def get_images_and_latents(dir, dataset, batch):
ae, ds = utils.load_ae(dir,
dataset,
batch,
all_aes.ALL_AES,
return_dataset=True)
with utils.HookReport.disable():
ae.eval_mode()
if dataset == 'celeba32':
images, latents, labels, shape = get_latents_and_labels(ae.eval_sess,
ae.eval_ops,
ds.test,
batches=batch)
elif dataset == 'lines2':
images, latents, labels, shape = get_latents_and_labels(ae.eval_sess,
ae.eval_ops,
ds.test,
batches=batch)
else:
images, latents, labels, shape = get_latents_and_labels(ae.eval_sess,
ae.eval_ops,
ds.train_once,
batches=None)
return images, latents, labels, shape
def main(argv):
del argv # Unused.
batch = FLAGS.batch
dataset = data.get_dataset(FLAGS.dataset, dict(batch_size=batch))
ae, ae_model = utils.load_ae(FLAGS.ae_dir, FLAGS.dataset, FLAGS.batch,
all_aes.ALL_AES)
with utils.HookReport.disable():
ae.eval_mode()
model = XLatentAE(dataset, FLAGS.train_dir)
model.train_dir = os.path.join(model.train_dir, ae_model)
model.ae = ae
model.train()
def main(argv):
del argv # Unused.
ae, ds = utils.load_ae(FLAGS.ae_dir, FLAGS.dataset, FLAGS.batch,
all_aes.ALL_AES, return_dataset=True)
with utils.HookReport.disable():
ae.eval_mode()
# Convert all test samples to latents and get the labels
test_latents, test_labels = get_latents_and_labels(ae.eval_sess,
ae.eval_ops,
ds.test)
print('Shape of test_labels = {}'.format(np.shape(test_labels)))
print('Shape of test_latents = {}'.format(np.shape(test_latents)))
train_latents, train_labels = get_latents_and_labels(ae.eval_sess,
ae.eval_ops,
ds.train_once)
print('Shape of train_labels = {}'.format(np.shape(train_labels)))
print('Shape of train_latents = {}'.format(np.shape(train_latents)))
with open('mnist256_features.pkl', 'wb') as file:
pkl.dump({'test_labels': test_labels, 'test_latents': test_latents,
'train_labels': train_labels, 'train_latents': train_latents}, file)
if not FLAGS.use_svd:
acc = cluster(train_latents, train_labels, test_latents, test_labels)
print('classification acc = {}'.format(acc))
return
if 0: # use PCA
print('PCA')
pca = sklearn.decomposition.PCA(train_latents.shape[1], whiten=True)
pca.fit(train_latents)
train_latents = pca.transform(train_latents)
test_latents = pca.transform(test_latents)
else:
print('SVD')
mean = train_latents.mean(axis=0)
train_latents -= mean
test_latents -= mean
s, vt = la.svd(train_latents, full_matrices=False)[-2:]
train_latents = (train_latents.dot(vt.T) / (s + 1e-5))
test_latents = (test_latents.dot(vt.T) / (s + 1e-5))
rank = train_latents.shape[1]
for x in range(FLAGS.n_try):
acc = cluster(train_latents[:, :rank].copy(),
train_labels[:, :rank].copy(),
test_latents[:, :rank].copy(),
test_labels[:, :rank].copy())
print('Rank %3d Inits %4d Accuracy %.2f' % (rank, FLAGS.n_init, 100 * acc))
def main(argv):
del argv # Unused.
batch = FLAGS.batch
dataset = data.get_dataset(FLAGS.dataset, dict(batch_size=batch))
ae, ae_model = utils.load_ae(FLAGS.ae_dir, FLAGS.dataset, FLAGS.batch,
all_aes.ALL_AES)
with utils.HookReport.disable():
ae.eval_mode()
model = XLatentAE(
dataset,
FLAGS.train_dir)
model.train_dir = os.path.join(model.train_dir, ae_model)
model.ae = ae
model.train()
def main(argv):
del argv # Unused.
ae, ds = utils.load_ae(FLAGS.ae_dir, FLAGS.dataset, FLAGS.batch,
all_aes.ALL_AES, return_dataset=True)
with utils.HookReport.disable():
ae.eval_mode()
# Convert all test samples to latents and get the labels
test_latents, test_labels = get_latents_and_labels(ae.eval_sess,
ae.eval_ops,
ds.test)
print('Shape of test_labels = {}'.format(np.shape(test_labels)))
print('Shape of test_latents = {}'.format(np.shape(test_latents)))
train_latents, train_labels = get_latents_and_labels(ae.eval_sess,
ae.eval_ops,
ds.train_once,
60000)
print('Shape of train_labels = {}'.format(np.shape(train_labels)))
print('Shape of train_latents = {}'.format(np.shape(train_latents)))
if not FLAGS.use_svd:
acc = cluster(train_latents, train_labels, test_latents, test_labels)
print('classification acc = {}'.format(acc))
return
if 0: # use PCA
print('PCA')
pca = sklearn.decomposition.PCA(train_latents.shape[1], whiten=True)
pca.fit(train_latents)
train_latents = pca.transform(train_latents)
test_latents = pca.transform(test_latents)
else:
print('SVD')
mean = train_latents.mean(axis=0)
train_latents -= mean
test_latents -= mean
s, vt = la.svd(train_latents, full_matrices=False)[-2:]
print('SVD Sigma', s)
train_latents = (train_latents.dot(vt.T) / (s + 1e-5))
test_latents = (test_latents.dot(vt.T) / (s + 1e-5))
rank = train_latents.shape[1]
for x in range(FLAGS.n_try):
acc = cluster(train_latents[:, :rank].copy(),
train_labels[:, :rank].copy(),
test_latents[:, :rank].copy(),
test_labels[:, :rank].copy())
print('Rank %3d Inits %4d Accuracy %.2f' % (rank, FLAGS.n_init, 100 * acc))
def main(argv):
del argv # Unused.
ae, ds = utils.load_ae(FLAGS.ae_dir,
FLAGS.dataset,
FLAGS.batch,
all_aes.ALL_AES,
return_dataset=True)
with utils.HookReport.disable():
ae.eval_mode()
# Convert all test samples to latents and get the labels
val_images, val_latents, val_labels = get_images_and_latents_and_labels(
ae.eval_sess, ae.eval_ops, ds.val)
print('Shape of val_labels = {}'.format(np.shape(val_labels)))
print('Shape of val_latents = {}'.format(np.shape(val_latents)))
test_images, test_latents, test_labels = get_images_and_latents_and_labels(
ae.eval_sess, ae.eval_ops, ds.test)
print('Shape of test_labels = {}'.format(np.shape(test_labels)))
print('Shape of test_latents = {}'.format(np.shape(test_latents)))
train_images, train_latents, train_labels = get_images_and_latents_and_labels(
ae.eval_sess, ae.eval_ops, ds.train_once)
print('Shape of train_labels = {}'.format(np.shape(train_labels)))
print('Shape of train_latents = {}'.format(np.shape(train_latents)))
train_val_test_images = [train_images, val_images, test_images]
train_val_test_labels = [train_labels, val_labels, test_labels]
train_val_test_latents = [train_latents, val_latents, test_latents]
# train_val_test_images = map(lambda x: unshift_and_unscale_all(x), train_val_test_images)
# if FLAGS.dataset == 'mnist32' or FLAGS.dataset == 'omniglot32':
# train_val_test_images = map(lambda x: unpad(x), train_val_test_images)
# if FLAGS.dataset == 'mnist32':
# train_val_test_labels = map(lambda y: np.argmax(y, axis=-1), train_val_test_labels)
ipdb.set_trace()
for (split, X, Y, Z) in zip(['train', 'val', 'test'],
train_val_test_images, train_val_test_labels,
train_val_test_latents):
save_split(split, X, Y, Z, ae)