Run the model on scans from ScanNet

[deli4iled]

Hi, I'd like to run the inference on ScanNet dataset, but I can't figure out how to generate the tfrecords. Could you please share the code to generate them on ScanNet dataset? Thanks

[KyriaAnnwyn]

@deli4iled Hi/ have you figured out how to generate tfrecord files for ScanComplete running? I've tried to train this Net using different dataset, but Have some error in my tfrecord files:

[KyriaAnnwyn]

@deli4iled @angeladai
I've tried to set batch-size to 1 to check that shapes are ok, but I still get the error, though I've checked that all the data I'm writing to tfrecorf has 16384 sizes.
Can't find where this multiplication by 8 comes from, when I use batch_size = 1

[KyriaAnnwyn]

I've found the error in my code for tfrecords generation it was in using np array of ints to fill bytes feature in the record. ^)

[livc]

@KyriaAnnwyn Hi, I'm working on this net with my own data, but I can't find the way to transfer my own data to the tfrecords, could you please share the code? Thanks.

[KyriaAnnwyn]

@livc Here is my code for tfrecords creation:
`
import numpy as np
import tensorflow as tf
import os
import argparse
import sys
import random

_RESOLUTIONS = ['5cm', '9cm', '19cm']
_INPUT_FEATURE = 'input_sdf'
_TARGET_FEATURE = 'target_df'
_TARGET_SEM_FEATURE = 'target_sem'
_HEIGHT_JITTER = [5, 3, 0]
hierarchy_level = 3

def float_feature(value):
"""Wrapper for inserting float features into Example proto."""
if not isinstance(value, (tuple, list)):
#value = [value]
print(value.shape)
return tf.train.Feature(float_list=tf.train.FloatList(value=value))

def int64_feature(value):
"""Wrapper for inserting int64 features into Example proto."""
if not isinstance(value, (tuple, list)):
value = [value]
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))

def bytes_feature(value):
"""Wrapper for inserting bytes features into Example proto."""
if not isinstance(value, (tuple, list)):
value = [value]
return tf.train.Feature(bytes_list=tf.train.BytesList(value=value))

def createTFRecord(args, pair, number):

key_input = RESOLUTIONS[hierarchy_level - 1] + '' + _INPUT_FEATURE
key_target = RESOLUTIONS[hierarchy_level - 1] + '' + _TARGET_FEATURE
key_target_sem = RESOLUTIONS[hierarchy_level - 1] + '' + _TARGET_SEM_FEATURE

#read full and hole tsdfs ans save then to example
insample = np.load(args.infolder + '/' + pair[0])
target = np.load(args.infolder + '/' + pair[1])
print(args.infolder + '/' + pair[0], args.infolder + '/' + pair[1])

center = [insample.shape[0]/2, insample.shape[1]/2, insample.shape[2]/2]
print(center)

#cut areas with specific size fron the input tsdfs, consider that the model shoul be somewhere in the center
target2write = []
insample2write = []
tfrecords_filename = args.outfolder + '/' + 'smpl_scancomplete_train_' + str(number) + '.tfrecords'
writer = tf.python_io.TFRecordWriter(tfrecords_filename)

for numsample in xrange(10*args.batch_size):
move = [random.randint(-(center[0] - args.dim_block / 2),center[0] - args.dim_block / 2),
random.randint(-(center[1] - args.height_block / 2), center[1] - args.height_block / 2),
random.randint(-(center[2] - args.dim_block / 2), center[2] - args.dim_block / 2)]

newCenter = center + move
insample_small = insample[(newCenter[0] - args.dim_block / 2):(newCenter[0] + args.dim_block / 2),
  (newCenter[1] - args.height_block / 2):(newCenter[1] + args.height_block / 2),
  (newCenter[2] - args.dim_block / 2):(newCenter[2] + args.dim_block / 2)]
target_small = target[(newCenter[0] - args.dim_block / 2):(newCenter[0] + args.dim_block / 2),
  (newCenter[1] - args.height_block / 2):(newCenter[1] + args.height_block / 2),
  (newCenter[2] - args.dim_block / 2):(newCenter[2] + args.dim_block / 2)]
#sem_dummy = np.ones(args.dim_block*args.dim_block*args.height_block, dtype = int)
#target2write.append(insample_small.ravel())
#insample2write.append(target_small.ravel())
#print(insample_small.shape)
#print(insample_small.ravel().shape)


#insample2write = np.array(insample2write).ravel()
#target2write = np.array(target2write).ravel()
#print(insample2write.shape)

insample_small = np.array(insample_small.ravel())
target_small = np.array(target_small.ravel())
semantics = np.ones(insample_small.size, dtype = np.int8)

print('Sizes:')
print(insample_small.size, target_small.size, semantics.size)

print('Build example')
example = tf.train.Example(features=tf.train.Features(feature={
      key_target: float_feature(target_small),
      key_target_sem: bytes_feature(semantics.tostring()),
      #key_target_sem: bytes_feature(semantics),
      key_input: float_feature(insample_small)
        }))
writer.write(example.SerializeToString())

writer.close()

def convertList(args):
substr = "_holes"
tsdfNames = [f for f in os.listdir(args.infolder) if substr in f]
name_pairs = []
number = 0

for name in tsdfNames:
targetN = name.replace(substr,'')
#targetN.replace(substr,'')
pair = (name, targetN)
name_pairs.append(pair)

for pair in name_pairs:
createTFRecord(args, pair, number)
number = number + 1

def parse_arguements(argv):
parser = argparse.ArgumentParser()

parser.add_argument('--infolder', type=str, help='Folder with input *.npy tsdfs',
default='../../TSDFBodyModelsBHNPY')
parser.add_argument('--outfolder', type=str, help='Folder with output tfrecords',
default='../../TSDFBodyModelsBHTFRecord')
#parser.add_argument('--pattern', type = str, help='File pattern for *.npy tsdfs', default='')
parser.add_argument('--batch_size', type=int, help='Number of blocks in each batch.', default=8)
parser.add_argument('--dim_block', type=int, help='Input/target block x/z dim.', default=32)
parser.add_argument('--height_block', type=int, help='Input/target block y dim.', default=16)

return parser.parse_args(argv)

if name == 'main':
convertList(parse_arguements(sys.argv[1:]))
`

[livc]

@KyriaAnnwyn Thank you!

[euzer]

thank you @KyriaAnnwyn for sharing this code. I got a question about how can i get numpy array of tsdfs ?
thanks !

[qiji77]

hello，
I have the same problem too,
Could you please tell me how to use my own dataset?(obj or images?)

[santhu937]

@qiji77
hi, did u get the solution to generate tfrecords from obj file?
I need to test on my own data.

[santhu937]

@KyriaAnnwyn can u please share the code file to convert obj file to tfrecords because I'm getting lot of errors with the above code.