|
import os |
|
import tensorflow as tf |
|
|
|
|
|
|
|
def build_data(split, shuffle=False): |
|
"""Build CLEVR dataset.""" |
|
file_path_base = './' |
|
|
|
|
|
def _parse_tfr_element(element): |
|
|
|
data = { |
|
'raw_image': tf.io.FixedLenFeature([], tf.string), |
|
'mask': tf.io.FixedLenFeature([], tf.string), |
|
'slot': tf.io.VarLenFeature(tf.float32), |
|
} |
|
|
|
content = tf.io.parse_single_example(element, data) |
|
|
|
raw_image = content['raw_image'] |
|
raw_mask = content['mask'] |
|
raw_slot = content['slot'] |
|
|
|
image_ori = tf.io.parse_tensor(raw_image, out_type=tf.uint8) |
|
image = tf.cast(image_ori, tf.float32) |
|
image = ((image / 255.0) - 0.5) * 2.0 |
|
|
|
|
|
raw_mask = tf.io.parse_tensor(raw_mask, out_type=tf.uint8) |
|
mask = tf.cast(raw_mask, tf.float32) |
|
mask = tf.clip_by_value(mask, 0.0, 1.0) |
|
slot = raw_slot.values |
|
return { |
|
"image_ori": image_ori, |
|
"raw_mask": raw_mask, |
|
"image": image, |
|
"mask": mask, |
|
"slot": slot |
|
} |
|
|
|
AUTOTUNE = tf.data.AUTOTUNE |
|
if split == "train": |
|
num_file = 100 |
|
file_path = os.path.join(file_path_base, 'train') |
|
filename = [ |
|
os.path.join(file_path, "train-part-{:0>3}.tfrecord".format(i)) |
|
for i in range(num_file) |
|
] |
|
elif split in ["collision", "blocking", "continuity"]: |
|
num_file = 6 |
|
file_path = os.path.join(file_path_base, "test") |
|
file_path = os.path.join(file_path, split) |
|
filename = [ |
|
os.path.join(file_path, "eval-part-{:0>3}.tfrecord".format(i)) |
|
for i in range(num_file) |
|
] |
|
elif split in ["permanence"]: |
|
num_file = 4 |
|
file_path = os.path.join(file_path_base, "test") |
|
file_path = os.path.join(file_path, split) |
|
filename = [ |
|
os.path.join(file_path, "eval-part-{:0>3}.tfrecord".format(i)) |
|
for i in range(4) |
|
] |
|
elif split == "eval": |
|
num_file = 4 |
|
file_path = os.path.join(file_path_base, "test") |
|
eval_list = ["collision", "blocking", "permanence", "continuity"] |
|
filename = [ |
|
file_path + i + '/' + "eval-part-000.tfrecord" for i in eval_list |
|
] |
|
else: |
|
raise ValueError("Error dataset type") |
|
if shuffle: |
|
filename = tf.data.Dataset.from_tensor_slices(filename) |
|
filename = filename.shuffle(num_file) |
|
ds = filename.interleave( |
|
lambda x: tf.data.TFRecordDataset(x, compression_type="GZIP"), |
|
cycle_length=num_file, |
|
block_length=1) |
|
ds = ds.shuffle(1000) |
|
else: |
|
ds = tf.data.TFRecordDataset(filename, compression_type="GZIP") |
|
ds = ds.map(_parse_tfr_element, num_parallel_calls=AUTOTUNE) |
|
return ds |
|
|
|
|
|
def load_data(batch_size, split, **kwargs): |
|
ds = build_data(split=split, **kwargs) |
|
ds = ds.batch(batch_size, drop_remainder=True) |
|
return ds |
|
|
