def train(args):
print(args)
numpy.random.seed(int(args['--seed']))
dataset = load_data.load_mnist_for_semi_sup(n_l=int(args['--num_labeled_samples']),
n_v=int(args['--num_validation_samples']))
x_train, t_train, ul_x_train = dataset[0]
x_test, t_test = dataset[2]
layer_sizes = [int(layer_size) for layer_size in args['--layer_sizes'].split('-')]
model = FNN_MNIST(layer_sizes=layer_sizes)
x = t_func.matrix()
ul_x = t_func.matrix()
t = t_func.ivector()
cost_semi = get_cost_type_semi(model, x, t, ul_x, args)
nll = costs.cross_entropy_loss(x=x, t=t, forward_func=model.forward_test)
error = costs.error(x=x, t=t, forward_func=model.forward_test)
optimizer = optimizers.ADAM(cost=cost_semi, params=model.params, alpha=float(args['--initial_learning_rate']))
index = t_func.iscalar()
ul_index = t_func.iscalar()
batch_size = int(args['--batch_size'])
ul_batch_size = int(args['--ul_batch_size'])
f_train = theano.function(inputs=[index, ul_index], outputs=cost_semi, updates=optimizer.updates,
givens={
x: x_train[batch_size * index:batch_size * (index + 1)],
t: t_train[batch_size * index:batch_size * (index + 1)],
ul_x: ul_x_train[ul_batch_size * ul_index:ul_batch_size * (ul_index + 1)]},
on_unused_input='ignore')
f_nll_train = theano.function(inputs=[index], outputs=nll,
givens={
x: x_train[batch_size * index:batch_size * (index + 1)],
t: t_train[batch_size * index:batch_size * (index + 1)]})
f_nll_test = theano.function(inputs=[index], outputs=nll,
givens={
x: x_test[batch_size * index:batch_size * (index + 1)],
t: t_test[batch_size * index:batch_size * (index + 1)]})
f_error_train = theano.function(inputs=[index], outputs=error,
givens={
x: x_train[batch_size * index:batch_size * (index + 1)],
t: t_train[batch_size * index:batch_size * (index + 1)]})
f_error_test = theano.function(inputs=[index], outputs=error,
givens={
x: x_test[batch_size * index:batch_size * (index + 1)],
t: t_test[batch_size * index:batch_size * (index + 1)]})
f_lr_decay = theano.function(inputs=[], outputs=optimizer.alpha,
updates={optimizer.alpha: theano.shared(
numpy.array(args['--learning_rate_decay']).astype(
theano.config.floatX)) * optimizer.alpha})
# Shuffle training set
randix = RandomStreams(seed=numpy.random.randint(1234)).permutation(n=x_train.shape[0])
update_permutation = OrderedDict()
update_permutation[x_train] = x_train[randix]
update_permutation[t_train] = t_train[randix]
f_permute_train_set = theano.function(inputs=[], outputs=x_train, updates=update_permutation)
# Shuffle unlabeled training set
ul_randix = RandomStreams(seed=numpy.random.randint(1234)).permutation(n=ul_x_train.shape[0])
update_ul_permutation = OrderedDict()
update_ul_permutation[ul_x_train] = ul_x_train[ul_randix]
f_permute_ul_train_set = theano.function(inputs=[], outputs=ul_x_train, updates=update_ul_permutation)
statuses = {'nll_train': [], 'error_train': [], 'nll_test': [], 'error_test': []}
n_train = x_train.get_value().shape[0]
n_test = x_test.get_value().shape[0]
n_ul_train = ul_x_train.get_value().shape[0]
l_i = 0
ul_i = 0
for epoch in range(int(args['--num_epochs'])):
# cPickle.dump((statuses, args), open('./trained_model/' + 'tmp-' + args['--save_filename'], 'wb'),
# cPickle.HIGHEST_PROTOCOL)
f_permute_train_set()
f_permute_ul_train_set()
for it in range(int(args['--num_batch_it'])):
f_train(l_i, ul_i)
l_i = 0 if l_i >= n_train / batch_size - 1 else l_i + 1
ul_i = 0 if ul_i >= n_ul_train / ul_batch_size - 1 else ul_i + 1
sum_nll_train = numpy.sum(numpy.array([f_nll_train(i) for i in range(int(n_train / batch_size))])) * batch_size
sum_error_train = numpy.sum(numpy.array([f_error_train(i) for i in range(int(n_train / batch_size))]))
sum_nll_test = numpy.sum(numpy.array([f_nll_test(i) for i in range(int(n_test / batch_size))])) * batch_size
sum_error_test = numpy.sum(numpy.array([f_error_test(i) for i in range(int(n_test / batch_size))]))
statuses['nll_train'].append(sum_nll_train / n_train)
statuses['error_train'].append(sum_error_train)
statuses['nll_test'].append(sum_nll_test / n_test)
statuses['error_test'].append(sum_error_test)
wlog("[Epoch] %d" % epoch)
acc = 1 - 1.0*statuses['error_test'][-1]/n_test
wlog("nll_test : %f error_test : %d accuracy:%f" % (statuses['nll_test'][-1], statuses['error_test'][-1], acc))
# writer.add_scalar("Test/Loss", statuses['nll_test'][-1], epoch * int(args['--num_batch_it']))
# writer.add_scalar("Test/Acc", acc, epoch * int(args['--num_batch_it']))
f_lr_decay()
# fine_tune batch stat
f_fine_tune = theano.function(inputs=[ul_index], outputs=model.forward_for_finetuning_batch_stat(x),
givens={x: ul_x_train[ul_batch_size * ul_index:ul_batch_size * (ul_index + 1)]})
[f_fine_tune(i) for i in range(n_ul_train // ul_batch_size)]
sum_nll_test = numpy.sum(numpy.array([f_nll_test(i) for i in range(n_test // batch_size)])) * batch_size
sum_error_test = numpy.sum(numpy.array([f_error_test(i) for i in range(n_test // batch_size)]))
statuses['nll_test'].append(sum_nll_test / n_test)
statuses['error_test'].append(sum_error_test)
acc = 1 - 1.0*statuses['error_test'][-1]/n_test
wlog("final nll_test: %f error_test: %d accuracy:%f" % (statuses['nll_test'][-1], statuses['error_test'][-1], acc))
def train(args):
print args
numpy.random.seed(int(args['--seed']))
if (args['--validation']):
dataset = load_data.load_mnist_for_validation(
n_v=int(args['--num_validation_samples']))
else:
dataset = load_data.load_mnist_full()
x_train, t_train = dataset[0]
x_test, t_test = dataset[1]
layer_sizes = [
int(layer_size) for layer_size in args['--layer_sizes'].split('-')
]
model = FNN_MNIST(layer_sizes=layer_sizes)
x = T.matrix()
t = T.ivector()
if (args['--cost_type'] == 'MLE'):
cost = costs.cross_entropy_loss(x=x,
t=t,
forward_func=model.forward_train)
elif (args['--cost_type'] == 'L2'):
cost = costs.cross_entropy_loss(x=x, t=t, forward_func=model.forward_train) \
+ costs.weight_decay(params=model.params, coeff=float(args['--lamb']))
elif (args['--cost_type'] == 'AT'):
cost = costs.adversarial_training(
x,
t,
model.forward_train,
'CE',
epsilon=float(args['--epsilon']),
lamb=float(args['--lamb']),
norm_constraint=args['--norm_constraint'],
forward_func_for_generating_adversarial_examples=model.
forward_no_update_batch_stat)
elif (args['--cost_type'] == 'VAT'):
cost = costs.virtual_adversarial_training(
x,
t,
model.forward_train,
'CE',
epsilon=float(args['--epsilon']),
norm_constraint=args['--norm_constraint'],
num_power_iter=int(args['--num_power_iter']),
forward_func_for_generating_adversarial_examples=model.
forward_no_update_batch_stat)
elif (args['--cost_type'] == 'VAT_finite_diff'):
cost = costs.virtual_adversarial_training_finite_diff(
x,
t,
model.forward_train,
'CE',
epsilon=float(args['--epsilon']),
norm_constraint=args['--norm_constraint'],
num_power_iter=int(args['--num_power_iter']),
forward_func_for_generating_adversarial_examples=model.
forward_no_update_batch_stat)
nll = costs.cross_entropy_loss(x=x, t=t, forward_func=model.forward_test)
error = costs.error(x=x, t=t, forward_func=model.forward_test)
optimizer = optimizers.ADAM(cost=cost,
params=model.params,
alpha=float(args['--initial_learning_rate']))
index = T.iscalar()
batch_size = int(args['--batch_size'])
f_train = theano.function(
inputs=[index],
outputs=cost,
updates=optimizer.updates,
givens={
x: x_train[batch_size * index:batch_size * (index + 1)],
t: t_train[batch_size * index:batch_size * (index + 1)]
})
f_nll_train = theano.function(
inputs=[index],
outputs=nll,
givens={
x: x_train[batch_size * index:batch_size * (index + 1)],
t: t_train[batch_size * index:batch_size * (index + 1)]
})
f_nll_test = theano.function(
inputs=[index],
outputs=nll,
givens={
x: x_test[batch_size * index:batch_size * (index + 1)],
t: t_test[batch_size * index:batch_size * (index + 1)]
})
f_error_train = theano.function(
inputs=[index],
outputs=error,
givens={
x: x_train[batch_size * index:batch_size * (index + 1)],
t: t_train[batch_size * index:batch_size * (index + 1)]
})
f_error_test = theano.function(
inputs=[index],
outputs=error,
givens={
x: x_test[batch_size * index:batch_size * (index + 1)],
t: t_test[batch_size * index:batch_size * (index + 1)]
})
f_lr_decay = theano.function(
inputs=[],
outputs=optimizer.alpha,
updates={
optimizer.alpha:
theano.shared(
numpy.array(args['--learning_rate_decay']).astype(
theano.config.floatX)) * optimizer.alpha
})
randix = RandomStreams(seed=numpy.random.randint(1234)).permutation(
n=x_train.shape[0])
update_permutation = OrderedDict()
update_permutation[x_train] = x_train[randix]
update_permutation[t_train] = t_train[randix]
f_permute_train_set = theano.function(inputs=[],
outputs=x_train,
updates=update_permutation)
statuses = {}
statuses['nll_train'] = []
statuses['error_train'] = []
statuses['nll_test'] = []
statuses['error_test'] = []
n_train = x_train.get_value().shape[0]
n_test = x_test.get_value().shape[0]
sum_nll_train = numpy.sum(
numpy.array([f_nll_train(i)
for i in xrange(n_train / batch_size)])) * batch_size
sum_error_train = numpy.sum(
numpy.array([f_error_train(i) for i in xrange(n_train / batch_size)]))
sum_nll_test = numpy.sum(
numpy.array([f_nll_test(i)
for i in xrange(n_test / batch_size)])) * batch_size
sum_error_test = numpy.sum(
numpy.array([f_error_test(i) for i in xrange(n_test / batch_size)]))
statuses['nll_train'].append(sum_nll_train / n_train)
statuses['error_train'].append(sum_error_train)
statuses['nll_test'].append(sum_nll_test / n_test)
statuses['error_test'].append(sum_error_test)
print "[Epoch]", str(-1)
print "nll_train : ", statuses['nll_train'][-1], "error_train : ", statuses['error_train'][-1], \
"nll_test : ", statuses['nll_test'][-1], "error_test : ", statuses['error_test'][-1]
print "training..."
make_sure_path_exists("./trained_model")
for epoch in xrange(int(args['--num_epochs'])):
cPickle.dump(
(statuses, args),
open('./trained_model/' + 'tmp-' + args['--save_filename'],
'wb'), cPickle.HIGHEST_PROTOCOL)
f_permute_train_set()
### update parameters ###
[f_train(i) for i in xrange(n_train / batch_size)]
#########################
sum_nll_train = numpy.sum(
numpy.array([f_nll_train(i)
for i in xrange(n_train / batch_size)])) * batch_size
sum_error_train = numpy.sum(
numpy.array(
[f_error_train(i) for i in xrange(n_train / batch_size)]))
sum_nll_test = numpy.sum(
numpy.array([f_nll_test(i)
for i in xrange(n_test / batch_size)])) * batch_size
sum_error_test = numpy.sum(
numpy.array([f_error_test(i)
for i in xrange(n_test / batch_size)]))
statuses['nll_train'].append(sum_nll_train / n_train)
statuses['error_train'].append(sum_error_train)
statuses['nll_test'].append(sum_nll_test / n_test)
statuses['error_test'].append(sum_error_test)
print "[Epoch]", str(epoch)
print "nll_train : ", statuses['nll_train'][-1], "error_train : ", statuses['error_train'][-1], \
"nll_test : ", statuses['nll_test'][-1], "error_test : ", statuses['error_test'][-1]
f_lr_decay()
### finetune batch stat ###
f_finetune = theano.function(
inputs=[index],
outputs=model.forward_for_finetuning_batch_stat(x),
givens={x: x_train[batch_size * index:batch_size * (index + 1)]})
[f_finetune(i) for i in xrange(n_train / batch_size)]
sum_nll_train = numpy.sum(
numpy.array([f_nll_train(i)
for i in xrange(n_train / batch_size)])) * batch_size
sum_error_train = numpy.sum(
numpy.array([f_error_train(i) for i in xrange(n_train / batch_size)]))
sum_nll_test = numpy.sum(
numpy.array([f_nll_test(i)
for i in xrange(n_test / batch_size)])) * batch_size
sum_error_test = numpy.sum(
numpy.array([f_error_test(i) for i in xrange(n_test / batch_size)]))
statuses['nll_train'].append(sum_nll_train / n_train)
statuses['error_train'].append(sum_error_train)
statuses['nll_test'].append(sum_nll_test / n_test)
statuses['error_test'].append(sum_error_test)
print "[after finetuning]"
print "nll_train : ", statuses['nll_train'][-1], "error_train : ", statuses['error_train'][-1], \
"nll_test : ", statuses['nll_test'][-1], "error_test : ", statuses['error_test'][-1]
###########################
make_sure_path_exists("./trained_model")
cPickle.dump((model, statuses, args),
open('./trained_model/' + args['--save_filename'], 'wb'),
cPickle.HIGHEST_PROTOCOL)
def train(latent_dim = 2, #dimension of latent variable z
z_prior = 'gaussian', # 'gaussian' or 'uniform'
lamb = 10., #ratio between reconstruction and adversarial cost
recon_obj_type = 'CE', #objective function on reconstruction ( 'CE'(cross ent.) or 'QE'(quadratic error) )
initlal_learning_rate = 0.002,
learning_rate_decay=1.0,
num_epochs=50,
batch_size=100,
save_filename='trained_model',
seed=1):
numpy.random.seed(seed=seed)
dataset = load_data.load_mnist_full()
x_train,_ = dataset[0]
x_test,_ = dataset[1]
model = AdversarialAutoencoderMNIST(latent_dim=latent_dim,z_prior=z_prior)
x = T.matrix()
loss_for_training,_,adv_loss_for_training = costs.adversarial_autoenc_loss(x=x,
enc_f=model.encode_train,
dec_f=model.decode_train,
disc_f=model.D_train,
p_z_sampler=model.sample_from_prior,
obj_type=recon_obj_type,
lamb=numpy.asarray(lamb,dtype=theano.config.floatX))
_,recon_loss,adv_loss = costs.adversarial_autoenc_loss(x=x,
enc_f=model.encode_test,
dec_f=model.decode_test,
disc_f=model.D_test,
p_z_sampler=model.sample_from_prior,
obj_type=recon_obj_type,
lamb=numpy.asarray(lamb,dtype=theano.config.floatX))
optimizer_recon = optimizers.ADAM(cost=loss_for_training,
params=model.model_params,
alpha=numpy.asarray(initlal_learning_rate,dtype=theano.config.floatX))
optimizer_adv = optimizers.ADAM(cost=adv_loss_for_training,
params=model.D_params,
alpha=numpy.asarray(initlal_learning_rate,dtype=theano.config.floatX))
index = T.iscalar()
f_training_model = theano.function(inputs=[index], outputs=loss_for_training, updates=optimizer_recon.updates,
givens={
x:x_train[batch_size*index:batch_size*(index+1)]})
f_training_discriminator = theano.function(inputs=[index], outputs=adv_loss_for_training, updates=optimizer_adv.updates,
givens={
x:x_train[batch_size*index:batch_size*(index+1)]})
f_recon_train = theano.function(inputs=[index], outputs=recon_loss,
givens={
x:x_train[batch_size*index:batch_size*(index+1)]})
f_adv_train = theano.function(inputs=[index], outputs=adv_loss,
givens={
x:x_train[batch_size*index:batch_size*(index+1)]})
f_recon_test = theano.function(inputs=[index], outputs=recon_loss,
givens={
x:x_test[batch_size*index:batch_size*(index+1)]})
f_adv_test = theano.function(inputs=[index], outputs=adv_loss,
givens={
x:x_test[batch_size*index:batch_size*(index+1)]})
f_lr_decay_recon = theano.function(inputs=[],outputs=optimizer_recon.alpha,
updates={optimizer_recon.alpha:theano.shared(numpy.array(learning_rate_decay).astype(theano.config.floatX))*optimizer_recon.alpha})
f_lr_decay_adv = theano.function(inputs=[],outputs=optimizer_adv.alpha,
updates={optimizer_adv.alpha:theano.shared(numpy.array(learning_rate_decay).astype(theano.config.floatX))*optimizer_adv.alpha})
randix = RandomStreams(seed=numpy.random.randint(1234)).permutation(n=x_train.shape[0])
f_permute_train_set = theano.function(inputs=[],outputs=x_train,updates={x_train:x_train[randix]})
statuses = {}
statuses['recon_train'] = []
statuses['adv_train'] = []
statuses['recon_test'] = []
statuses['adv_test'] = []
n_train = x_train.get_value().shape[0]
n_test = x_test.get_value().shape[0]
sum_recon_train = numpy.sum(numpy.array([f_recon_train(i) for i in xrange(n_train/batch_size)]))*batch_size
sum_adv_train = numpy.sum(numpy.array([f_adv_train(i) for i in xrange(n_train/batch_size)]))*batch_size
sum_recon_test = numpy.sum(numpy.array([f_recon_test(i) for i in xrange(n_test/batch_size)]))*batch_size
sum_adv_test = numpy.sum(numpy.array([f_adv_test(i) for i in xrange(n_test/batch_size)]))*batch_size
statuses['recon_train'].append(sum_recon_train/n_train)
statuses['adv_train'].append(sum_adv_train/n_train)
statuses['recon_test'].append(sum_recon_test/n_test)
statuses['adv_test'].append(sum_adv_test/n_test)
print "[Epoch]",str(-1)
print "recon_train : " , statuses['recon_train'][-1], "adv_train : ", statuses['adv_train'][-1], \
"recon_test : " , statuses['recon_test'][-1], "adv_test : ", statuses['adv_test'][-1]
z = model.encode_test(input=x)
f_enc = theano.function(inputs=[],outputs=z,givens={x:dataset[1][0]})
def plot_latent_variable(epoch):
output = f_enc()
plt.figure(figsize=(8,8))
color=cm.rainbow(numpy.linspace(0,1,10))
for l,c in zip(range(10),color):
ix = numpy.where(dataset[1][1].get_value()==l)[0]
plt.scatter(output[ix,0],output[ix,1],c=c,label=l,s=8,linewidth=0)
plt.xlim([-5.0,5.0])
plt.ylim([-5.0,5.0])
plt.legend(fontsize=15)
plt.savefig('z_epoch' + str(epoch) + '.pdf')
print "training..."
make_sure_path_exists("./trained_model")
for epoch in xrange(num_epochs):
cPickle.dump((model,statuses),open('./trained_model/'+'tmp-' + save_filename,'wb'),cPickle.HIGHEST_PROTOCOL)
f_permute_train_set()
### update parameters ###
for i in xrange(n_train/batch_size):
### Optimize model and discriminator alternately ###
f_training_discriminator(i)
f_training_model(i)
#########################
if(latent_dim == 2):
plot_latent_variable(epoch=epoch)
sum_recon_train = numpy.sum(numpy.array([f_recon_train(i) for i in xrange(n_train/batch_size)]))*batch_size
sum_adv_train = numpy.sum(numpy.array([f_adv_train(i) for i in xrange(n_train/batch_size)]))*batch_size
sum_recon_test = numpy.sum(numpy.array([f_recon_test(i) for i in xrange(n_test/batch_size)]))*batch_size
sum_adv_test = numpy.sum(numpy.array([f_adv_test(i) for i in xrange(n_test/batch_size)]))*batch_size
statuses['recon_train'].append(sum_recon_train/n_train)
statuses['adv_train'].append(sum_adv_train/n_train)
statuses['recon_test'].append(sum_recon_test/n_test)
statuses['adv_test'].append(sum_adv_test/n_test)
print "[Epoch]",str(epoch)
print "recon_train : " , statuses['recon_train'][-1], "adv_train : ", statuses['adv_train'][-1], \
"recon_test : " , statuses['recon_test'][-1], "adv_test : ", statuses['adv_test'][-1]
f_lr_decay_recon()
f_lr_decay_adv()
make_sure_path_exists("./trained_model")
cPickle.dump((model,statuses),open('./trained_model/'+save_filename,'wb'),cPickle.HIGHEST_PROTOCOL)
return model,statuses