def read_one_datum(fqueue, dim, key_num=36, hm_dim=64, nStack=4):
reader = tf.TFRecordReader()
key, value = reader.read(fqueue)
basics = tf.parse_single_example(
value,
features={
# 'key2d': tf.FixedLenFeature([key_num * 2], tf.float32),
'key2d': tf.FixedLenFeature([key_num * 2], tf.float32),
'image': tf.FixedLenFeature([], tf.string)
})
image = basics['image']
image = tf.decode_raw(image, tf.uint8)
image.set_shape([3 * dim * dim])
image = tf.reshape(image, [dim, dim, 3])
# appy resizing and hm generation
# image = tf.image.resize_images(image, size=[256, 256]) # on need.... stick with 64 * 64... instead of 256 by 256, try...
# print("img_shape sanity check",image)
key = basics['key2d']
key_resh = tf.reshape(key, shape=(key_num, 2))
key_resh *= hm_dim # to match hm_dimension
key_hm = ut._tf_generate_hm(
hm_dim, hm_dim, key_resh) # tensor, (64, 64, 36), dtype=tf.float32
temp = [key_hm for _ in range(nStack)]
key_hm = tf.stack(temp, axis=0)
# print("sanity check : key_hm", key_hm)
# print("sanity check - image", image)
return image, key_hm
def read_one_datum(fqueue, dim, key_num=36, hm_dim=64, nStack=4):
reader = tf.TFRecordReader()
key, value = reader.read(fqueue)
basics = tf.parse_single_example(value,
features={
'key2d':
tf.FixedLenFeature([key_num * 2],
tf.float32),
'key3d':
tf.FixedLenFeature([key_num * 3],
tf.float32),
'image':
tf.FixedLenFeature([], tf.string)
})
image = basics['image']
image = tf.decode_raw(image, tf.uint8)
image.set_shape([3 * dim * dim])
image = tf.reshape(image, [dim, dim, 3])
k2d = basics['key2d']
k3d = basics['key3d']
# convert k2d to HMs(4stacks for sHG)
key_resh = tf.reshape(k2d, shape=(key_num, 2))
key_resh *= hm_dim # to match hm_dimension
key_hm = ut._tf_generate_hm(
hm_dim, hm_dim, key_resh) # tensor, (64, 64, 36), dtype=tf.float32
temp = [key_hm for _ in range(nStack)]
key_hm = tf.stack(temp, axis=0)
return image, key_hm, k3d
def read_one_datum(fqueue, dim, key_num=36, hm_dim=64, nStack=4):
reader = tf.TFRecordReader()
key, value = reader.read(fqueue)
basics = tf.parse_single_example(value, features={
# 'key2d': tf.FixedLenFeature([key_num * 2], tf.float32),
'key2d': tf.FixedLenFeature([key_num * 2], tf.float32),
'image': tf.FixedLenFeature([], tf.string)})
image = basics['image']
image = tf.decode_raw(image, tf.uint8)
image.set_shape([3 * dim * dim])
image = tf.reshape(image, [dim, dim, 3])
# appy resizing and hm generation
image = tf.image.resize_images(image, size=[256, 256])
key = basics['key2d']
key_resh = tf.reshape(key*hm_dim, shape=(key_num, 2))
key_resh *= hm_dim
key_hm = ut._tf_generate_hm(hm_dim, hm_dim, key_resh)
temp = [key_hm for _ in range(nStack)]
key_hm = tf.stack(temp, axis=0)
# print(key_hm.get_shape().as_list())
return image, key_hm