def build_summary(metrics, labels, logits, learning_rate):
true_image = map_to_image(utils.tile_images(labels))
pred_image = map_to_image(
utils.tile_images(tf.to_float(tf.nn.sigmoid(logits) > 0.5)))
return tf.summary.merge([
tf.summary.scalar('learning_rate', learning_rate),
tf.summary.scalar('loss', metrics['loss']),
tf.summary.scalar('iou', metrics['iou']),
tf.summary.scalar('roc_auc', metrics['roc_auc']),
tf.summary.scalar('pr_auc', metrics['pr_auc']),
tf.summary.image('true_image', tf.expand_dims(true_image, 0)),
tf.summary.image('pred_image', tf.expand_dims(pred_image, 0))
])
def draw(trainer=None, model=None, batch=None, prefix=None):
# save input & predicted images
with chainer.using_config('enable_backprop', False):
with chainer.using_config('train', False):
directory = trainer.out
if model.dataset_name == 'shapenet':
images_in, _, viewpoints, _, _ = converter(batch)
images_prediction = model.predict_and_render(
images_in, viewpoints).data.get().squeeze()
images_prediction_cp = model.predict_and_render(
images_in, viewpoints[::-1]).data.get().squeeze()
images_ref = None
cmin_in = -1
elif model.dataset_name == 'pascal':
images_in, images_ref, rotation_matrices, rotation_matrices_random, labels = converter(
batch)
images_prediction = model.predict_and_render(
images_in, rotation_matrices).data.get().squeeze()
images_prediction_cp = model.predict_and_render(
images_in, rotation_matrices_random).data.get().squeeze()
cmin_in = 0
# input images
image = utils.tile_images(images_in.get())
scipy.misc.toimage(image, cmin=cmin_in, cmax=1).save(
os.path.join(directory, '%s_images_in.png' % prefix))
if images_ref is not None:
image = utils.tile_images(images_ref.get())
scipy.misc.toimage(image, cmin=cmin_in, cmax=1).save(
os.path.join(directory, '%s_images_ref.png' % prefix))
# predicted images
images_prediction = utils.tile_images(images_prediction)
if model.no_texture:
images_prediction = 1 - images_prediction
scipy.misc.toimage(images_prediction, cmin=0, cmax=1).save(
os.path.join(directory, '%s_prediction.png' % prefix))
# predicted images from viewpoints
images_prediction_cp = utils.tile_images(images_prediction_cp)
if model.no_texture:
images_prediction_cp = 1 - images_prediction_cp
scipy.misc.toimage(images_prediction_cp, cmin=0, cmax=1).save(
os.path.join(directory, '%s_prediction_cp.png' % prefix))
def render(self, mode='human'):
imgs = self.get_images()
bigimg = tile_images(imgs)
if mode == 'human':
self.get_viewer().imshow(bigimg)
return self.get_viewer().isopen
elif mode == 'rgb_array':
return bigimg
else:
raise NotImplementedError
def render(self, mode='human'):
for pipe in self.remotes:
pipe.send(('render', None))
imgs = [pipe.recv() for pipe in self.remotes]
bigimg = tile_images(imgs)
if mode == 'human':
import cv2
cv2.imshow('vecenv', bigimg[:, :, ::-1])
cv2.waitKey(1)
elif mode == 'rgb_array':
return bigimg
else:
raise NotImplementedError
def generate_between_classes(
model: torch.nn.Module,
img_size: list,
classes: int,
saveto: str,
n_classes: int,
cuda: bool = False,
) -> None:
y = np.zeros((1, n_classes), dtype="float32")
y[:, classes] = 1 / len(classes)
y = np.repeat(y, 10, axis=0)
gen = utils.tile_images(generate(model, img_size, y, cuda=cuda), n_rows=1)
imageio.imsave(saveto, gen.astype("uint8"))
def render(self, mode="human"):
for pipe in self.remotes:
pipe.send(("render", None))
imgs = [pipe.recv() for pipe in self.remotes]
bigimg = tile_images(imgs)
if mode == "human":
import cv2
cv2.imshow("vecenv", bigimg[:, :, ::-1])
cv2.waitKey(1)
elif mode == "rgb_array":
return bigimg
else:
raise NotImplementedError
def render(self, mode='human'):
for pipe in self.remotes:
pipe.send(('render', None))
if self.experiment_lvl == 'ego':
normal = []
for pipe in self.remotes:
res = pipe.recv()
normal.append(res[0]['normal'])
imgs = np.stack(normal)
else:
imgs = [pipe.recv() for pipe in self.remotes]
bigimg = tile_images(imgs)
if mode == 'human':
import cv2
cv2.imshow('vecenv', bigimg[:, :, ::-1])
cv2.waitKey(1)
elif mode == 'rgb_array':
return bigimg
else:
raise NotImplementedError
def render(self, mode='human'):
for pipe in self.remotes:
pipe.send(('render', None))
imgs = [pipe.recv() for pipe in self.remotes]
bigimg = tile_images(imgs)
if mode == 'human':
#import cv2
#cv2.imshow('vecenv', bigimg[:,:,::-1])
#cv2.waitKey(1)
if self.viewer is None:
from gym.envs.classic_control import rendering
self.viewer = rendering.SimpleImageViewer()
#print(imgs)
self.viewer.imshow(bigimg[:, :, ::-1])
#print(bigimg[:, :, ::-1].shape)
return bigimg[:, :, ::-1]
elif mode == 'rgb_array':
return bigimg
else:
raise NotImplementedError
def test_RBM_timep(self, learning_rate=0.1, training_epochs=10,
batch_size=50,
n_chains=20, n_samples=5, output_folder='rbm_plots',
n_hidden=500):
"""
Demonstrate how to train and afterwards sample from it using Theano.
:param learning_rate: learning rate used for training the RBM
:param training_epochs: number of epochs used for training
:param dataset: path the the pickled dataset
:param batch_size: size of a batch used to train the RBM
:param n_chains: number of parallel Gibbs chains to be used for sampling
:param n_samples: number of samples to plot for each chain
from DLRBM_2D_temporal import *
funcs = DLRBM_2D_temporal()
funcs.test_RBM_timep()
"""
traindata_path= 'allLpatches_10x10x5.pklz' #'allLpatches_subs_smaller.pklz' #'allLpatches.pklz'
trainUdata_path= 'allUpatches_10x10x5.pklz'#'allUpatches_subs_smaller.pklz' #'allUpatches.pklz'
labeldata_path= 'allLabels_10x10x5.pklz' #'allLabels_subs_smaller.pklz' #'allLabels.pklz'
#############
## LOAD datasets
#############
datasets = self.load_wUdata(traindata_path, labeldata_path, trainUdata_path)
train_set_x, train_set_y = datasets[0]
np_train_x, np_train_y = datasets[3]
valid_set_x, valid_set_y = datasets[1]
np_valid_x, np_valid_y = datasets[4]
test_set_x, test_set_y = datasets[2]
np_test_x, np_test_y = datasets[5]
# inpect one image class 1
Vol = np_train_x[0].reshape(5,10,10)
imgslicestime = [Vol[0,:,:], Vol[1,:,:], Vol[2,:,:], Vol[3,:,:], Vol[4,:,:]]
subslicestime = []
# Display image
fig, ax = plt.subplots(nrows=4, ncols=4, figsize=(16, 8))
for k in range(1,5):
ax[k-1,0].imshow(imgslicestime[k], cmap=plt.cm.gray)
ax[k-1,0].set_axis_off()
ax[k-1,0].set_adjustable('box-forced')
# Display Original histogram
ax[k-1,1].hist(imgslicestime[k].ravel(), bins=50, color='black')
ax[k-1,1].ticklabel_format(axis='y', style='scientific', scilimits=(0, 0))
ax[k-1,1].set_xlabel('original')
# substract volume
subVol = np.asarray(imgslicestime[k]) - np.asarray(imgslicestime[0])
subslicestime.append(subVol)
# Display normalized histogram
ax[k-1,2].hist(subVol.ravel(), bins=50, color='black')
ax[k-1,2].ticklabel_format(axis='y', style='scientific', scilimits=(0, 0))
ax[k-1,2].set_xlabel('substracted histogram')
# display normalized 0-1
ax[k-1,3].imshow(subVol, cmap=plt.cm.gray)
ax[k-1,3].set_axis_off()
ax[k-1,3].set_adjustable('box-forced')
plt.show(block=False)
#################
# Substract pathces from pre-contrast
#################
subsnp_train_x = []
subsnp_valid_x = []
subsnp_test_x = []
for img in np_train_x:
Vol = img.reshape(5,10,10)
imgslicestime = [Vol[0,:,:], Vol[1,:,:], Vol[2,:,:], Vol[3,:,:], Vol[4,:,:]]
subslicestime = []
for k in range(1,5):
# substract volume
subVol = np.asarray(imgslicestime[k]) - np.asarray(imgslicestime[0])
subslicestime.append(subVol)
#append
subsnp_train_x.append( np.asarray(subslicestime).reshape(4*10*10) )
for img in np_valid_x:
Vol = img.reshape(5,10,10)
imgslicestime = [Vol[0,:,:], Vol[1,:,:], Vol[2,:,:], Vol[3,:,:], Vol[4,:,:]]
subslicestime = []
for k in range(1,5):
# substract volume
subVol = np.asarray(imgslicestime[k]) - np.asarray(imgslicestime[0])
subslicestime.append(subVol)
#append
subsnp_valid_x.append( np.asarray(subslicestime).reshape(4*10*10) )
for img in np_test_x:
Vol = img.reshape(5,10,10)
imgslicestime = [Vol[0,:,:], Vol[1,:,:], Vol[2,:,:], Vol[3,:,:], Vol[4,:,:]]
subslicestime = []
for k in range(1,5):
# substract volume
subVol = np.asarray(imgslicestime[k]) - np.asarray(imgslicestime[0])
subslicestime.append(subVol)
#append
subsnp_test_x.append( np.asarray(subslicestime).reshape(4*10*10) )
# train_set_x, train_set_y = self.shared_dataset(subsnp_train_x, np_train_y)
# valid_set_x, valid_set_y = self.shared_dataset(subsnp_valid_x, np_valid_y)
# test_set_x, test_set_y = self.shared_dataset(subsnp_test_x, np_test_y)
#
# subsdatasets = [(train_set_x, train_set_y),
# (valid_set_x, valid_set_y),
# (test_set_x, test_set_y) ]
print "n_chains %d " % n_chains
print "n_hidden %d " % n_hidden
print "training_epochs %d " % training_epochs
print "batch_size %d " % batch_size
print "learning_rate %f " % learning_rate
print "n_samples %d " % n_samples
print "output_folder %s " % output_folder
##############################
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] // batch_size
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 30))
# initialize storage for the persistent chain (state = hidden
# layer of chain)
persistent_chain = theano.shared(numpy.zeros((batch_size, n_hidden),
dtype=theano.config.floatX),
borrow=True)
# construct the RBM class
rbm = RBM(input=x, n_visible=5*10*10,
n_hidden=n_hidden, numpy_rng=rng, theano_rng=theano_rng)
# get the cost and the gradient corresponding to one step of CD-15
cost, updates = rbm.get_cost_updates(lr=learning_rate,
persistent=persistent_chain, k=15)
#################################
# Training the RBM #
#################################
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
# it is ok for a theano function to have no output
# the purpose of train_rbm is solely to update the RBM parameters
train_rbm = theano.function(
[index],
cost,
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size]
},
name='train_rbm'
)
plotting_time = 0.
start_time = timeit.default_timer()
# go through training epochs
for epoch in range(training_epochs):
# go through the training set
mean_cost = []
for batch_index in range(n_train_batches):
mean_cost += [train_rbm(batch_index)]
print('Training epoch %d, cost is ' % epoch, numpy.mean(mean_cost))
# Plot filters after each training epoch
plotting_start = timeit.default_timer()
# Construct image from the weight matrix
Wrbm = rbm.W.get_value(borrow=True).T
fig, axes = plt.subplots(ncols=4, nrows=1, figsize=(16, 6))
for k in range(1,5):
image = utils.tile_images(
X=Wrbm.reshape( Wrbm.shape[0], 5, 10, 10)[:,k,:,:],
img_shape=(10,10),
tile_shape=(25,20),
tile_spacing=(0, 0) )
im = axes[k-1].imshow(image, vmin=np.min(image.ravel()), vmax=np.max(image.ravel()), interpolation='nearest', cmap="Greys_r")
axes[k-1].get_xaxis().set_visible(False)
axes[k-1].get_yaxis().set_visible(False)
cax,kw = mpl.colorbar.make_axes([ax for ax in axes.flat])
plt.colorbar(im, cax=cax, **kw)
plt.show(block=False)
fig.savefig('filters_at_epoch_%i.png' % epoch)
plotting_stop = timeit.default_timer()
plotting_time += (plotting_stop - plotting_start)
end_time = timeit.default_timer()
pretraining_time = (end_time - start_time) - plotting_time
print ('Plotting took %f minutes' % (plotting_time / 60.))
print ('Training took %f minutes' % (pretraining_time / 60.))
#####################################
# save the best model
#####################################
with open('bestRBM_10x10x5.obj', 'wb') as fp:
pickle.dump(rbm, fp)
#####################################
# Open best model
#####################################
with open('bestRBM_10x10x5.obj', 'rb') as fp:
rbm = pickle.load(fp)
#################################
# Sampling from the RBM #
#################################
# find out the number of test samples
number_of_test_samples = test_set_x.get_value(borrow=True).shape[0]
# pick random test examples, with which to initialize the persistent chain
test_idx = rng.randint(number_of_test_samples - n_chains)
persistent_vis_chain = theano.shared(
numpy.asarray(
test_set_x.get_value(borrow=True)[test_idx:test_idx + n_chains],
dtype=theano.config.floatX
)
)
plot_every = 1000
# define one step of Gibbs sampling (mf = mean-field) define a
# function that does `plot_every` steps before returning the
# sample for plotting
(
[
presig_hids,
hid_mfs,
hid_samples,
presig_vis,
vis_mfs,
vis_samples
],
updates
) = theano.scan(
rbm.gibbs_vhv,
outputs_info=[None, None, None, None, None, persistent_vis_chain],
n_steps=plot_every
)
# add to updates the shared variable that takes care of our persistent
# chain :.
updates.update({persistent_vis_chain: vis_samples[-1]})
# construct the function that implements our persistent chain.
# we generate the "mean field" activations for plotting and the actual
# samples for reinitializing the state of our persistent chain
sample_fn = theano.function(
[],
[
vis_mfs[-1],
vis_samples[-1]
],
updates=updates,
name='sample_fn'
)
# create a space to store the image for plotting ( we need to leave
# room for the tile_spacing as well)
sub_pre = numpy.zeros(
(11 * n_samples + 1, 11 * n_chains - 1),
dtype='uint8')
sub_post1 = numpy.zeros(
(11 * n_samples + 1, 11 * n_chains - 1),
dtype='uint8')
sub_post2 = numpy.zeros(
(11 * n_samples + 1, 11 * n_chains - 1),
dtype='uint8')
sub_post3 = numpy.zeros(
(11 * n_samples + 1, 11 * n_chains - 1),
dtype='uint8')
sub_post4 = numpy.zeros(
(11 * n_samples + 1, 11 * n_chains - 1),
dtype='uint8')
fig, axes = plt.subplots(ncols=1, nrows=5)
for idx in range(n_samples):
# generate `plot_every` intermediate samples that we discard,
# because successive samples in the chain are too correlated
vis_mf, vis_sample = sample_fn()
print(' ... plotting sample %d' % idx)
Xvis = vis_mf.reshape( vis_mf.shape[0],5,100)
sub_pre[11 * idx:11 * idx + 10, :] = utils.tile_raster_images(
X=Xvis[:,0,:],
img_shape=(10, 10),
tile_shape=(1, n_chains),
tile_spacing=(1, 1)
)
sub_post1[11 * idx:11 * idx + 10, :] = utils.tile_raster_images(
X=Xvis[:,1,:],
img_shape=(10, 10),
tile_shape=(1, n_chains),
tile_spacing=(1, 1)
)
sub_post2[11 * idx:11 * idx + 10, :] = utils.tile_raster_images(
X=Xvis[:,2,:],
img_shape=(10, 10),
tile_shape=(1, n_chains),
tile_spacing=(1, 1)
)
sub_post3[11 * idx:11 * idx + 10, :] = utils.tile_raster_images(
X=Xvis[:,3,:],
img_shape=(10, 10),
tile_shape=(1, n_chains),
tile_spacing=(1, 1)
)
sub_post4[11 * idx:11 * idx + 10, :] = utils.tile_raster_images(
X=Xvis[:,4,:],
img_shape=(10, 10),
tile_shape=(1, n_chains),
tile_spacing=(1, 1)
)
# construct image
axes[0].imshow( Image.fromarray( sub_pre), cmap="Greys_r")
axes[1].imshow( Image.fromarray( sub_post1), cmap="Greys_r")
axes[2].imshow( Image.fromarray( sub_post2), cmap="Greys_r")
axes[3].imshow( Image.fromarray( sub_post3), cmap="Greys_r")
axes[4].imshow( Image.fromarray( sub_post4), cmap="Greys_r")
axes[0].get_xaxis().set_visible(False)
axes[0].get_yaxis().set_visible(False)
axes[1].get_xaxis().set_visible(False)
axes[1].get_yaxis().set_visible(False)
axes[2].get_xaxis().set_visible(False)
axes[2].get_yaxis().set_visible(False)
axes[3].get_xaxis().set_visible(False)
axes[3].get_yaxis().set_visible(False)
axes[4].get_xaxis().set_visible(False)
axes[4].get_yaxis().set_visible(False)
plt.show()
item='test'
#show and save
fig.savefig('samples_fromchain'+str(item)+'.pdf')
return
def main(args):
dset = load_data(args.dataset, args.test)
if not args.test:
val_images = test_images = dset.validation.images
else:
val_images = dset.test.images
def get_val_batch():
idcs = np.arange(val_images.shape[0])
np.random.shuffle(idcs)
return val_images[idcs[:args.batch_size * 2]]
compute_fid, _fid_lc = load_fid(dset.test.images, args)
use_restored_s1 = len(args.restore_s1_from) > 0
if use_restored_s1:
args_json = json.load(
open(os.path.join(args.restore_s1_from, 'hps.txt')))
assert args_json['dataset'] == args.dataset and \
args_json['sd_param'] == args.sd_param
keys_to_restore = ['sd_param', 'z_dims_s1', 'skip_last_fc']
vars(args).update(dict((k, args_json[k]) for k in keys_to_restore))
tf.set_random_seed(args.seed)
np.random.seed(args.seed)
G = Graph(args, val_images.shape[1], val_images.shape[3])
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
if use_restored_s1:
s1_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=FIRST_STAGE)
saver = tfv1.train.Saver(var_list=s1_vars)
ckpt_dir = tf.train.get_checkpoint_state(
args.restore_s1_from).model_checkpoint_path
print('RESTORING S1 FROM', ckpt_dir)
saver.restore(sess, ckpt_dir)
saver = tfv1.train.Saver(keep_checkpoint_every_n_hours=2, max_to_keep=1)
print(args)
B = 100
if args.decoder_warmup and not use_restored_s1:
print('Warmup stage:')
es = EarlyStopper(args.n_iter_s0 // (B * 4))
with LogContext(args.n_iter_s0 // B, logdir=args.dir,
tfsummary=True) as ctx:
for i in ctx:
for j in range(B): # training
x_mb, _ = dset.train.next_batch(args.batch_size)
G.train_step_0(sess, x_mb, args.lr * 2, i)
x_mb = G.process_batch(get_val_batch())
to_log = sess.run({**G.optimizer_s0.print}, {
G.x_ph: x_mb,
G.is_training_ph: False
})
ctx.log_scalars(to_log, list(to_log))
if es.add_check(to_log['__loss']):
print('Early stopping')
break
if not use_restored_s1:
print('First stage:')
es = EarlyStopper(args.n_iter_s1 // (B * 8))
with LogContext(args.n_iter_s1 // B, logdir=args.dir,
tfsummary=True) as ctx:
for i in ctx:
for j in range(B): # training
lr = compute_lr(args.lr, args.lr_halve_every_s1, i * B + j)
x_mb, _ = dset.train.next_batch(args.batch_size)
G.train_step_1(sess, x_mb, lr, i)
# plot validation
x_mb = G.process_batch(get_val_batch())
to_log = sess.run({**G.optimizer_s1.print}, {
G.x_ph: x_mb,
G.is_training_ph: False
})
# if i % 80 == 0:
# fids, _ = G.get_fids('s1', sess, val_images, compute_fid)
# to_log.update(fids)
ctx.log_scalars(to_log, list(to_log))
if args.s1_early_stop and es.add_check(to_log['__loss']):
print('Early stop')
break
e_start = args.n_iter_s1 // B
N = dset.train.images.shape[0]
B = (N + args.batch_size - 1) // args.batch_size
G.gen_qzx(sess, dset.train.images)
if args.production:
if use_restored_s1:
fids, _ = G.get_fids('s1', sess, val_images, compute_fid)
print('S1 FID:', fids)
else:
saver.save(sess,
os.path.join(args.dir, 'model'),
global_step=e_start)
print('Model saved')
if args.decoder_warmup_s2:
print('S2 warmup stage:')
es = EarlyStopper(args.n_iter_s0 // (B * 4))
with LogContext(args.n_iter_s0 // B, logdir=args.dir,
tfsummary=True) as ctx:
for i in ctx:
idcs = np.arange(N)
np.random.shuffle(idcs)
z_samples = np.random.normal(G.s1_mu_z, G.s1_sig_z)[idcs]
for j in range(B): # training
G.train_step_02(sess, z_samples[j * B:(j + 1) * B],
args.lr, i)
to_log = sess.run(G.optimizer_s02.print, {
G.model_s1.z: z_samples,
G.is_training_ph: False
})
ctx.log_scalars(to_log, list(to_log))
if es.add_check(to_log['__loss']):
print('Early stopping')
break
with LogContext(args.n_iter_s2 // B + 1,
logdir=args.dir,
tfsummary=True,
n_ep_start=e_start) as ctx:
for i in ctx:
idcs = np.arange(N)
np.random.shuffle(idcs)
z_samples = np.random.normal(G.s1_mu_z, G.s1_sig_z)[idcs]
for j in range(B): # training
lr = compute_lr(args.lr, args.lr_halve_every_s2, i * B + j)
z_mb = z_samples[j * B:(j + 1) * B]
G.train_step_2(sess, z_mb, lr, i)
# plot validation
x_mb = G.process_batch(get_val_batch())
to_log = sess.run(G.toprint_s2, {
G.x_ph: x_mb,
G.is_training_ph: False
})
if i % 80 == 0:
fids, _ = G.get_fids('s2', sess, val_images, compute_fid)
to_log.update(fids)
ctx.log_scalars(to_log, list(to_log))
if (i * 100) % args.save_every == 0 and args.save_every > 0:
saver.save(sess,
os.path.join(args.dir, 'model'),
global_step=i + e_start)
print('Model saved')
print('Test/validation:')
test_images = val_images
_, sample_images = G.get_fids('s2', sess, val_images, compute_fid)
im_tiled = tile_images(sample_images[:100].reshape((100, G.HW, G.HW, G.C)))
save_image(os.path.join(args.dir, 'sampled.png'), im_tiled)
score = compute_fid_inception(dset, sess, G)
print('FID SCORE (Bin Dai\'s) =', score)
def test_SdA_timep(self, pretraining_epochs, pretrain_lr, batch_size,
training_epochs, finetune_lr,
corruption_levels,
hidden_layers_sizes, hidden_layers_sidelen, output_folder):
"""
Demonstrates how to train and test a stochastic denoising autoencoder.
:type learning_rate: float
:param learning_rate: learning rate used in the finetune stage
(factor for the stochastic gradient)
:type pretraining_epochs: int
:param pretraining_epochs: number of epoch to do pretraining
:type pretrain_lr: float
:param pretrain_lr: learning rate to be used during pre-training
:type n_iter: int
:param n_iter: maximal number of iterations ot run the optimizer
:type dataset: string
:param dataset: path the the pickled dataset
"""
traindata_path= 'allLpatches_10x10x5.pklz' #'allLpatches_subs_smaller.pklz' #'allLpatches.pklz'
trainUdata_path= 'allUpatches_10x10x5.pklz'#'allUpatches_subs_smaller.pklz' #'allUpatches.pklz'
labeldata_path= 'allLabels_10x10x5.pklz' #'allLabels_subs_smaller.pklz' #'allLabels.pklz'
#############
## LOAD datasets
#############
datasets = self.load_wUdata(traindata_path, labeldata_path, trainUdata_path)
train_set_x, train_set_y = datasets[0]
np_train_x, np_train_y = datasets[3]
valid_set_x, valid_set_y = datasets[1]
np_valid_x, np_valid_y = datasets[4]
test_set_x, test_set_y = datasets[2]
np_test_x, np_test_y = datasets[5]
# inpect one image class 1
Vol = np_train_x[0].reshape(5,10,10)
imgslicestime = [Vol[0,:,:], Vol[1,:,:], Vol[2,:,:], Vol[3,:,:], Vol[4,:,:]]
subslicestime = []
# Display image
fig, ax = plt.subplots(nrows=4, ncols=4, figsize=(12, 4))
for k in range(1,5):
ax[k-1,0].imshow(imgslicestime[k], cmap=plt.cm.gray)
ax[k-1,0].set_axis_off()
ax[k-1,0].set_adjustable('box-forced')
# Display Original histogram
ax[k-1,1].hist(imgslicestime[k].ravel(), bins=50, color='black')
ax[k-1,1].ticklabel_format(axis='y', style='scientific', scilimits=(0, 0))
ax[k-1,1].set_xlabel('original')
# substract volume
subVol = np.asarray(imgslicestime[k]) - np.asarray(imgslicestime[0])
subslicestime.append(subVol)
# Display normalized histogram
ax[k-1,2].hist(subVol.ravel(), bins=50, color='black')
ax[k-1,2].ticklabel_format(axis='y', style='scientific', scilimits=(0, 0))
ax[k-1,2].set_xlabel('substracted histogram')
# display normalized 0-1
ax[k-1,3].imshow(subVol, cmap=plt.cm.gray)
ax[k-1,3].set_axis_off()
ax[k-1,3].set_adjustable('box-forced')
plt.close()
#################
# Substract pathces from pre-contrast
#################
subsnp_train_x = []
subsnp_valid_x = []
subsnp_test_x = []
for img in np_train_x:
Vol = img.reshape(5,10,10)
imgslicestime = [Vol[0,:,:], Vol[1,:,:], Vol[2,:,:], Vol[3,:,:], Vol[4,:,:]]
subslicestime = []
for k in range(1,5):
# substract volume
subVol = np.asarray(imgslicestime[k]) - np.asarray(imgslicestime[0])
subslicestime.append(subVol)
#append
subsnp_train_x.append( np.asarray(subslicestime).reshape(4*10*10) )
for img in np_valid_x:
Vol = img.reshape(5,10,10)
imgslicestime = [Vol[0,:,:], Vol[1,:,:], Vol[2,:,:], Vol[3,:,:], Vol[4,:,:]]
subslicestime = []
for k in range(1,5):
# substract volume
subVol = np.asarray(imgslicestime[k]) - np.asarray(imgslicestime[0])
subslicestime.append(subVol)
#append
subsnp_valid_x.append( np.asarray(subslicestime).reshape(4*10*10) )
for img in np_test_x:
Vol = img.reshape(5,10,10)
imgslicestime = [Vol[0,:,:], Vol[1,:,:], Vol[2,:,:], Vol[3,:,:], Vol[4,:,:]]
subslicestime = []
for k in range(1,5):
# substract volume
subVol = np.asarray(imgslicestime[k]) - np.asarray(imgslicestime[0])
subslicestime.append(subVol)
#append
subsnp_test_x.append( np.asarray(subslicestime).reshape(4*10*10) )
train_set_x, train_set_y = self.shared_dataset(subsnp_train_x, np_train_y)
valid_set_x, valid_set_y = self.shared_dataset(subsnp_valid_x, np_valid_y)
test_set_x, test_set_y = self.shared_dataset(subsnp_test_x, np_test_y)
subsdatasets = [(train_set_x, train_set_y),
(valid_set_x, valid_set_y),
(test_set_x, test_set_y) ]
#################
# To normalize find mean and std, then convert to float in [0-1] range
# ref on mininatch normalization: chrome-extension://oemmndcbldboiebfnladdacbdfmadadm/https://arxiv.org/pdf/1502.03167v3.pdf
# stackedOver: http://stats.stackexchange.com/questions/211436/why-do-we-normalize-images-by-subtracting-the-datasets-image-mean-and-not-the-c?rq=1
#################
# Xtrain = np.asarray(np_train_x)
# muXtrain = np.mean(Xtrain)
# varXtrain = np.var(Xtrain)
# print(muXtrain,varXtrain)
#
# Xvalid = np.asarray(np_valid_x)
# muXvalid = np.mean(Xvalid)
# varXvalid = np.var(Xvalid)
# print(muXvalid,varXvalid)
#
# Xtest = np.asarray(np_test_x)
# muXtest = np.mean(Xtest)
# varXtest = np.var(Xtest)
# print(muXtest,varXtest)
#
# fnp_train_x = []
# fnp_valid_x = []
# fnp_test_x = []
# for img in np_train_x:
# #img=np_train_x[1]
# fimg = (img - muXtrain)/np.sqrt(varXtrain+0.00001)
# rVol = exposure.rescale_intensity(fimg.reshape(4,10,10), out_range=(0, 1))
# img = sitk.GetImageFromArray(rVol)
# aimg = sitk.Cast(img,sitk.sitkFloat32)
# sitkVol = sitk.GetArrayFromImage(aimg)
# img = sitkVol.reshape(400)
# fnp_train_x.append(img)
# for img in np_valid_x:
# fimg = (img - muXvalid)/np.sqrt(varXvalid+0.00001)
# rVol = exposure.rescale_intensity(fimg.reshape(4,10,10), out_range=(0, 1))
# img = sitk.GetImageFromArray(rVol)
# aimg = sitk.Cast(img,sitk.sitkFloat32)
# sitkVol = sitk.GetArrayFromImage(aimg)
# img = sitkVol.reshape(400)
# fnp_valid_x.append(img)
# for img in np_test_x:
# fimg = (img - muXtest)/np.sqrt(varXtest+0.00001)
# rVol = exposure.rescale_intensity(fimg.reshape(4,10,10), out_range=(0, 1))
# img = sitk.GetImageFromArray(rVol)
# aimg = sitk.Cast(img,sitk.sitkFloat32)
# sitkVol = sitk.GetArrayFromImage(aimg)
# img = sitkVol.reshape(400)
# fnp_test_x.append(img)
#
#################
## plot ooptional
#################
# fig, axes = plt.subplots(ncols=4, nrows=2)
# axes[0,0].imshow(np_train_x[1].reshape(5,10,10)[0,:,:], cmap="Greys_r")
# axes[0,1].imshow(np_train_x[1].reshape(5,10,10)[1,:,:], cmap="Greys_r")
# axes[0,2].imshow(np_train_x[1].reshape(5,10,10)[2,:,:], cmap="Greys_r")
# axes[0,3].imshow(np_train_x[1].reshape(5,10,10)[3,:,:], cmap="Greys_r")
# axes[1,0].imshow( fnp_train_x[1].reshape(5,10,10)[0,:,:], cmap="Greys_r")
# axes[1,1].imshow( fnp_train_x[1].reshape(5,10,10)[1,:,:], cmap="Greys_r")
# axes[1,2].imshow( fnp_train_x[1].reshape(5,10,10)[2,:,:], cmap="Greys_r")
# axes[1,3].imshow( fnp_train_x[1].reshape(5,10,10)[3,:,:], cmap="Greys_r")
# train_set_x, train_set_y = self.shared_dataset(fnp_train_x, np_train_y)
# valid_set_x, valid_set_y = self.shared_dataset(fnp_valid_x, np_valid_y)
# test_set_x, test_set_y = self.shared_dataset(fnp_test_x, np_test_y)
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0]
n_train_batches //= batch_size
# numpy random generator
numpy_rng = np.random.RandomState(89677)
print('... building the Stacked Autoenconder model')
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
# construct the stacked denoising autoencoder class
sda = SdA(
numpy_rng=numpy_rng,
n_ins = 4*10*10, # before 30*30*4
hidden_layers_sizes=hidden_layers_sizes,
corruption_levels=corruption_levels,
n_outs=2
)
#########################
# PRETRAINING THE MODEL #
#########################
dA_avg_costs = []
dA_iter = []
layer_i = []
print('... getting the pretraining functions')
pretraining_fns = sda.pretraining_functions(train_set_x=train_set_x,
np_train_y=np_train_y,
batch_size=batch_size)
print('... pre-training the model')
start_time = timeit.default_timer()
## Pre-train layer-wise
for i in range(sda.n_layers):
# go through pretraining epochs
meanc = []
for epoch in range(pretraining_epochs):
# go through the training set
c = [];
if( epoch % 100 == 0):
pretrain_lr = pretrain_lr/10
for batch_index in range(n_train_batches):
c.append(pretraining_fns[i](index=batch_index,
corruption=corruption_levels[i],
lr=pretrain_lr))
# append
dA_avg_costs.append( np.mean(c) )
dA_iter.append(epoch)
layer_i.append(i)
print('Pre-training layer %i, epoch %d, cost %f' % (i, epoch, np.mean(c)))
# check that Cost does not improve more than 1% on the training set after an epoch
meanc.append( np.mean(c) )
if (epoch > 195):
decrCost = abs(meanc[epoch-1] - meanc[epoch])/meanc[epoch]*100
if(decrCost <= 0.01):
break
end_time = timeit.default_timer()
print(('The pretraining code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.)))
#####################################
# Plot images in 2D
#####################################
Xtmp = sda.dA_layers[0].W.get_value(borrow=True).T
imagefilters = []
for k in range(4):
imgX = Xtmp.reshape( Xtmp.shape[0], 4, 10, 10)[:,k,:,:]
image = tile_images(X=imgX , img_shape=(10, 10),
tile_shape=(10, 10),
tile_spacing=(1, 1))
# append to all
imagefilters.append( image )
##############
# Format
#################
LLdata = [float(L) for L in dA_avg_costs]
LLiter = [float(it) for it in dA_iter]
LLilayer = [ilayer for ilayer in layer_i]
dfpredata = pd.DataFrame( LLdata )
dfpredata.columns = ['LL_iter']
dfpredata['iter'] = LLiter
dfpredata['layer'] = LLilayer
########################
# FINETUNING THE MODEL #
########################
# for fine tunning make minibatch size = 1
# compute number of minibatches for training, validation and testing
#batch_size = 4
#n_train_batches = train_set_x.get_value(borrow=True).shape[0]//4
# get the training, validation and testing function for the model
print('... getting the finetuning functions')
train_fn, validate_model, test_model = sda.build_finetune_functions(
datasets=subsdatasets,
batch_size=batch_size,
learning_rate=finetune_lr
)
print('... finetunning the Stacked model')
############
### for plotting likelihood or cost, accumulate returns of train_model
############
minibatch_avg_costs = []
minibatch_iter = []
minibatch_loss = []
# early-stopping parameters
patience = 1000 * n_train_batches # look as this many examples regardless
patience_increase = 2. # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience // 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_validation_loss = np.inf
test_score = 0.
start_time = timeit.default_timer()
done_looping = False
epoch = 0
while (epoch < training_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in range(n_train_batches):
minibatch_avg_cost = train_fn(minibatch_index)
iter = (epoch - 1) * n_train_batches + minibatch_index
if (iter + 1) % validation_frequency == 0:
validation_losses = validate_model()
this_validation_loss = np.mean(validation_losses)
print('epoch %i, minibatch %i/%i, validation error %f %%' %
(epoch, minibatch_index + 1,
n_train_batches,
this_validation_loss * 100.))
##############
# append
#################
minibatch_avg_costs.append(minibatch_avg_cost)
minibatch_iter.append(iter)
minibatch_loss.append(this_validation_loss*100)
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
#improve patience if loss improvement is good enough
if (
this_validation_loss < best_validation_loss *
improvement_threshold
):
patience = max(patience, iter * patience_increase)
# save best validation score and iteration number
best_validation_loss = this_validation_loss
best_iter = iter
# test it on the test set
test_losses = test_model()
test_score = np.mean(test_losses)
print((' epoch %i, minibatch %i/%i, test error of '
'best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if patience <= iter:
done_looping = True
break
end_time = timeit.default_timer()
print 'Optimization complete with best validation score of %f, on iteration %i, with test performance %f' % ((best_validation_loss * 100.), (best_iter + 1), (test_score * 100.))
print(('The training code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.)))
##############
# Format
#################
LLdata = [float(L) for L in minibatch_avg_costs]
LLiter = [float(it) for it in minibatch_iter]
LLoss = [float(l) for l in minibatch_loss]
dfinedata = pd.DataFrame( LLdata )
dfinedata.columns = ['LL_iter']
dfinedata['iter'] = LLiter
dfinedata['loss'] = LLoss
###############
## Predictions
###############
# get training data in numpy format
X,y = subsnp_train_x, np_train_y #datasets[3]
Xtrain = np.asarray(X)
ytrain = utils.makeMultiClass(y)
# get valid data in numpy format
X,y = subsnp_valid_x, np_valid_y # #datasets[4]
Xvalid = np.asarray(X)
yvalid = utils.makeMultiClass(y)
# get test data in numpy format
X,y = subsnp_test_x, np_test_y #datasets[5]
Xtest = np.asarray(X)
ytest = utils.makeMultiClass(y)
###############
# predicting using the SDA
###############
# in train
predtrain = sda.predict_functions(Xtrain).argmax(1)
# let's see how the network did
y = ytrain.argmax(1)
e0 = 0.0; y0 = len([0 for yi in range(len(y)) if y[yi]==0])
e1 = 0.0; y1 = len([1 for yi in range(len(y)) if y[yi]==1])
for i in range(len(y)):
if(y[i] == 1):
e1 += y[i]==predtrain[i]
if(y[i] == 0):
e0 += y[i]==predtrain[i]
# printing the result, this structure should result in 80% accuracy
Acutrain0=100*e0/y0
Acutrain1=100*e1/y1
print "Train Accuracy for class 0: %2.2f%%"%(Acutrain0)
print "Train Accuracy for class 1: %2.2f%%"%(Acutrain1)
# in Valid
predvalid = sda.predict_functions(Xvalid).argmax(1)
# let's see how the network did
y = yvalid.argmax(1)
e0 = 0.0; y0 = len([0 for yi in range(len(y)) if y[yi]==0])
e1 = 0.0; y1 = len([1 for yi in range(len(y)) if y[yi]==1])
for i in range(len(y)):
if(y[i] == 1):
e1 += y[i]==predvalid[i]
if(y[i] == 0):
e0 += y[i]==predvalid[i]
# printing the result, this structure should result in 80% accuracy
Acuvalid0=100*e0/y0
Acuvalid1=100*e1/y1
print "Valid Accuracy for class 0: %2.2f%%"%(Acuvalid0)
print "Valid Accuracy for class 1: %2.2f%%"%(Acuvalid1)
# in Xtest
predtest = sda.predict_functions(Xtest).argmax(1)
# let's see how the network did
y = ytest.argmax(1)
e0 = 0.0; y0 = len([0 for yi in range(len(y)) if y[yi]==0])
e1 = 0.0; y1 = len([1 for yi in range(len(y)) if y[yi]==1])
for i in range(len(y)):
if(y[i] == 1):
e1 += y[i]==predtest[i]
if(y[i] == 0):
e0 += y[i]==predtest[i]
# printing the result, this structure should result in 80% accuracy
Acutest0=100*e0/y0
Acutest1=100*e1/y1
print "Test Accuracy for class 0: %2.2f%%"%(Acutest0)
print "Test Accuracy for class 1: %2.2f%%"%(Acutest1)
return [dfpredata, dfinedata, imagefilters, sda,
Acutrain0, Acutrain1,
Acuvalid0, Acuvalid1,
Acutest0, Acutest1]
def fit(
train_loader: torch.utils.data.DataLoader,
val_loader: torch.utils.data.DataLoader,
model: torch.nn.Module,
exp_path: str,
label_preprocess: Callable,
loss_fn: torch.autograd.Function,
label2onehot: Callable,
n_classes: int = 10,
optimizer: str = "adam",
learning_rate: float = 1e-4,
cuda: bool = True,
patience: int = 10,
max_epochs: int = 200,
resume: bool = False,
):
def _save_img(x: np.array, filename: str):
"save numpy array representaion of image to image file"
Image.fromarray((255 * x).astype("uint8")).save(filename)
def _run_epoch(dataloader: torch.utils.data.DataLoader, training: bool):
p_bar = ProgressBar()
losses = []
mean_outs = []
# import pdb; pdb.set_trace()
# batch_images = [bs, ch, w, h], batch_targets = [bs, n_classes] OHE
for batch_images, batch_targets in p_bar(dataloader):
batch_labels = label_preprocess(batch_images)
if cuda:
batch_images, batch_targets = batch_images.cuda(), batch_targets.cuda()
batch_labels = batch_labels.cuda()
if training:
optimizer.zero_grad()
model.train()
else:
model.eval()
predicted_labels = model(batch_images, batch_targets)
# cross entropy between image's pixel value vs predicted
loss = loss_fn(predicted_labels, batch_labels)
# track mean output
predicted_labels = predicted_labels.data.cpu().numpy()
# TODO: why do we need to keep track of mean?
mean_outs.append(
np.mean(np.argmax(predicted_labels, axis=1)) / predicted_labels.shape[1]
)
if training:
loss.backward()
optimizer.step()
losses.append(loss.data.cpu().numpy())
utils.clearline()
return float(np.mean(losses)), np.mean(mean_outs)
if cuda:
model = model.cuda()
exp_path = Path(exp_path)
if not os.path.isdir(exp_path):
os.makedirs(exp_path)
statsfile = exp_path / "stats.json"
optimizer = {
"adam": torch.optim.Adam(model.parameters(), lr=learning_rate),
"sgd": torch.optim.SGD(model.parameters(), lr=learning_rate, momentum=0.9),
"adamax": torch.optim.Adamax(model.parameters(), lr=learning_rate),
}[optimizer.lower()]
# load a single example from the iterator to get the image size
x = train_loader.sampler.data_source.__getitem__(0)[0]
img_size = list(x.numpy().shape[1:])
if not resume:
stats = {
"loss": {"train": [], "val": []},
"mean_output": {"train": [], "val": []},
}
best_val = np.inf
stall = 0
start_epoch = 0
generated = []
plots = []
else:
with open(statsfile, "r") as js:
stats = json.load(js)
best_val = np.min(stats["loss"]["val"])
stall = len(stats["loss"]["val"]) - np.argmin(stats["loss"]["val"]) - 1
start_epoch = len(stats["loss"]["val"]) - 1
generated = list(np.load(exp_path / "generated.npy"))
plots = list(np.load(exp_path / "generated_plots.npy"))
print(f"Resuming from epoch {start_epoch}")
for epoch in range(start_epoch, max_epochs):
# Training
t0 = time.time()
loss, mean_out = _run_epoch(train_loader, training=True)
time_per_example = (time.time() - t0) / len(train_loader.dataset)
stats["loss"]["train"].append(loss)
stats["mean_output"]["train"].append(mean_out)
print(
(
f"Epoch {epoch}: Training loss = {loss:.4f} mean output = {mean_out:.2f} "
f"{time_per_example*1000:.2f} msec/example"
)
)
# Validation
t0 = time.time()
loss, mean_out = _run_epoch(val_loader, training=False)
time_per_example = (time.time() - t0) / len(val_loader.dataset)
stats["loss"]["val"].append(loss)
stats["mean_output"]["val"].append(mean_out)
print(
(
f" Validation loss = {loss:.4f} mean output = {mean_out:.2f} "
f"{time_per_example*1000:.2f} msec/example"
)
)
# Generate images and update gif
new_frame = utils.tile_images(
generate_images(model, img_size, n_classes, label2onehot, cuda)
)
generated.append(new_frame)
# Update gif with loss plot
plot_frame = plot_loss(stats["loss"]["train"], stats["loss"]["val"])
if new_frame.ndim == 2:
new_frame = np.repeat(new_frame[:, :, np.newaxis], 3, axis=2)
nw = int(new_frame.shape[1] * plot_frame.shape[0] / new_frame.shape[0])
new_frame = resize(
new_frame,
[plot_frame.shape[0], nw],
order=0,
preserve_range=True,
mode="constant",
)
plots.append(
np.concatenate(
(plot_frame.astype("uint8"), new_frame.astype("uint8")), axis=1
)
)
# Save gif arrays so it can resume training if interrupted
np.save(exp_path / "generated.npy", generated)
np.save(exp_path / "generated_plots.npy", plots)
# Save stats and update training curves
with open(statsfile, "w") as sf:
json.dump(stats, sf)
utils.plot_stats(stats, exp_path)
# Early stopping
torch.save(model, exp_path / "last_checkpoint.pth")
if loss < best_val:
best_val = loss
stall = 0
torch.save(model, exp_path / "best_checkpoint.pth")
imageio.imsave(
exp_path / "best_generated.jpeg", generated[-1].astype("uint8")
)
imageio.imsave(
exp_path / "best_generated_plots.jpeg", plots[-1].astype("uint8")
)
imageio.mimsave(
exp_path / "generated.gif",
np.array(generated),
format="gif",
loop=0,
fps=2,
)
imageio.mimsave(
exp_path / "generated_plot.gif",
np.array(plots),
format="gif",
loop=0,
fps=2,
)
else:
stall += 1
if stall >= patience:
break
def main(args):
from tensorflow.examples.tutorials.mnist import input_data
if not args.test:
mnist = input_data.read_data_sets(
'data/',
one_hot=False,
source_url='http://yann.lecun.com/exdb/mnist/')
val_images = test_images = mnist.validation.images
else:
mnist = input_data.read_data_sets(
'data/',
one_hot=False,
source_url='http://yann.lecun.com/exdb/mnist/',
validation_size=0)
val_images = mnist.test.images
np.random.seed(args.seed)
tf.set_random_seed(args.seed)
if args.do_fid:
compute_fid, _fid_lc = load_fid(
mnist.test.images, args, binarize=(args.observation != 'normal'))
# digit placeholder
x_ph = tf.placeholder(tf.float32, shape=(None, 784))
is_training_ph = tf.placeholder(tf.bool, shape=(), name='is_training')
z_dims = args.z_dims if args.latent == 'euc' else args.z_dims + 1
activation = {'elu': tf.nn.elu, 'tanh': tf.nn.tanh}[args.ae_activation]
if args.implicit:
x1 = tf.tile(x_ph, [args.n_particles, 1])
model_cls = models.ImplicitAE
model = model_cls(x=x1,
h_dim=args.h_dims,
z_dim=z_dims,
eps_dim=args.eps_dims,
latent_space=args.latent,
ae_activation=args.ae_activation,
energy_activation=args.energy_activation,
rescale_sph_latent=args.rescale_sph_latent,
observation=args.observation,
gan=(args.model == 'wae-gan'),
is_training_ph=is_training_ph)
optimizer = {
'vae': optimizers.ImplicitVAE,
'wae': optimizers.ImplicitWAE,
'wae-gan': optimizers.GanWAE,
'wae-mmd': optimizers.MMDWAE,
}[args.model](model, args)
ais_method = args.ais_method if args.latent == 'euc' else 'riem_ld'
log_lh_sym = log_lh_ais(model, optimizer, ais_method)
else:
assert not args.model.startswith('wae')
dist = {'euc': 'normal', 'sph': 'vmf'}[args.latent]
model = models.ExplicitAE(x=x_ph,
h_dim=args.h_dims,
z_dim=z_dims,
distribution=dist,
rescale_sph_latent=args.rescale_sph_latent)
optimizer = optimizers.ExplicitVAE(model, args)
log_lh_lb_sym = log_likelihood(model, optimizer, n=1000)
ais_method = args.ais_method if args.latent == 'euc' else 'riem_ld'
log_lh_sym = log_lh_ais(model, optimizer, ais_method)
# === FOR VIS ===
if args.latent == 'euc':
dist_pz = tf.distributions.Normal(tf.zeros_like(model.z),
tf.ones_like(model.z))
pz_sample = dist_pz.sample()
elif args.latent == 'sph':
dist_pz = HypersphericalUniform(model.z_dim - 1, dtype=model.x.dtype)
pz_sample = dist_pz.sample([tf.shape(model.z)[0]])
x_sample_sym = tf.nn.sigmoid(model._decoder(pz_sample))
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
saver = tfv1.train.Saver(keep_checkpoint_every_n_hours=2, max_to_keep=2)
print(args)
with LogContext(args.n_iter // 100, logdir=args.dir,
tfsummary=True) as ctx:
for i in ctx:
for j in range(100):
# training
x_mb, _ = mnist.train.next_batch(args.batch_size)
if args.observation != 'normal':
x_mb = (x_mb > np.random.random(size=x_mb.shape)).astype(
np.float32)
optimizer.step(sess, {x_ph: x_mb, is_training_ph: True})
# plot validation
x_mb = val_images
if args.observation != 'normal':
x_mb = (x_mb > np.random.random(size=x_mb.shape)).astype(
np.float32)
to_log = sess.run({**optimizer.print}, {
x_ph: x_mb,
is_training_ph: False
})
ctx.log_scalars(to_log, list(to_log))
if i % 80 == 0 and i > 0:
if not args.implicit:
idcs = np.arange(x_mb.shape[0])
np.random.shuffle(idcs)
idcs = idcs[:48]
loglh_val = sess.run(log_lh_sym, {
model.x: x_mb[idcs],
is_training_ph: False
})
to_log.update({
'log_lh_mean':
np.mean(loglh_val),
'log_lh_sd':
np.std(loglh_val) / (idcs.shape[0]**0.5)
})
if args.do_fid:
sample_images = sess.run(
x_sample_sym, {
model.x: np.zeros_like(
val_images[:10000]).astype('f'),
is_training_ph: False
})
fid_score = compute_fid(sample_images)
to_log['fid'] = fid_score
print(i, to_log)
if (i * 100) % args.save_every == 0 and args.save_every > 0:
saver.save(sess,
os.path.join(args.dir, 'model'),
global_step=i)
print('Model saved')
print('Test/validation:')
test_images = val_images
sample_images = sess.run(
x_sample_sym, {
model.x: np.zeros_like(val_images[:10000]).astype('f'),
is_training_ph: False
})
im_tiled = tile_images(sample_images[:100].reshape((100, 28, 28)))
save_image(os.path.join(args.dir, 'sampled.png'), im_tiled)
if args.do_fid:
fid_score = compute_fid(sample_images)
print('FID SCORE =', fid_score)
return
if not args.implicit:
x_mb = val_images
if args.observation != 'normal':
# dynamic binarization
x_mb = (x_mb > np.random.random(size=x_mb.shape)).astype(
np.float32)
print_ = {**optimizer.print}
print_['LL'] = log_lh_lb_sym
print(sess.run(print_, {model.x: x_mb, is_training_ph: False}))
lsum = 0
lden = 0
idc = np.arange(test_images.shape[0])
np.random.shuffle(idc)
test_images = test_images[idc]
with LogContext(test_images.shape[0] // args.batch_size) as ctx:
for j in ctx:
ti = test_images[j * args.batch_size:(j + 1) * args.batch_size]
if args.observation != 'normal':
ti = (ti > np.random.random(size=ti.shape)).astype(np.float32)
lhs = sess.run(log_lh_sym, {model.x: ti, is_training_ph: False})
lsum += lhs.mean() * ti.shape[0]
lden += ti.shape[0]
if j % 10 == 0:
ctx.log_scalars({'avg': lsum / lden}, ['avg'])
print('AIS lb = {}'.format(lsum / lden))
from utils import tile_images, save_image
import os, sys
import numpy as np
from matplotlib import pyplot as plt
dir_ = sys.argv[1]
ims = []
for j in range(100):
im = plt.imread(os.path.join(dir_, '{}.png'.format(j)))
ims.append(im)
ims = np.array(ims)
save_image(os.path.join(dir_, 'all.png'), tile_images(ims))