how to apply a model to calculate the saliency of the point cloud?

Question

I have a point-clould of which I would like to meassure saliency.using this model after train here is the code:

import tensorflow as tf
import numpy as np
import argparse
import socket
import importlib
import time
import os
import scipy.misc
import sys
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(BASE_DIR)
sys.path.append(os.path.join(BASE_DIR, 'models'))
sys.path.append(os.path.join(BASE_DIR, 'utils'))
import provider
import pc_util

parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=0, help='GPU to use [default: GPU 0]')
parser.add_argument('--model', default='pointnet_cls',
                help='Model name: pointnet_cls or pointnet_cls_basic [default: 
pointnet_cls]')
parser.add_argument('--batch_size', type=int, default=4, help='Batch Size during 
evaluating, recommend: 1,2,4,8 [default: 4]')
parser.add_argument('--num_point', type=int, default=1024, help='Point Number 
[256/512/1024/2048] [default: 1024]')
parser.add_argument('--model_path', default='log/model.ckpt',
                help='model checkpoint file path [default: log/model.ckpt]')
parser.add_argument('--dump_dir', default='dump', help='dump folder path [dump]')
parser.add_argument('--visu', action='store_true', default=False, help='Whether to dump 
image for error case [default: False]')
parser.add_argument('--total_droppoints', type=int, default=100, help='Number of drop 
points [default: 100]')
parser.add_argument('--num_per_iteration', type=int, default=5,
                help='Number of points dropped per iteration, where 
mod(total_droppoints,num_per_iteration) need to be 0 [default: 5]')
parser.add_argument('--lowdrop', action='store_true', default=False, help='Whether to 
use 
lowdrop, True: lowdrop, False:highdrop [default: False]')
parser.add_argument('--pred_labels', action='store_true', default=False, help='Whether 
to 
compute the gradients with the predicted labels, True: predicted labels, False: ground 
truth label [default: False]')
parser.add_argument('--scale_parameter', type=float, default=1, help='Scale parameter 
alpha [default: 1]')
FLAGS = parser.parse_args()

GPU_INDEX = FLAGS.gpu
MODEL = importlib.import_module(FLAGS.model)  # import network module
BATCH_SIZE = FLAGS.batch_size
NUM_POINT = FLAGS.num_point
MODEL_PATH = FLAGS.model_path
DUMP_DIR = FLAGS.dump_dir
TOTAL_DROP = FLAGS.total_droppoints
PER_ITER = FLAGS.num_per_iteration
SCALE_ALPHA = FLAGS.scale_parameter
PRED_LABELS = FLAGS.pred_labels
if not os.path.exists(DUMP_DIR): os.mkdir(DUMP_DIR)
LOG_FOUT = open(os.path.join(DUMP_DIR, 'log_evaluate.txt'), 'w')
LOG_FOUT.write(str(FLAGS) + '\n')



NUM_CLASSES = 40
SHAPE_NAMES = [line.rstrip() for line in \
           open(os.path.join(BASE_DIR, 
'data/modelnet40_ply_hdf5_2048/shape_names.txt'))]

HOSTNAME = socket.gethostname()

# ModelNet40 official train/test split
TRAIN_FILES = provider.getDataFiles( \
os.path.join(BASE_DIR, 'data/modelnet40_ply_hdf5_2048/train_files.txt'))
TEST_FILES = provider.getDataFiles( \
os.path.join(BASE_DIR, 'data/modelnet40_ply_hdf5_2048/test_files.txt'))


def log_string(out_str):
LOG_FOUT.write(out_str + '\n')
LOG_FOUT.flush()
print(out_str)


def saliency_map_and_evaluate(num_votes=1):
is_training = False

with tf.device('/gpu:' + str(GPU_INDEX)):
    pointclouds_pl, labels_pl = MODEL.placeholder_inputs(BATCH_SIZE, NUM_POINT)
    is_training_pl = tf.placeholder(tf.bool, shape=())

    # simple model
    pred, end_points = MODEL.get_model(pointclouds_pl, is_training_pl)
    loss = MODEL.get_loss(pred, labels_pl, end_points)
    if not PRED_LABELS:
        classify_loss = MODEL.get_loss_classification_only(pred, labels_pl, end_points)  
#Compute loss via true labels
    else:
        labels_pre = tf.argmax(pred, axis=1, output_type=tf.int32)
        classify_loss = MODEL.get_loss_classification_only(pred, labels_pre, end_points) 
#Compute loss via predict labels

    #Compute the loss's gradients about the points' coordinates
    #Only classification loss is used here
    gradients=tf.gradients(classify_loss,pointclouds_pl)

    # Add ops to save and restore all the variables.
    saver = tf.train.Saver()

# Create a session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
config.log_device_placement = True
sess = tf.Session(config=config)

# Restore variables from disk.
saver.restore(sess, MODEL_PATH)
log_string("Model restored.")

ops = {'pointclouds_pl': pointclouds_pl,
       'labels_pl': labels_pl,
       'is_training_pl': is_training_pl,
       'pred': pred,
       'loss': loss,
       'gradients':gradients}

sali_eval_one_epoch(sess, ops, num_votes)


def sali_eval_one_epoch(sess, ops, num_votes=1, topk=1):
error_cnt = 0
is_training = False
total_correct = 0
total_seen = 0
loss_sum = 0
total_seen_class = [0 for _ in range(NUM_CLASSES)]
total_correct_class = [0 for _ in range(NUM_CLASSES)]
fout = open(os.path.join(DUMP_DIR, 'pred_label.txt'), 'w')
for fn in range(len(TEST_FILES)):
    log_string('----' + str(fn) + '----')
    current_data, current_label = provider.loadDataFile(TEST_FILES[fn])
    current_data = current_data[:, 0:NUM_POINT, :]
    current_label = np.squeeze(current_label)
    print(current_data.shape)
    #Initial adversatial data with current data
    adversarial_data = current_data

    file_size = current_data.shape[0]
    num_batches = file_size // BATCH_SIZE
    print(file_size)

    for batch_idx in range(num_batches):
        start_idx = batch_idx * BATCH_SIZE
        end_idx = (batch_idx + 1) * BATCH_SIZE
        cur_batch_size = end_idx - start_idx

        # Aggregating BEG
        batch_loss_sum = 0  # sum of losses for the batch
        batch_pred_sum = np.zeros((cur_batch_size, NUM_CLASSES))  # score for classes
        batch_pred_classes = np.zeros((cur_batch_size, NUM_CLASSES))  # 0/1 for classes

        #Drop points in a cyclic pattern
        data_now = current_data[start_idx:end_idx, :, :]   # store the present batch 
 data
        for num_drop in range(TOTAL_DROP//PER_ITER):
            rotated_data = provider.rotate_point_cloud_by_angle(data_now,0)
            feed_dict = {ops['pointclouds_pl']: rotated_data,
                     ops['labels_pl']: current_label[start_idx:end_idx],
                     ops['is_training_pl']: is_training}
            gradients_val = sess.run(ops['gradients'],feed_dict=feed_dict)
            #Compute the saliency score
            gradients_val = np.reshape(gradients_val,(BATCH_SIZE,NUM_POINT,-1))
            point_radius = np.sqrt(np.sum(np.square(data_now),-1))
            saliency_score = - 
 np.power(point_radius,SCALE_ALPHA)*np.sum(data_now*gradients_val,-1)

            if FLAGS.lowdrop:
                saliency_score *=-1

            # Drop points according to the saliency score
            # Here, instead of directly deleting high/low score points, the points' 
 coordinates are replaced with 0
            drop_index = np.argpartition(saliency_score[:,0:NUM_POINT- 
num_drop*PER_ITER], -PER_ITER, axis=1)
            for point_idx in range(BATCH_SIZE):
                data_now[point_idx:point_idx + 1, 0:NUM_POINT - num_drop * PER_ITER, :] 
 = \
                    
data_now[point_idx:point_idx+1,drop_index[point_idx:point_idx+1,0:NUM_POINT - num_drop * 
PER_ITER],:]
                data_now[point_idx:point_idx+1, NUM_POINT - (num_drop + 1) * PER_ITER:, 
 :] = 0
        adversarial_data[start_idx:end_idx, :, :] = data_now
        print(batch_idx)


        for vote_idx in range(num_votes):
            # rotated_data = 
 provider.rotate_point_cloud_by_angle(current_data[start_idx:end_idx, :, :],
            #                                                     vote_idx / 
float(num_votes) * np.pi * 2)
            rotated_data = 
provider.rotate_point_cloud_by_angle(adversarial_data[start_idx:end_idx, :, :],
                                                                vote_idx / 
float(num_votes) * np.pi * 2)
            feed_dict = {ops['pointclouds_pl']: rotated_data,
                         ops['labels_pl']: current_label[start_idx:end_idx],
                         ops['is_training_pl']: is_training}
            loss_val, pred_val = sess.run([ops['loss'], ops['pred']],
                                          feed_dict=feed_dict)
            batch_pred_sum += pred_val
            batch_pred_val = np.argmax(pred_val, 1)
            for el_idx in range(cur_batch_size):
                batch_pred_classes[el_idx, batch_pred_val[el_idx]] += 1
            batch_loss_sum += (loss_val * cur_batch_size / float(num_votes))
        # pred_val_topk = np.argsort(batch_pred_sum, axis=-1) 
 [:,-1*np.array(range(topk))-1]
        # pred_val = np.argmax(batch_pred_classes, 1)
        pred_val = np.argmax(batch_pred_sum, 1)
        # Aggregating END

        correct = np.sum(pred_val == current_label[start_idx:end_idx])
        # correct = np.sum(pred_val_topk[:,0:topk] == label_val)
        total_correct += correct
        total_seen += cur_batch_size
        loss_sum += batch_loss_sum

        for i in range(start_idx, end_idx):
            l = current_label[i]
            total_seen_class[l] += 1
            total_correct_class[l] += (pred_val[i - start_idx] == l)
            fout.write('%d, %d\n' % (pred_val[i - start_idx], l))

            if pred_val[i - start_idx] != l and FLAGS.visu:  # ERROR CASE, DUMP!
                img_filename = '%d_label_%s_pred_%s.jpg' % (error_cnt, SHAPE_NAMES[l],
                                                            SHAPE_NAMES[pred_val[i - 
 start_idx]])
                img_filename = os.path.join(DUMP_DIR, img_filename)
                output_img = pc_util.point_cloud_three_views(np.squeeze(current_data[i, 
:, :]))
                scipy.misc.imsave(img_filename, output_img)
                error_cnt += 1

log_string('eval mean loss: %f' % (loss_sum / float(total_seen)))
log_string('eval accuracy: %f' % (total_correct / float(total_seen)))
log_string('eval avg class acc: %f' % (
    np.mean(np.array(total_correct_class) / np.array(total_seen_class, 
dtype=np.float))))

class_accuracies = np.array(total_correct_class) / np.array(total_seen_class, 
dtype=np.float)
for i, name in enumerate(SHAPE_NAMES):
    log_string('%10s:\t%0.3f' % (name, class_accuracies[i]))


if __name__ == '__main__':
os.environ['CUDA_VISIBLE_DEVICES'] = str(GPU_INDEX)
with tf.Graph().as_default():
    saliency_map_and_evaluate(num_votes=1)
LOG_FOUT.close()

I need to apply this model to calculate the saliency of the point cloud What is the common way to do this? how to apply a model to calculate the saliency of the point cloud then generate the score? thank you