def test_predict(self):
unlabeled_dataset = SparseDataset("One pic", 'UNLABELEDBATCH', 1, 1)
network = create_dC2()
num_of_inputs = 5
nClasses = 40
renderSize = 40
test_file = ('SparseConvNet/Data/ModelNet/night_stand/'
'train/night_stand_0180.off')
for i in range(num_of_inputs):
unlabeled_dataset.add_picture(Off3DPicture(test_file, renderSize))
matrix_of_preds = network.predict(unlabeled_dataset)
self.assertEqual(matrix_of_preds.shape, (num_of_inputs, nClasses))
def test_predict(self):
unlabeled_dataset = SparseDataset("One pic", 'UNLABELEDBATCH', 1, 1)
network = create_dC2()
num_of_inputs = 5
nClasses = 40
renderSize = 40
test_file = ('SparseConvNet/Data/ModelNet/night_stand/'
'train/night_stand_0180.off')
for i in range(num_of_inputs):
unlabeled_dataset.add_picture(Off3DPicture(test_file, renderSize))
matrix_of_preds = network.predict(unlabeled_dataset)
self.assertEqual(matrix_of_preds.shape, (num_of_inputs, nClasses))
def generate_modelnet_dataset(full=False, limit=-1):
number_of_features = 1
renderSize = 40
if full:
data_folder = "SparseConvNet/Data/ModelNet/"
else:
data_folder = "SparseConvNet/Data/_ModelNet/"
class_folders = os.listdir(data_folder)
class_folders.sort()
number_of_classes = len(class_folders)
sparse_dataset = SparseDataset("ModelNet (Train subset)", 'TRAINBATCH',
number_of_features, number_of_classes)
for class_id, folder in enumerate(class_folders):
dirpath = os.path.join(data_folder, folder, 'train')
for _count, filename in enumerate(os.listdir(dirpath)):
if _count > limit > 0:
break
sparse_dataset.add_picture(Off3DPicture(
os.path.join(dirpath, filename), renderSize, label=class_id))
# sparse_dataset.repeatSamples(10)
return sparse_dataset
def generate_modelnet_dataset(full=False, limit=-1):
number_of_features = 1
renderSize = 40
if full:
data_folder = "SparseConvNet/Data/ModelNet/"
else:
data_folder = "SparseConvNet/Data/_ModelNet/"
class_folders = os.listdir(data_folder)
class_folders.sort()
number_of_classes = len(class_folders)
sparse_dataset = SparseDataset("ModelNet (Train subset)", 'TRAINBATCH',
number_of_features, number_of_classes)
for class_id, folder in enumerate(class_folders):
dirpath = os.path.join(data_folder, folder, 'train')
for _count, filename in enumerate(os.listdir(dirpath)):
if _count > limit > 0:
break
sparse_dataset.add_picture(
Off3DPicture(os.path.join(dirpath, filename),
renderSize,
label=class_id))
# sparse_dataset.repeatSamples(10)
return sparse_dataset
def test_dataset_creation(self):
with self.assertRaises(ValueError):
SparseDataset("Testing DataSet", "FakeBatch", 1, 2)
print("Successfully caught Value Exception.")
ds = SparseDataset("Testing DataSet", "TRAINBATCH", 1, 2)
self.assertEqual(ds.name, "Testing DataSet")
def train(ds,
network,
experiment_hash,
batch_size=150,
test_every_n_batches=100,
unique_classes_in_batch=5,
lr_policy=nop,
momentum_policy=nop,
pair_taking_method=0,
render_size=40,
weights_dir='./weights',
in_batch_sample_selection=False,
norm_type='cosine',
L=1,
epoch=0,
epoch_limit=None,
batch_iteration_hook=nop,
epoch_iteration_hook=nop,
pairs_limit=None):
linear_triplet_loss, ltl_grad, norm = get_functions(norm_type=norm_type,
margin=L)
ds.summary()
gen = ds.generate_triplets(batch_size=batch_size,
unique_classes_in_batch=unique_classes_in_batch,
method=pair_taking_method,
limit=pairs_limit)
weights_temp = os.path.join(weights_dir, experiment_hash)
print('Taking {} batches in to dataset'.format(test_every_n_batches))
if epoch_limit is None:
if pairs_limit is None:
total_pairs_num = sum(
map(lambda l: len(l) * (len(l) - 1), ds.classes))
else:
total_pairs_num = ds.get_limit(pairs_limit) * ds.class_count
total_number_of_epochs = int(
np.ceil(total_pairs_num / (batch_size / 3.0) /
test_every_n_batches))
else:
total_number_of_epochs = epoch_limit
for _ in tqdm(xrange(total_number_of_epochs),
total=total_number_of_epochs,
unit="epoch"):
train_ds = SparseDataset(ds.name + " train",
'TRAINBATCH',
1,
ds.class_count,
shuffle=False)
ranges_for_all = []
for batch_samples, ranges in tqdm(islice(gen, test_every_n_batches),
leave=False,
desc="Creating dataset"):
ranges_for_all.append(ranges)
for _sample in batch_samples:
train_ds.add_picture(Off3DPicture(_sample, render_size))
if not ranges_for_all:
break
batch_gen = network.batch_generator(train_ds, batch_size)
learning_rate = lr_policy(epoch)
momentum = momentum_policy(epoch)
for bid, (batch,
_ranges) in tqdm(enumerate(izip(batch_gen, ranges_for_all)),
leave=False,
unit='batch',
total=test_every_n_batches):
activation = network.processBatchForward(batch)
feature_vectors = activation['features']
delta = np.zeros_like(feature_vectors)
batch_loss = []
for _offset, _range in zip(3 * np.cumsum([0] + _ranges)[:-1],
_ranges):
if in_batch_sample_selection:
one_class_ids = np.arange(2 * _range) + _offset
other_class_ids = np.arange(2 * _range,
3 * _range) + _offset
while one_class_ids.any():
anchor = one_class_ids[0]
positive_id = weighted_sampling_of_best(
np.apply_along_axis(
partial(norm, feature_vectors[anchor]), 1,
feature_vectors[one_class_ids[1:]]),
best='max')
positive_id += 1
negative_id = weighted_sampling_of_best(
np.apply_along_axis(
partial(norm, feature_vectors[anchor]), 1,
feature_vectors[other_class_ids]),
best='min')
triplet_slice = [
anchor, one_class_ids[positive_id],
other_class_ids[negative_id]
]
one_class_ids = np.delete(one_class_ids,
[0, positive_id])
other_class_ids = np.delete(other_class_ids,
negative_id)
delta[triplet_slice] = ltl_grad(
feature_vectors[triplet_slice])
batch_loss.append(
linear_triplet_loss(
feature_vectors[triplet_slice]))
else:
for _i in range(_range):
triplet_slice = _offset + (np.arange(3) * _range) + _i
delta[triplet_slice] = ltl_grad(
feature_vectors[triplet_slice])
batch_loss.append(
linear_triplet_loss(
feature_vectors[triplet_slice]))
batch_iteration_hook(batch_loss=batch_loss, epoch=epoch, bid=bid)
network.processBatchBackward(batch,
delta,
learningRate=learning_rate,
momentum=momentum)
network.saveWeights(weights_temp, epoch)
epoch_iteration_hook(_network=network,
learning_rate=learning_rate,
momentum=momentum,
epoch=epoch,
weights_path="{}_epoch-{}.cnn".format(
weights_temp, epoch))
epoch += 1
del train_ds
def train(ds, network, experiment_hash,
batch_size=150, test_every_n_batches=100,
unique_classes_in_batch=5,
lr_policy=nop,
momentum_policy=nop,
pair_taking_method=0,
render_size=40, weights_dir='./weights',
in_batch_sample_selection=False, norm_type='cosine', L=1,
epoch=0, epoch_limit=None,
batch_iteration_hook=nop, epoch_iteration_hook=nop,
pairs_limit=None):
linear_triplet_loss, ltl_grad, norm = get_functions(norm_type=norm_type,
margin=L)
ds.summary()
gen = ds.generate_triplets(batch_size=batch_size,
unique_classes_in_batch=unique_classes_in_batch,
method=pair_taking_method, limit=pairs_limit)
weights_temp = os.path.join(weights_dir, experiment_hash)
print('Taking {} batches in to dataset'.format(test_every_n_batches))
if epoch_limit is None:
if pairs_limit is None:
total_pairs_num = sum(map(lambda l: len(l) * (len(l) - 1),
ds.classes))
else:
total_pairs_num = ds.get_limit(pairs_limit) * ds.class_count
total_number_of_epochs = int(np.ceil(
total_pairs_num
/ (batch_size / 3.0) / test_every_n_batches))
else:
total_number_of_epochs = epoch_limit
for _ in tqdm(xrange(total_number_of_epochs),
total=total_number_of_epochs, unit="epoch"):
train_ds = SparseDataset(
ds.name + " train", 'TRAINBATCH', 1, ds.class_count, shuffle=False)
ranges_for_all = []
for batch_samples, ranges in tqdm(islice(gen, test_every_n_batches),
leave=False,
desc="Creating dataset"):
ranges_for_all.append(ranges)
for _sample in batch_samples:
train_ds.add_picture(Off3DPicture(_sample, render_size))
if not ranges_for_all:
break
batch_gen = network.batch_generator(train_ds, batch_size)
learning_rate = lr_policy(epoch)
momentum = momentum_policy(epoch)
for bid, (batch, _ranges) in tqdm(
enumerate(izip(batch_gen, ranges_for_all)),
leave=False, unit='batch', total=test_every_n_batches):
activation = network.processBatchForward(batch)
feature_vectors = activation['features']
delta = np.zeros_like(feature_vectors)
batch_loss = []
for _offset, _range in zip(3 * np.cumsum([0] + _ranges)[:-1], _ranges):
if in_batch_sample_selection:
one_class_ids = np.arange(2 * _range) + _offset
other_class_ids = np.arange(2 * _range, 3 * _range) + _offset
while one_class_ids.any():
anchor = one_class_ids[0]
positive_id = weighted_sampling_of_best(
np.apply_along_axis(
partial(norm, feature_vectors[anchor]), 1,
feature_vectors[one_class_ids[1:]]),
best='max')
positive_id += 1
negative_id = weighted_sampling_of_best(
np.apply_along_axis(
partial(norm, feature_vectors[anchor]), 1,
feature_vectors[other_class_ids]), best='min')
triplet_slice = [anchor,
one_class_ids[positive_id],
other_class_ids[negative_id]]
one_class_ids = np.delete(one_class_ids, [0, positive_id])
other_class_ids = np.delete(other_class_ids, negative_id)
delta[triplet_slice] = ltl_grad(
feature_vectors[triplet_slice])
batch_loss.append(linear_triplet_loss(
feature_vectors[triplet_slice]))
else:
for _i in range(_range):
triplet_slice = _offset + (np.arange(3) * _range) + _i
delta[triplet_slice] = ltl_grad(feature_vectors[triplet_slice])
batch_loss.append(linear_triplet_loss(feature_vectors[triplet_slice]))
batch_iteration_hook(
batch_loss=batch_loss,
epoch=epoch,
bid=bid
)
network.processBatchBackward(batch, delta,
learningRate=learning_rate,
momentum=momentum)
network.saveWeights(weights_temp, epoch)
epoch_iteration_hook(_network=network,
learning_rate=learning_rate,
momentum=momentum,
epoch=epoch,
weights_path="{}_epoch-{}.cnn".format(
weights_temp, epoch))
epoch += 1
del train_ds