def __init__(self, model, config_file, train=True):
self.model = model
self.config = utils.get_yaml_contents(config_file)
self.is_train = train
self.exclude_blocks = [
'Mixed_1', 'Mixed_2', 'Mixed_3', 'Mixed_4', 'Mixed_5', 'Mixed_6',
'maxpool4', 'maxpool7', 'inception_fc0', 'inception_fc1',
'dp0_inception_fc', 'dp1_inception_fc', 'pc_fc0', 'pc_fc1',
'dp0_pc_fc', 'dp1_pc_fc', 'dp_combined', 'fc1', 'dp1', 'logits',
'Logits', 'Predictions', 'Logits/biases', 'agg', 'agg/biases',
'transform_net1', 'dgcnn1', 'dgcnn2', 'dgcnn3', 'dgcnn4', 'fc2',
'fc3', 'dp2'
]
if self.model == "3DmFV":
self.m = self.config['models']['3DmFV']['n_gaussians']
self.gmm_variance = self.config['models']['3DmFV']['gmm_variance']
self.gmm = utils.get_3d_grid_gmm(
subdivisions=[self.m, self.m, self.m],
variance=self.gmm_variance)
self.set_config()
self.set_logdir()
self.load_train_data()
self.input_dim = 6 if self.model_config.pointcloud_color else 3
self.load_model()
cos = cos_ang[expert_to_use, range(len(expert_to_use))]
ang_err_batch = np.rad2deg(np.arccos(np.abs(cos))) # unoriented
ang_err.append(ang_err_batch)
loss_sum += loss_val
test_writer.add_summary(summary, step)
total_seen += BATCH_SIZE
mean_loss = loss_sum / float(test_num_batchs)
log_string('eval mean loss: %f' % (mean_loss))
ang_err = np.reshape(ang_err, [n_shapes, PATCHES_PER_SHAPE])
rms = np.sqrt(np.mean(np.square(ang_err), axis=1))
mean_rms = np.mean(rms)
log_string('eval mean rms: %f' % (mean_rms))
tf.summary.scalar('Eval RMS angle', mean_rms)
return mean_loss
if __name__ == "__main__":
gmm = utils.get_3d_grid_gmm(
subdivisions=[N_GAUSSIANS, N_GAUSSIANS, N_GAUSSIANS],
variance=GMM_VARIANCE)
pickle.dump(gmm, open(os.path.join(LOG_DIR, 'gmm.p'), "wb"))
train(gmm)
LOG_FOUT.close()
default='data/test/PMD_M20.pcd')
parser.add_argument(
'--dataset',
help='cloud to test',
default='/media/emha/HDD4/pointcloud_dataset/pcd/asus/test')
# /home/emha/pmd_multiview_dataset/
# /media/emha/HDD1/Dataset/rgbd-dataset-51/
args = parser.parse_args()
cloud = args.pcd
data_folder = args.dataset
gmm = None
if fe is not "mc":
n_gaussian = int(''.join(filter(str.isdigit, fe)))
print(n_gaussian)
gmm = utils.get_3d_grid_gmm(
subdivisions=[n_gaussian, n_gaussian, n_gaussian],
variance=0.05)
acc = evaluate_maxrdd_fv_cls(data_folder,
classifier,
objects='all',
color=True,
gmm=gmm,
feature_extraction=fe,
mc_feature=True)
#acc = evaluate_cls_using_pgz(data_folder, gmm, classifier, fe, [10,13])
accuracies.append(acc)
indices.append(int(idx))
plt.scatter(indices, accuracies)