syn_id = snc_category_to_synth_id()[class_name[0]]
class_dir = osp.join(top_in_dir, syn_id)
_, _, pc_data_test = load_and_split_all_point_clouds_under_folder(
class_dir, n_threads=8, file_ending='.ply', verbose=True)
for i in range(1, len(class_name)):
syn_id = snc_category_to_synth_id()[class_name[i]]
class_dir = osp.join(top_in_dir, syn_id)
_, _, pc_data_test_curr = load_and_split_all_point_clouds_under_folder(
class_dir, n_threads=8, file_ending='.ply', verbose=True)
pc_data_test.merge(pc_data_test_curr)
# Load configuration
train_dir = osp.join(top_out_dir, flags.train_folder)
restore_epoch = 500
conf = Conf.load(osp.join(train_dir, 'configuration'))
conf.encoder_args['return_layer_before_symmetry'] = True
# Reload a saved model
reset_tf_graph()
ae = SNetPointNetAutoEncoder(conf.experiment_name, conf)
ae.restore_model(train_dir, epoch=restore_epoch, verbose=True)
n_sample_points = conf.n_samp[0]
# create evaluation dir
eval_dir = create_dir(osp.join(train_dir, "eval"))
# sample point clouds
_, sampled_pc, sample_idx, _ = ae.get_samples(
for i in range(1, len(class_name)):
syn_id = snc_category_to_synth_id()[class_name[i]]
class_dir = osp.join(top_in_dir, syn_id)
pc_data_train_curr, pc_data_val_curr, _ = load_and_split_all_point_clouds_under_folder(
class_dir, n_threads=8, file_ending='.ply', verbose=True)
pc_data_train.merge(pc_data_train_curr)
pc_data_val.merge(pc_data_val_curr)
if flags.object_class == 'multi':
pc_data_train.shuffle_data(seed=55)
pc_data_val.shuffle_data(seed=55)
ae_dir = osp.join(top_out_dir, flags.ae_folder)
# Load autoencoder configuration
conf = Conf.load(osp.join(ae_dir, 'configuration'))
# Update autoencoder configuration
conf.ae_dir = ae_dir
conf.ae_name = 'autoencoder'
conf.restore_ae = flags.restore_ae
conf.ae_restore_epoch = 500
conf.fixed_ae = flags.fixed_ae
if conf.fixed_ae:
conf.encoder_args[
'b_norm_decay'] = 1. # for avoiding the update of batch normalization moving_mean and moving_variance parameters
conf.decoder_args[
'b_norm_decay'] = 1. # for avoiding the update of batch normalization moving_mean and moving_variance parameters
conf.decoder_args[
'b_norm_decay_finish'] = 1. # for avoiding the update of batch normalization moving_mean and moving_variance parameters
class_dir = osp.join(top_in_dir, syn_id)
_, _, pc_data_test = load_and_split_all_point_clouds_under_folder(
class_dir, n_threads=8, file_ending=".ply", verbose=True
)
for i in range(1, len(class_name)):
syn_id = snc_category_to_synth_id()[class_name[i]]
class_dir = osp.join(top_in_dir, syn_id)
_, _, pc_data_test_curr = load_and_split_all_point_clouds_under_folder(
class_dir, n_threads=8, file_ending=".ply", verbose=True
)
pc_data_test.merge(pc_data_test_curr)
# load configuration
train_dir = osp.join(top_out_dir, flags.train_folder)
conf = Conf.load(osp.join(train_dir, "configuration"))
# update configuration
conf.use_batch_size_for_place_holder = False
conf.encoder_args["return_layer_before_symmetry"] = False
conf.hard_projection = flags.hard_projection
conf.pc_size = [flags.n_sample_points]
conf.n_samp = [flags.n_sample_points, 3]
# reload a saved model
reset_tf_graph()
ae = SampleNetProgressivePointNetAE(conf.experiment_name, conf)
ae.restore_model(train_dir, epoch=flags.restore_epoch, verbose=True)
n_input_points = conf.n_input[0]
n_sample_points = conf.n_samp[0]
class_dir = osp.join(top_in_dir, syn_id)
_, _, pc_data_test = load_and_split_all_point_clouds_under_folder(
class_dir, n_threads=8, file_ending=".ply", verbose=True
)
for i in range(1, len(class_name)):
syn_id = snc_category_to_synth_id()[class_name[i]]
class_dir = osp.join(top_in_dir, syn_id)
_, _, pc_data_test_curr = load_and_split_all_point_clouds_under_folder(
class_dir, n_threads=8, file_ending=".ply", verbose=True
)
pc_data_test.merge(pc_data_test_curr)
# load train configuration
train_dir = create_dir(osp.join(top_out_dir, flags.train_folder))
conf = Conf.load(train_dir + "/configuration")
conf.use_fps = flags.use_fps
conf.n_sample_points = flags.n_sample_points
# build AE Model
reset_tf_graph()
ae = PointNetAutoEncoder(conf.experiment_name, conf)
# reload a saved model
ae.restore_model(train_dir, epoch=flags.restore_epoch, verbose=True)
n_sample_points = flags.n_sample_points
# create evaluation dir
eval_dir = create_dir(osp.join(train_dir, "eval"))
pc_data_val_curr,
_,
) = load_and_split_all_point_clouds_under_folder(class_dir,
n_threads=8,
file_ending=".ply",
verbose=True)
pc_data_train.merge(pc_data_train_curr)
pc_data_val.merge(pc_data_val_curr)
if flags.object_class == "multi":
pc_data_train.shuffle_data(seed=55)
pc_data_val.shuffle_data(seed=55)
# load autoencoder configuration
ae_dir = osp.join(top_out_dir, flags.ae_folder)
conf = Conf.load(osp.join(ae_dir, "configuration"))
# update autoencoder configuration
conf.ae_dir = ae_dir
conf.ae_name = "autoencoder"
conf.restore_ae = flags.restore_ae
conf.ae_restore_epoch = flags.ae_restore_epoch
conf.fixed_ae = flags.fixed_ae
if conf.fixed_ae:
conf.encoder_args[
"b_norm_decay"] = 1.0 # for avoiding the update of batch normalization moving_mean and moving_variance parameters
conf.decoder_args[
"b_norm_decay"] = 1.0 # for avoiding the update of batch normalization moving_mean and moving_variance parameters
conf.decoder_args[
"b_norm_decay_finish"] = 1.0 # for avoiding the update of batch normalization moving_mean and moving_variance parameters
# load default training parameters
train_params = default_train_params()
encoder, decoder, enc_args, dec_args = mlp_architecture_ala_iclr_18(
n_pc_points, bneck_size)
train_dir = create_dir(osp.join(top_out_dir, flags.train_folder))
conf = Conf(
n_input=[n_pc_points, 3],
loss=ae_loss,
training_epochs=train_params["training_epochs"],
batch_size=train_params["batch_size"],
denoising=train_params["denoising"],
learning_rate=train_params["learning_rate"],
train_dir=train_dir,
loss_display_step=train_params["loss_display_step"],
saver_step=train_params["saver_step"],
z_rotate=train_params["z_rotate"],
encoder=encoder,
decoder=decoder,
encoder_args=enc_args,
decoder_args=dec_args,
)
conf.experiment_name = experiment_name
conf.held_out_step = 5 # how often to evaluate/print(out loss on)
# held_out data (if they are provided in ae.train()).
conf.class_name = class_name
conf.use_fps = flags.use_fps
conf.n_sample_points = flags.n_sample_points
conf.n_samp_out = [2048, 3]
conf.training_epochs = flags.training_epochs
# Load default training parameters
train_params = default_train_params()
encoder, decoder, enc_args, dec_args = mlp_architecture_ala_iclr_18(
n_pc_points, bneck_size)
train_dir = create_dir(osp.join(top_out_dir, flags.train_folder))
conf = Conf(n_input=[n_pc_points, 3],
loss=ae_loss,
training_epochs=train_params['training_epochs'],
batch_size=train_params['batch_size'],
denoising=train_params['denoising'],
learning_rate=train_params['learning_rate'],
train_dir=train_dir,
loss_display_step=train_params['loss_display_step'],
saver_step=train_params['saver_step'],
z_rotate=train_params['z_rotate'],
encoder=encoder,
decoder=decoder,
encoder_args=enc_args,
decoder_args=dec_args)
conf.experiment_name = experiment_name
conf.held_out_step = 5 # How often to evaluate/print out loss on
# held_out data (if they are provided in ae.train()).
conf.class_name = class_name
conf.use_fps = flags.use_fps
conf.n_sample_points = flags.n_sample_points
conf.n_samp_out = [2048, 3]
conf.save(osp.join(train_dir, 'configuration'))
