def inference(model_config, eval_config, dataset_config, data_split,
ckpt_indices):
# Overwrite the defaults
dataset_config = config_builder.proto_to_obj(dataset_config)
dataset_config.data_split = data_split
dataset_config.data_split_dir = 'training'
if data_split == 'test':
dataset_config.data_split_dir = 'testing'
eval_config.eval_mode = 'test'
eval_config.evaluate_repeatedly = False
dataset_config.has_labels = False
# Enable this to see the actually memory being used
eval_config.allow_gpu_mem_growth = True
eval_config = config_builder.proto_to_obj(eval_config)
# Grab the checkpoint indices to evaluate
eval_config.ckpt_indices = ckpt_indices
# Remove augmentation during evaluation in test mode
dataset_config.aug_list = []
# Build the dataset object
dataset = DatasetBuilder.build_kitti_dataset(dataset_config,
use_defaults=False)
# Setup the model
model_name = model_config.model_name
# Overwrite repeated field
model_config = config_builder.proto_to_obj(model_config)
# Switch path drop off during evaluation
model_config.path_drop_probabilities = [1.0, 1.0]
with tf.Graph().as_default():
if model_name == 'mlod_model':
model = MlodModel(model_config,
train_val_test=eval_config.eval_mode,
dataset=dataset)
elif model_name == 'rpn_model':
model = RpnModel(model_config,
train_val_test=eval_config.eval_mode,
dataset=dataset)
else:
raise ValueError('Invalid model name {}'.format(model_name))
model_evaluator = Evaluator(model, dataset_config, eval_config)
model_evaluator.run_latest_checkpoints()
def set_up_model():
test_pipeline_config_path = mlod.root_dir() + \
'/configs/mlod_exp_example.config'
model_config, train_config, _, dataset_config = \
config_builder.get_configs_from_pipeline_file(
test_pipeline_config_path, is_training=True)
dataset_config = config_builder.proto_to_obj(dataset_config)
train_val_test = 'test'
dataset_config.data_split = 'test'
dataset_config.data_split_dir = 'testing'
dataset_config.has_labels = False
dataset_config.aug_list = []
dataset = DatasetBuilder.build_kitti_dataset(dataset_config,
use_defaults=False)
model_name = model_config.model_name
if model_name == 'rpn_model':
model = RpnModel(model_config,
train_val_test=train_val_test,
dataset=dataset)
elif model_name == 'mlod_model':
model = MlodModel(model_config,
train_val_test=train_val_test,
dataset=dataset)
else:
raise ValueError('Invalid model_name')
return model
def set_up_model_test_mode(pipeline_config_path, data_split):
"""Returns the model and its config in test mode."""
model_config, _, _, dataset_config = \
config_builder.get_configs_from_pipeline_file(
pipeline_config_path, is_training=False)
dataset = DatasetBuilder.build_kitti_dataset(dataset_config,
use_defaults=False)
# Overwrite the defaults
dataset_config = config_builder.proto_to_obj(dataset_config)
# Use the validation set
dataset_config.data_split = data_split
dataset_config.data_split_dir = 'training'
if data_split == 'test':
dataset_config.data_split_dir = 'testing'
# Remove augmentation when in test mode
dataset_config.aug_list = []
# Build the dataset object
dataset = DatasetBuilder.build_kitti_dataset(dataset_config,
use_defaults=False)
model_name = model_config.model_name
if model_name == 'rpn_model':
model = RpnModel(model_config, train_val_test='test', dataset=dataset)
elif model_name == 'mlod_model':
model = MlodModel(model_config, train_val_test='test', dataset=dataset)
else:
raise ValueError('Invalid model_name')
return model, model_config
def test_path_drop_weights(self):
# Tests the effect of path-drop on network's feature maps.
# It sets up a minimal-training process to check the
# feature before and after running the 'train_op' while
# path-drop is in effect.
train_val_test = 'train'
# overwrite the training iterations
self.train_config.max_iterations = 2
self.train_config.overwrite_checkpoints = True
# Overwrite path drop probabilities
model_config = config_builder.proto_to_obj(self.model_config)
model_config.path_drop_probabilities = [0.0, 0.8]
with tf.Graph().as_default():
# Set a graph-level seed
tf.set_random_seed(1245)
model = RpnModel(model_config,
train_val_test=train_val_test,
dataset=self.dataset)
prediction_dict = model.build()
losses_dict, total_loss = model.loss(prediction_dict)
global_summaries = set([])
# Optimizer
training_optimizer = optimizer_builder.build(
self.train_config.optimizer, global_summaries)
train_op = slim.learning.create_train_op(total_loss,
training_optimizer)
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
for step in range(1, self.train_config.max_iterations):
feed_dict = model.create_feed_dict()
if step == 1:
current_feature_maps = sess.run(model.img_feature_maps,
feed_dict=feed_dict)
exp_feature_maps = current_feature_maps
train_op_loss = sess.run(train_op, feed_dict=feed_dict)
print('Step {}, Total Loss {:0.3f} '.format(
step, train_op_loss))
updated_feature_maps = sess.run(model.img_feature_maps,
feed_dict=feed_dict)
# The feature maps should have remained the same since
# the image path was dropped
np.testing.assert_array_almost_equal(updated_feature_maps,
exp_feature_maps,
decimal=4)
def set_up_video_dataset(dataset_config, dataset_dir, data_split,
data_split_dir):
# Overwrite fields
dataset_config.name = 'kitti_video'
dataset_config.dataset_dir = dataset_dir
dataset_config.data_split = data_split
dataset_config.data_split_dir = data_split_dir
dataset_config.has_labels = False
# Overwrite repeated fields
dataset_config = config_builder_util.proto_to_obj(dataset_config)
dataset_config.aug_list = []
dataset = DatasetBuilder.build_kitti_dataset(dataset_config,
use_defaults=False)
return dataset
def test_disable_path_drop(self):
# Test path drop is disabled when the probabilities
# are set to 1.0.
train_val_test = 'train'
# Overwrite path drop probabilities
model_config = config_builder.proto_to_obj(self.model_config)
model_config.path_drop_probabilities = [1.0, 1.0]
with tf.Graph().as_default():
model = RpnModel(model_config,
train_val_test=train_val_test,
dataset=self.dataset)
model.build()
# These variables are set during path drop only
# in the case of no path-drop, they should be non-existence
self.assertFalse(hasattr(model, 'img_path_drop_mask'))
self.assertFalse(hasattr(model, 'bev_path_drop_mask'))
def main():
""" Converts a set of network predictions into text files required for
KITTI evaluation.
"""
##############################
# Options
##############################
checkpoint_name = 'mlod_exp_example'
# data_split = 'val'
data_split = 'val_half'
global_steps = None
# global_steps = [28000, 19000, 33000, 34000]
score_threshold = 0.1
save_2d = False # Save 2D predictions
save_3d = True # Save 2D and 3D predictions together
# Checkpoints below this are skipped
min_step = 20000
# Object Type
obj_type = 'obj'
##############################
# End of Options
##############################
# Parse experiment config
pipeline_config_file = \
mlod.root_dir() + '/data/outputs/' + checkpoint_name + \
'/' + checkpoint_name + '.config'
_, _, _, dataset_config = \
config_builder_util.get_configs_from_pipeline_file(
pipeline_config_file, is_training=False)
# Overwrite defaults
dataset_config = config_builder_util.proto_to_obj(dataset_config)
dataset_config.data_split = data_split
dataset_config.aug_list = []
if data_split == 'test':
dataset_config.data_split_dir = 'testing'
dataset = DatasetBuilder.build_kitti_dataset(dataset_config,
use_defaults=False)
# Get available prediction folders
predictions_root_dir = mlod.root_dir() + '/data/outputs/' + \
checkpoint_name + '/predictions'
proposals_root_dir = predictions_root_dir + \
'/proposals_and_scores/' + dataset.data_split
print('Converting proposals from', proposals_root_dir)
if not global_steps:
global_steps = os.listdir(proposals_root_dir)
global_steps.sort(key=int)
print('Checkpoints found ', global_steps)
for step_idx in range(len(global_steps)):
global_step = global_steps[step_idx]
# Skip first checkpoint
if int(global_step) < min_step:
continue
final_predictions_dir = proposals_root_dir + \
'/' + str(global_step)
# 2D and 3D prediction directories
kitti_predictions_2d_dir = predictions_root_dir + \
'/kitti_proposals_2d/' + \
dataset.data_split + '/' + \
str(score_threshold) + '/' + \
str(global_step) + '/data'
kitti_proposals_3d_dir = predictions_root_dir + \
'/kitti_proposals_3d/' + \
dataset.data_split + '/' + \
str(score_threshold) + '/' + \
str(global_step) + '/data'
if save_2d and not os.path.exists(kitti_predictions_2d_dir):
os.makedirs(kitti_predictions_2d_dir)
if save_3d and not os.path.exists(kitti_proposals_3d_dir):
os.makedirs(kitti_proposals_3d_dir)
# Do conversion
num_samples = dataset.num_samples
num_valid_samples = 0
print('\nGlobal step:', global_step)
print('Converting proposals from:', final_predictions_dir)
if save_2d:
print('2D Detections saved to:', kitti_predictions_2d_dir)
if save_3d:
print('Proposals saved to:', kitti_proposals_3d_dir)
for sample_idx in range(num_samples):
# Print progress
sys.stdout.write('\rConverting {} / {}'.format(
sample_idx + 1, num_samples))
sys.stdout.flush()
sample_name = dataset.sample_names[sample_idx]
prediction_file = sample_name + '.txt'
kitti_predictions_2d_file_path = kitti_predictions_2d_dir + \
'/' + prediction_file
kitti_predictions_3d_file_path = kitti_proposals_3d_dir + \
'/' + prediction_file
predictions_file_path = final_predictions_dir + \
'/' + prediction_file
# If no predictions, skip to next file
if not os.path.exists(predictions_file_path):
if save_2d:
np.savetxt(kitti_predictions_2d_file_path, [])
if save_3d:
np.savetxt(kitti_predictions_3d_file_path, [])
continue
all_predictions = np.loadtxt(predictions_file_path)
# Swap l, w for predictions where w > l
swapped_indices = all_predictions[:, 4] > all_predictions[:, 3]
fixed_predictions = np.copy(all_predictions)
fixed_predictions[swapped_indices,
3] = all_predictions[swapped_indices, 4]
fixed_predictions[swapped_indices,
4] = all_predictions[swapped_indices, 3]
fixed_predictions[swapped_indices, 6] = np.pi / 2
score_filter = all_predictions[:, 7] >= score_threshold
all_predictions = fixed_predictions[score_filter]
# If no predictions, skip to next file
if len(all_predictions) == 0:
if save_2d:
np.savetxt(kitti_predictions_2d_file_path, [])
if save_3d:
np.savetxt(kitti_predictions_3d_file_path, [])
continue
# Project to image space
sample_name = prediction_file.split('.')[0]
img_idx = int(sample_name)
# Load image for truncation
image = Image.open(dataset.get_rgb_image_path(sample_name))
stereo_calib_p2 = calib_utils.read_calibration(
dataset.calib_dir, img_idx).p2
boxes = []
image_filter = []
for i in range(len(all_predictions)):
box_3d = all_predictions[i, 0:7]
img_box = box_3d_projector.project_to_image_space(
box_3d,
stereo_calib_p2,
truncate=True,
image_size=image.size,
discard_before_truncation=False)
# Skip invalid boxes (outside image space)
if img_box is None:
image_filter.append(False)
else:
image_filter.append(True)
boxes.append(img_box)
boxes = np.asarray(boxes)
all_predictions = all_predictions[image_filter]
# If no predictions, skip to next file
if len(boxes) == 0:
if save_2d:
np.savetxt(kitti_predictions_2d_file_path, [])
if save_3d:
np.savetxt(kitti_predictions_3d_file_path, [])
continue
num_valid_samples += 1
# To keep each value in its appropriate position, an array of zeros
# (N, 16) is allocated but only values [4:16] are used
kitti_predictions = np.zeros([len(boxes), 16])
# Truncation and Occlusion are always empty (see below)
# Alpha (Not computed)
kitti_predictions[:, 3] = -10 * np.ones(
(len(kitti_predictions)), dtype=np.int32)
# 2D predictions
kitti_predictions[:, 4:8] = boxes[:, 0:4]
# 3D predictions
# (l, w, h)
kitti_predictions[:, 8] = all_predictions[:, 5]
kitti_predictions[:, 9] = all_predictions[:, 4]
kitti_predictions[:, 10] = all_predictions[:, 3]
# (x, y, z)
kitti_predictions[:, 11:14] = all_predictions[:, 0:3]
# (ry, score)
kitti_predictions[:, 14:16] = all_predictions[:, 6:8]
# Round detections to 3 decimal places
kitti_predictions = np.round(kitti_predictions, 3)
# Empty Truncation, Occlusion
kitti_empty_1 = -1 * np.ones(
(len(kitti_predictions), 2), dtype=np.int32)
# Empty 3D (x, y, z)
kitti_empty_2 = -1 * np.ones(
(len(kitti_predictions), 3), dtype=np.int32)
# Empty 3D (h, w, l)
kitti_empty_3 = -1000 * np.ones(
(len(kitti_predictions), 3), dtype=np.int32)
# Empty 3D (ry)
kitti_empty_4 = -10 * np.ones(
(len(kitti_predictions), 1), dtype=np.int32)
# Create Type Array
obj_types = [obj_type for i in range(len(kitti_predictions))]
# Stack 2D predictions text
kitti_text_2d = np.column_stack([
obj_types, kitti_empty_1, kitti_predictions[:, 3:8],
kitti_empty_2, kitti_empty_3, kitti_empty_4,
kitti_predictions[:, 15]
])
# Stack 3D predictions text
kitti_text_3d = np.column_stack(
[obj_types, kitti_empty_1, kitti_predictions[:, 3:16]])
# Save to text files
if save_2d:
np.savetxt(kitti_predictions_2d_file_path,
kitti_text_2d,
newline='\r\n',
fmt='%s')
if save_3d:
np.savetxt(kitti_predictions_3d_file_path,
kitti_text_3d,
newline='\r\n',
fmt='%s')
print('\nNum valid:', num_valid_samples)
print('Num samples:', num_samples)
model_config, _, eval_config, dataset_config = \
config_builder.get_configs_from_pipeline_file(
args.pipeline_config_path,
is_training=False)
# Overwrite data split
dataset_config.data_split = 'trainval'
dataset_config.data_split_dir = 'training'
dataset_config.has_labels = True
# Set CUDA device id
os.environ['CUDA_VISIBLE_DEVICES'] = args.device
# Convert to object to overwrite repeated fields
dataset_config = config_builder.proto_to_obj(dataset_config)
# Remove augmentation during evaluation
dataset_config.aug_list = ['flipping']
# Build the dataset object
dataset = DatasetBuilder.build_kitti_dataset(dataset_config,
use_defaults=False)
# Setup the model
model_name = model_config.model_name
# Convert to object to overwrite repeated fields
model_config = config_builder.proto_to_obj(model_config)
# Switch path drop off during evaluation