(labels=tf.ones_like(real_logit), logits=real_logit) +
sigmoid_xent(labels=tf.zeros_like(fake_logit), logits=fake_logit))
loss_domain = 0.5 * tf.reduce_mean(
sigmoid_xent(labels=tf.zeros_like(real_logit), logits=real_logit) +
sigmoid_xent(labels=tf.ones_like(fake_logit), logits=fake_logit))
else:
loss_disc = constant(0)
loss_domain = constant(0)
# Virtual adversarial training (turn off src in non-VADA phase)
loss_src_vat = vat_loss(T.src_x, src_p, nn.classifier) if args.sw > 0 and args.dirt == 0 else constant(0)
loss_trg_vat = vat_loss(T.trg_x, trg_p, nn.classifier) if args.tw > 0 else constant(0)
# Generate images and process generated images
trg_gen_x = nn.trg_generator(T.trg_z)
trg_gen_e = nn.classifier(trg_gen_x, phase=True, enc_phase=1, enc_trim=args.etrim, reuse=True, internal_update=True)
trg_gen_g = nn.classifier(trg_gen_e, phase=True, gen_trim=args.gtrim, gen_phase=1, enc_trim=args.etrim, reuse=True, internal_update=True)
trg_gen_p = nn.classifier(trg_gen_g, phase=True, gen_trim=args.gtrim, reuse=True, internal_update=True)
# Feature matching loss function for generator
loss_trg_gen_fm = tf.reduce_mean(tf.square(tf.reduce_mean(trg_g, axis=0) - tf.reduce_mean(trg_gen_g, axis=0))) if args.dirt == 0 else constant(0)
# Unsupervised loss function
if args.dirt == 0:
logit_real = tf.reduce_logsumexp(trg_p, axis=1)
logit_fake = tf.reduce_logsumexp(trg_gen_p, axis=1)
dis_loss_real = -0.5*tf.reduce_mean(logit_real) + 0.5*tf.reduce_mean(tf.nn.softplus(logit_real))
dis_loss_fake = 0.5*tf.reduce_mean(tf.nn.softplus(logit_fake))
loss_trg_usv = dis_loss_real + dis_loss_fake # UnSuperVised loss function
else:
loss_trg_usv = constant(0)
# Evaluation (EMA)
ema = tf.train.ExponentialMovingAverage(decay=0.998)
var_class = tf.get_collection('trainable_variables', 'class/')
ema_op = ema.apply(var_class)
ema_p = nn.classifier(T.test_x, enc_phase=1, enc_trim=0, phase=False, reuse=True, getter=tb.tfutils.get_getter(ema))
# Teacher model (a back-up of EMA model)
teacher_p = nn.classifier(T.test_x, enc_phase=1, enc_trim=0, phase=False, scope='teacher')
var_main = tf.get_collection('variables', 'class/(?!.*ExponentialMovingAverage:0)')
var_teacher = tf.get_collection('variables', 'teacher/(?!.*ExponentialMovingAverage:0)')
teacher_assign_ops = []
for t, m in zip(var_teacher, var_main):
ave = ema.average(m)
ave = ave if ave else m
teacher_assign_ops += [tf.assign(t, ave)]
update_teacher = tf.group(*teacher_assign_ops)
teacher = tb.function(T.sess, [T.test_x], tf.nn.softmax(teacher_p))
# Accuracies
src_acc = basic_accuracy(T.src_y, src_p)
trg_acc = basic_accuracy(T.trg_y, trg_p)
ema_acc = basic_accuracy(T.test_y, ema_p)
fn_ema_acc = tb.function(T.sess, [T.test_x, T.test_y], ema_acc)
# Optimizer
dw = constant(args.dw) if args.dirt == 0 else constant(0)
cw = constant(1) if args.dirt == 0 else constant(args.bw)
sw = constant(args.sw) if args.dirt == 0 else constant(0)
tw = constant(args.tw)
uw = constant(args.uw) if args.dirt == 0 else constant(0)
loss_main = (dw * loss_domain +
cw * loss_src_class +
sw * loss_src_vat +
tw * loss_trg_cent +
tw * loss_trg_vat +
uw * loss_trg_usv)
var_main = tf.get_collection('trainable_variables', 'class')
train_main = tf.train.AdamOptimizer(args.lr, 0.5).minimize(loss_main, var_list=var_main)
train_main = tf.group(train_main, ema_op)
# Optimizer for feature discriminator
if args.dw > 0 and args.dirt == 0:
var_disc = tf.get_collection('trainable_variables', 'disc_real')
train_disc = tf.train.AdamOptimizer(args.lr, 0.5).minimize(loss_disc, var_list=var_disc)
else:
train_disc = constant(0)
# Optimizer for generators
if args.dirt == 0:
fmw = constant(1)
loss_trg_gen = (fmw * loss_trg_gen_fm)
var_trg_gen = tf.get_collection('trainable_variables', 'trg_gen')
trg_gen_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope='trg_gen')
with tf.control_dependencies(trg_gen_update_ops):
train_trg_gen = tf.train.AdamOptimizer(args.lr, 0.5).minimize(loss_trg_gen, var_list=var_trg_gen)
train_gen = train_trg_gen
else:
fmw = constant(0)
train_gen = constant(0)
# Summarizations
summary_disc = [tf.summary.scalar('domain/loss_disc', loss_disc),]
summary_main = [tf.summary.scalar('domain/loss_domain', loss_domain),
tf.summary.scalar('class/loss_src_class', loss_src_class),
tf.summary.scalar('class/loss_trg_cent', loss_trg_cent),
tf.summary.scalar('class/loss_trg_usv', loss_trg_usv),
tf.summary.scalar('lipschitz/loss_trg_vat', loss_trg_vat),
tf.summary.scalar('lipschitz/loss_src_vat', loss_src_vat),
tf.summary.scalar('hyper/dw', dw),
tf.summary.scalar('hyper/cw', cw),
tf.summary.scalar('hyper/sw', sw),
tf.summary.scalar('hyper/tw', tw),
tf.summary.scalar('hyper/uw', uw),
tf.summary.scalar('hyper/fmw', fmw),
tf.summary.scalar('acc/src_acc', src_acc),
tf.summary.scalar('acc/trg_acc', trg_acc)]
summary_gen = [tf.summary.scalar('gen/loss_trg_gen_fm', loss_trg_gen_fm),
tf.summary.image('gen/trg_gen_img', trg_gen_x),]
# Merge summaries
summary_disc = tf.summary.merge(summary_disc)
summary_main = tf.summary.merge(summary_main)
summary_gen = tf.summary.merge(summary_gen)
# Saved ops
c = tf.constant
T.ops_print = [c('disc'), loss_disc,
c('domain'), loss_domain,
c('class'), loss_src_class,
c('cent'), loss_trg_cent,
c('trg_vat'), loss_trg_vat,
c('src_vat'), loss_src_vat,
c('src'), src_acc,
c('trg'), trg_acc]
T.ops_disc = [summary_disc, train_disc]
T.ops_main = [summary_main, train_main]
T.ops_gen = [summary_gen , train_gen]
T.fn_ema_acc = fn_ema_acc
T.teacher = teacher
T.update_teacher = update_teacher
T.trg_gen_x = trg_gen_x
T.trg_gen_p = trg_gen_p
T.src_p = src_p
T.trg_p = trg_p
T.ema_p = ema_p
return T