第1页 / 共18页
第2页 / 共18页
第3页 / 共18页
第4页 / 共18页
第5页 / 共18页
第6页 / 共18页
第7页 / 共18页
第8页 / 共18页
SGPN中train.py训练代码详细解读.docx
Train.py
# see issue #152
Train.py
import argparse
import tensorflow as tf
import numpy as np
import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
import sys
BASE_DIR = os.path.dirname(os.path.abspath('__file__'))
print (BASE_DIR)
sys.path.append(BASE_DIR)
sys.path.append(os.path.dirname(BASE_DIR))
sys.path.append(os.path.join(BASE_DIR, 'utils'))
sys.path.append(os.path.join(BASE_DIR, 'models'))
import provider
import model
这里有所改动
sys.path.append(os.path.join(BASE_DIR, '../../'))
sys.path.append(os.path.join(BASE_DIR, '../../utils'))
sys.path.append(os.path.join(BASE_DIR, '../../models'))
import provider
from models import model
改为:
sys.path.append(os.path.join(BASE_DIR, 'utils'))
sys.path.append(os.path.join(BASE_DIR, 'models'))
import provider
import model
# Parsing Arguments （解析论据）
parser = argparse.ArgumentParser()
# Experiment Settings （实验设置）
parser.add_argument('--gpu', type=str, default="1", help='GPU to use [default: GPU 1]')
parser.add_argument('--wd', type=float, default=0.9, help='Weight Decay [Default: 0.0]')
parser.add_argument('--epoch', type=int, default=200, help='Number of epochs [default: 50]')
parser.add_argument('--batch', type=int, default=4, help='Batch Size during training [default: 4]')
parser.add_argument('--point_num', type=int, default=4096, help='Point Number')
parser.add_argument('--group_num', type=int, default=50, help='Maximum Group Number in one
pc')
parser.add_argument('--cate_num', type=int, default=13, help='Number of categories')
parser.add_argument('--margin_same', type=float, default=10., help='Double hinge loss margin:
same semantic')
parser.add_argument('--margin_diff',
different semantic')
type=float, default=80., help='Double hinge loss margin:
这里给实验的一些参数设值。
wd = 0.9 (Weight 衰减)
epoch = 200
batch = 4 (培训期间的批量大小)
point_num = 4096 (一台电脑的最大组数)
group_num = 50
cate_num = 13 (分类数量)
margin_same = 10 (双铰链损失余量：相同的语义)
margin_diff = 80 (双铰链损失余量：不同的语义)
# Input&Output Settings （input&output 设置）
parser.add_argument('--output_dir',
type=str,
help='Directory that stores all training logs and trained models')
parser.add_argument('--input_list', type=str, default='data/train_hdf5_file_list.txt', help='Input data
list file')
parser.add_argument('--restore_model',
help='Pretrained model')
FLAGS = parser.parse_args(args=[])
default='checkpoint/stanford_sem_seg',
type=str,
default='checkpoint/stanford_ins_seg',
input&output 参数的设置
output_dir (存储所有培训日志和训练模型的目录)
input_list (输入数据列表文件)
restore_model (预训练模型)
os.environ["CUDA_VISIBLE_DEVICES"] = FLAGS.gpu
TRAINING_FILE_LIST = FLAGS.input_list
PRETRAINED_MODEL_PATH = os.path.join(FLAGS.restore_model, 'trained_models/')
PRETRAINED_MODEL_PATH (预测模型路径) (trained_models 文件夹内)
POINT_NUM = FLAGS.point_num
BATCH_SIZE = FLAGS.batch
OUTPUT_DIR = FLAGS.output_dir
if not os.path.exists(OUTPUT_DIR):
os.makedirs(OUTPUT_DIR)
如果没有 OUTPUT_DIR 的路径 stanford_sem_seg 这个文件夹，创建 stanford_sem_seg 文件
夹。
NUM_GROUPS = FLAGS.group_num
NUM_CATEGORY = FLAGS.cate_num
print('#### Batch Size: {0}'.format(BATCH_SIZE))
print('#### Point Number: {0}'.format(POINT_NUM))
print('#### Training using GPU: {0}'.format(FLAGS.gpu))
DECAY_STEP = 800000.
DECAY_RATE = 0.5
LEARNING_RATE_CLIP = 1e-6
BASE_LEARNING_RATE = 1e-4
MOMENTUM = 0.9
梯度下降的参数数值设置。
TRAINING_EPOCHES = FLAGS.epoch
MARGINS = [FLAGS.margin_same, FLAGS.margin_diff]
print('### Training epoch: {0}'.format(TRAINING_EPOCHES))
MODEL_STORAGE_PATH = os.path.join(OUTPUT_DIR, 'trained_models')
if not os.path.exists(MODEL_STORAGE_PATH):
os.mkdir(MODEL_STORAGE_PATH)
LOG_STORAGE_PATH = os.path.join(OUTPUT_DIR, 'logs')
if not os.path.exists(LOG_STORAGE_PATH):
os.mkdir(LOG_STORAGE_PATH)
SUMMARIES_FOLDER = os.path.join(OUTPUT_DIR, 'summaries')
if not os.path.exists(SUMMARIES_FOLDER):
os.mkdir(SUMMARIES_FOLDER)
在 stanford_sem_seg 这个文件夹中是否有 trained_models、logs、summaries 这三个文件夹，
没有的话创建这些文件夹。
LOG_DIR = FLAGS.output_dir
if not os.path.exists(LOG_DIR):
os.mkdir(LOG_DIR)
os.system('cp %s %s' % (os.path.join(BASE_DIR, 'models/model.py'), LOG_DIR))
model def
os.system('cp %s %s' % (os.path.join(BASE_DIR,
procedure
'train.py'), LOG_DIR))
# bkp of
# bkp of train
调用 model.py 中定义过的函数
运行 train 的程序
def printout(flog, data):
print(data)
flog.write(data + '\n')
定义的 printout 函数
具体作用：后面看了才知道
def train():
with tf.Graph().as_default():
with tf.device('/gpu:' + str(FLAGS.gpu)):
batch = tf.Variable(0, trainable=False, name='batch')
learning_rate = tf.train.exponential_decay(
# base learning rate
# global_var indicating the number of steps
BASE_LEARNING_RATE,
batch * BATCH_SIZE,
DECAY_STEP, # step size
DECAY_RATE,
staircase=True
# decay rate
# Stair-case or continuous decreasing)
learning_rate = tf.maximum(learning_rate, LEARNING_RATE_CLIP)
计 算 指 数 衰 减 的 学 习 率 。 训 练 时 学 习 率 最 好 随 着 训 练 衰 减 。 函 数
tf.train.exponential_decay(learning_rate, global_step, decay_steps, decay_rate, staircase=False,
name=None)为指数衰减函数。
BASE_LEARNING_RATE = 1e-4；BATCH_SIZE = 4；DECAY_STEP = 800000；DECAY_RATE
= 0.5
计算公式如下：
decayed_learning_rate = learning_rate *decay_rate ^ (global_step / decay_steps)
此处 global_step = batch * BATCH_SIZE
lr_op = tf.summary.scalar('learning_rate', learning_rate)
这里是为了实现 tensorboard 可视化 learning_rate
pointclouds_ph,
ptsseglabel_ph,
pts_seglabel_mask_ph,
pts_group_mask_ph, alpha_ph = model. placeholder_inputs(BATCH_SIZE,
POINT_NUM, NUM_GROUPS, NUM_CATEGORY)
ptsgroup_label_ph,
这里的 placeholder_inputs 调用的是 model.py 中的函数（如下）：
def placeholder_inputs(batch_size, num_point, num_group, num_cate):
if num_point == 0:
pointclouds_ph = tf.placeholder(tf.float32, shape=(batch_size, None, 9))
else:
pointclouds_ph = tf.placeholder(tf.float32, shape=(batch_size, num_point, 9))
pts_seglabels_ph = tf.placeholder(tf.int32, shape=(batch_size, num_point, num_cate))
pts_grouplabels_ph = tf.placeholder(tf.float32, shape=(batch_size, num_point, num_group))
pts_seglabel_mask_ph = tf.placeholder(tf.float32, shape=(batch_size, num_point))
pts_group_mask_ph = tf.placeholder(tf.float32, shape=(batch_size, num_point))
alpha_ph = tf.placeholder(tf.float32, shape=())
pts_seglabels_ph,
return
pts_grouplabels_ph,
pointclouds_ph,
pts_group_mask_ph, alpha_ph
pts_seglabel_mask_ph,
pointclouds_ph：(batch_size, num_point, 9)
ptsseglabel_ph：(batch_size, num_point, num_cate)
注：这里的 ptsseglabel_ph 以及 ptsgroup_label_ph 和 model.py 中的 pts_seglabels_ph 以及
pts_grouplabels_ph 不一样
ptsgroup_label_ph：(batch_size, num_point, num_group)
pts_seglabel_mask_ph：(batch_size, num_point)
pts_group_mask_ph：(batch_size, num_point)
alpha_ph：shape = ()未指定形状,则可以输入任何形状的张量。
is_training_ph = tf.placeholder(tf.bool, shape=())
is_training_ph：shape = ()未指定形状,则可以输入任何形状的张量。
labels = {'ptsgroup': ptsgroup_label_ph,
'semseg': ptsseglabel_ph,
'semseg_mask': pts_seglabel_mask_ph,
'group_mask': pts_group_mask_ph}
labels(标签)：这是一段字典，'ptsgroup'为 key，ptsgroup_label_ph 为 value。
net_output = model.get_model(pointclouds_ph, is_training_ph,
group_cate_num=NUM_CATEGORY, m=MARGINS[0])
loss, grouperr, same, same_cnt, diff, diff_cnt, pos, pos_cnt =
model.get_loss(net_output, labels, alpha_ph, MARGINS)
is_training_ph 上面有定义，shape = ()未指定形状,则可以输入任何形状的张量。
group_cate_num = NUM_CATEGORY = cate_num = 13
m = MARGINS[0]
这里的 get_model 调用的是 model.py 中的函数（如下）：
def get_model(point_cloud, is_training, group_cate_num=50, m=10., bn_decay=None):
#input: point_cloud: BxNx9 (XYZ, RGB, NormalizedXYZ)
batch_size = point_cloud.get_shape()[0].value
print(point_cloud.get_shape())
F = pointnet.get_model(point_cloud, is_training, bn=True, bn_decay=bn_decay)
# Semantic prediction（语义预测）
Fsem = tf_util.conv2d(F, 128, [1, 1], padding='VALID', stride=[1, 1], bn=False,
ptssemseg_logits = tf_util.conv2d(Fsem, group_cate_num,
[1, 1], padding='VALID',
stride=[1, 1], activation_fn=None, scope='ptssemseg_logits')
is_training=is_training, scope='Fsem')
ptssemseg_logits = tf.squeeze(ptssemseg_logits, [2])
ptssemseg = tf.nn.softmax(ptssemseg_logits, name="ptssemseg")
# Similarity matrix（相似矩阵）
Fsim = tf_util.conv2d(F, 128,
[1, 1], padding='VALID',
is_training=is_training, scope='Fsim')
stride=[1, 1], bn=False,
Fsim = tf.squeeze(Fsim, [2])
r = tf.reduce_sum(Fsim * Fsim, 2)
r = tf.reshape(r, [batch_size, -1, 1])
print(r.get_shape(),Fsim.get_shape())
D = r - 2 * tf.matmul(Fsim, tf.transpose(Fsim, perm=[0, 2, 1])) + tf.transpose(r, perm=[0, 2,
1])
# simmat_logits = tf.maximum(D, 0.)
simmat_logits = tf.maximum(m * D, 0.)
# Confidence Map（置信度图）
Fconf = tf_util.conv2d(F, 128,
[1, 1], padding='VALID',
stride=[1, 1], bn=False,
conf_logits = tf_util.conv2d(Fconf,
1,
is_training=is_training, scope='Fsconf')
1],
activation_fn=None, scope='conf_logits')
[1,
padding='VALID',
stride=[1,
1],
conf_logits = tf.squeeze(conf_logits, [2])
conf = tf.nn.sigmoid(conf_logits, name="confidence")
return {'semseg': ptssemseg,
'semseg_logits': ptssemseg_logits,
'simmat': simmat_logits,
'conf': conf,
'conf_logits': conf_logits}
input: point_cloud: BxNx9 (XYZ, RGB, NormalizedXYZ)
.get_shape()[i].value：直接获取 TensorShape 变量的第 i 个维度值，比如 a 的 shape（2,3,4），
则 a.get_shape()[1].value = 3。这里的 point_cloud.get_shape()[0].value 是为了得到 point_cloud
的第一个的维度值 B。
def get_model(point_cloud, is_training, bn=True, bn_decay=None):
""" ConvNet baseline, input is BxNx9 gray image """
batch_size = point_cloud.get_shape()[0].value
num_point = point_cloud.get_shape()[1].value
input_image = tf.expand_dims(point_cloud, -1)
# CONV
net = tf_util.conv2d(input_image, 64, [1, 9], padding='VALID', stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv1',
bn_decay=bn_decay)
net = tf_util.conv2d(net, 64, [1, 1], padding='VALID', stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv2',
bn_decay=bn_decay)
net = tf_util.conv2d(net, 64, [1, 1], padding='VALID', stride=[1, 1],
is_training=is_training,
bn=bn,
scope='conv3',
bn_decay=bn_decay)
net = tf_util.conv2d(net, 128, [1, 1], padding='VALID', stride=[1, 1],
bn=bn, is_training=is_training, scope='conv4', bn_decay=bn_decay)
points_feat1 = tf_util.conv2d(net, 1024, [1, 1], padding='VALID', stride=[1, 1],
bn=bn, is_training=is_training, scope='conv5', bn_decay=bn_decay)
# MAX
pc_feat1
=
tf_util.max_pool2d(points_feat1,
[num_point,
1],
padding='VALID',
scope='maxpool1')
# FC
pc_feat1 = tf.reshape(pc_feat1, [batch_size, -1])
pc_feat1 = tf_util.fully_connected(pc_feat1, 256, bn=bn, is_training=is_training, scope='fc1',
bn_decay=bn_decay)
pc_feat1 = tf_util.fully_connected(pc_feat1, 128, bn=bn, is_training=is_training, scope='fc2',
bn_decay=bn_decay)
print(pc_feat1)
# CONCAT
pc_feat1_expand = tf.tile(tf.reshape(pc_feat1, [batch_size, 1, 1, -1]), [1, num_point, 1, 1])
points_feat1_concat = tf.concat(axis=3, values=[points_feat1, pc_feat1_expand])
# CONV
net = tf_util.conv2d(points_feat1_concat, 512, [1, 1], padding='VALID', stride=[1, 1], bn=bn,
is_training=is_training, scope='conv6')
256,
is_training=is_training, scope='conv7')
net = tf_util.conv2d(net,
[1,
1],
padding='VALID',
stride=[1,
1],
bn=bn,
# net = tf_util.dropout(net, keep_prob=0.7, is_training=is_training, scope='dp1')
# net = tf_util.conv2d(net, 13, [1, 1], padding='VALID', stride=[1, 1],
# activation_fn=None, scope='conv8')
# net = tf.squeeze(net, [2])
return net
def conv2d(inputs,num_output_channels,kernel_size,scope,stride=[1, 1],padding='SAME',
use_xavier=True,stddev=1e-3,weight_decay=0.0,activation_fn=tf.nn.relu,bn=False,
bn_decay=None,is_training=None):与 tf_util.conv2d(F, 128, [1, 1], padding='VALID', stride=[1,
1], bn=False, is_training=is_training, scope='Fsem')的参数一一对应
这里的 get_ loss 调用的是 model.py 中的函数（如下）
def get_loss(net_output, labels, alpha=10., margin=[1.,2.]):
"""
input:
net_output:{'semseg', 'semseg_logits','simmat','conf','conf_logits'}
labels:{'ptsgroup', 'semseg','semseg_mask','group_mask'}
"""
pts_group_label = labels['ptsgroup']
pts_semseg_label = labels['semseg']
group_mask = tf.expand_dims(labels['group_mask'], dim=2)
pred_confidence_logits = net_output['conf']
pred_simmat = net_output['simmat']
# Similarity Matrix loss（相似性矩阵损失）
B = pts_group_label.get_shape()[0]
N = pts_group_label.get_shape()[1]
onediag = tf.ones([B,N], tf.float32)
group_mat_label = tf.matmul(pts_group_label,tf.transpose(pts_group_label, perm=[0, 2, 1]))
#BxNxN: (i,j) if i and j in the same group（如果 i 和 j 在同一组中）
group_mat_label = tf.matrix_set_diag(group_mat_label,onediag)
sem_mat_label = tf.cast(tf.matmul(pts_semseg_label,tf.transpose(pts_semseg_label, perm=[0,
2, 1])), tf.float32) #BxNxN: (i,j) if i and j are the same semantic
category
sem_mat_label = tf.matrix_set_diag(sem_mat_label,onediag)
samesem_mat_label = sem_mat_label
diffsem_mat_label = tf.subtract(1.0, sem_mat_label)
samegroup_mat_label = group_mat_label
diffgroup_mat_label = tf.subtract(1.0, group_mat_label)
diffgroup_samesem_mat_label = tf.multiply(diffgroup_mat_label, samesem_mat_label)
diffgroup_diffsem_mat_label = tf.multiply(diffgroup_mat_label, diffsem_mat_label)
num_samegroup = tf.reduce_sum(samegroup_mat_label)
num_diffgroup_samesem = tf.reduce_sum(diffgroup_samesem_mat_label)
num_diffgroup_diffsem = tf.reduce_sum(diffgroup_diffsem_mat_label)
# Double hinge loss（双铰链损失）
C_same = tf.constant(margin[0], name="C_same") # same semantic category
C_diff = tf.constant(margin[1], name="C_diff") # different semantic category
pos = tf.multiply(samegroup_mat_label, pred_simmat) # minimize distances if in the same
group
neg_samesem = alpha * tf.multiply(diffgroup_samesem_mat_label,
neg_diffsem = tf.multiply(diffgroup_diffsem_mat_label, tf.maximum(tf.subtract(C_diff,
tf.maximum(tf.subtract(C_same, pred_simmat), 0))
pred_simmat), 0))
simmat_loss = neg_samesem + neg_diffsem + pos
group_mask_weight = tf.matmul(group_mask, tf.transpose(group_mask, perm=[0, 2, 1]))
# simmat_loss = tf.add(simmat_loss, pos)
simmat_loss = tf.multiply(simmat_loss, group_mask_weight)
simmat_loss = tf.reduce_mean(simmat_loss)
相关推荐
2023年江西萍乡中考道德与法治真题及答案.doc
2012年重庆南川中考生物真题及答案.doc
2013年江西师范大学地理学综合及文艺理论基础考研真题.doc
2020年四川甘孜小升初语文真题及答案I卷.doc
2020年注册岩土工程师专业基础考试真题及答案.doc
2023-2024学年福建省厦门市九年级上学期数学月考试题及答案.doc
2021-2022学年辽宁省沈阳市大东区九年级上学期语文期末试题及答案.doc
2022-2023学年北京东城区初三第一学期物理期末试卷及答案.doc
2018上半年江西教师资格初中地理学科知识与教学能力真题及答案.doc
2012年河北国家公务员申论考试真题及答案-省级.doc
2020-2021学年江苏省扬州市江都区邵樊片九年级上学期数学第一次质量检测试题及答案.doc
2022下半年黑龙江教师资格证中学综合素质真题及答案.doc
