def create_test_data(data_dir, data_types):
"""performs moving samples from several datasets to test dataset"""
destination_data_type = "test2017"
destination_dir = get_imgs_dir_filename(data_dir, destination_data_type)
# create test img dir if does not exist
if not os.path.exists(destination_dir):
os.makedirs(destination_dir)
destination_json_filename = get_annotations_json_filename(
data_dir, destination_data_type, prefix="cleaned")
destination_samples = {"imgs": [], "anns": []}
for data_type in data_types.values():
json_filename = get_annotations_json_filename(data_dir,
data_type,
prefix="cleaned")
splitted_samples = get_splitted_samples(json_filename)
destination_samples["imgs"].extend(
splitted_samples["destination"]["imgs"])
destination_samples["anns"].extend(
splitted_samples["destination"]["anns"])
json_data = get_json_data(splitted_samples["origin"])
save_json(json_data, json_filename)
origin_dir = get_imgs_dir_filename(data_dir, data_type)
move_images(origin_dir, destination_dir,
splitted_samples["destination"]["imgs"])
json_data = get_json_data(destination_samples)
save_json(json_data, destination_json_filename)
def create_cleaned_annotations(data_dir, data_types):
"""perform cleaning and saving of annotations"""
for data_type in data_types.values():
old_json_filename = get_annotations_json_filename(data_dir, data_type)
imgs_dir = get_imgs_dir_filename(data_dir, data_type)
edited_json = get_edited_json(old_json_filename, imgs_dir)
new_json_filename = get_annotations_json_filename(
data_dir, data_type, "cleaned")
with open(new_json_filename, "w") as file:
json.dump(edited_json, file)
def check_dataset(data_dir, data_type, captions_per_image):
"""check dataset for correctness"""
hdf5_filename = get_hdf5_filename(data_dir, data_type)
json_filename = get_annotations_json_filename(data_dir,
data_type,
prefix="cleaned")
coco = COCO(json_filename)
img_ids = coco.getImgIds()
print(len(img_ids))
captions_json_filename = get_captions_json_filename(data_dir,
data_type,
prefix="cleaned")
with open(captions_json_filename) as cap_file:
captions = json.load(cap_file)["anns"]
random_indices = np.random.choice(range(len(captions)), 5)
rand_caps = [captions[idx] for idx in random_indices]
random_img_indices = [idx // captions_per_image for idx in random_indices]
with h5py.File(hdf5_filename, "r") as hdf5_file:
images = hdf5_file["images"]
print(len(images))
rand_images = [images[idx] for idx in random_img_indices]
for i, img in enumerate(rand_images):
img = img.transpose((1, 2, 0))
imgplot = plt.imshow(img)
plt.title(rand_caps[i]["caption"])
plt.show()
def create_image_hdf5_datasets(data_dir, data_types, captions_per_image):
"""create hdf5 dataset for image directory"""
for data_type in data_types.values():
json_filename = get_annotations_json_filename(data_dir,
data_type,
prefix="cleaned")
coco = COCO(json_filename)
img_ids = coco.getImgIds()
hdf5_filename = get_hdf5_filename(data_dir, data_type)
with h5py.File(hdf5_filename, "a") as hdf5_file:
# Make a note of the number of captions we are sampling per image
hdf5_file.attrs["captions_per_image"] = captions_per_image
# Create dataset inside HDF5 file to store images
dataset_shape = (len(img_ids), 3, 256, 256)
images = hdf5_file.create_dataset("images",
shape=dataset_shape,
dtype="uint8",
compression="gzip",
compression_opts=9)
for i, img_id in enumerate(img_ids):
img_filename = coco.imgs[img_id]["file_name"]
impath = get_img_filename(data_dir, data_type, img_filename)
img = get_transformed_img(impath)
# Save image to HDF5 file
images[i] = img
if i % 1000 == 0:
print("{}-th iteration is written".format(i))
print("{} is written".format(hdf5_filename))
def create_trimmed_annotations(data_dir, data_types, captions_per_image):
"""limit captions per image and remove images from json filename"""
for data_type in data_types.values():
json_filename = get_annotations_json_filename(data_dir,
data_type,
prefix="cleaned")
samples = get_list_trimmed_annotations(json_filename,
captions_per_image)
json_data = get_json_data(samples)
save_json(json_data, json_filename)
def get_image_paths(data_dir, data_type):
"""extract paths for images in right order"""
json_filename = get_annotations_json_filename(data_dir,
data_type,
prefix="cleaned")
coco = COCO(json_filename)
img_ids = coco.getImgIds()
img_paths = []
for img_id in img_ids:
img_filename = coco.imgs[img_id]["file_name"]
impath = get_img_filename(data_dir, data_type, img_filename)
img_paths.append(impath)
return img_paths
def create_only_captions_json(data_dir, data_types):
"""create json that only contains info about captions"""
for data_type in data_types.values():
ann_json_filename = get_annotations_json_filename(data_dir,
data_type,
prefix="cleaned")
coco = COCO(ann_json_filename)
json_data = {}
json_data["anns"] = list(coco.anns.values())
caption_json_filename = get_captions_json_filename(data_dir,
data_type,
prefix="cleaned")
save_json(json_data, caption_json_filename)
def create_image_hdf5_datasets_in_parallel_mpi(data_dir, data_types,
captions_per_image):
"""create hdf5 dataset for image directory"""
comm = MPI.COMM_WORLD
num_processes = comm.size
# The process ID (integer 0-3 for 4-process run)
rank = comm.rank
for data_type in data_types.values():
json_filename = get_annotations_json_filename(data_dir,
data_type,
prefix="cleaned")
coco = COCO(json_filename)
img_ids = coco.getImgIds()
hdf5_filename = get_hdf5_filename(data_dir, data_type)
with h5py.File(hdf5_filename, "a", driver='mpio',
comm=comm) as hdf5_file:
# Make a note of the number of captions we are sampling per image
hdf5_file.attrs["captions_per_image"] = captions_per_image
# Create dataset inside HDF5 file to store images
dataset_shape = (len(img_ids), 3, 256, 256)
images = hdf5_file.create_dataset("images",
shape=dataset_shape,
dtype="uint8")
for i, img_id in enumerate(img_ids):
if i % num_processes == rank:
img_filename = coco.imgs[img_id]["file_name"]
impath = get_img_filename(data_dir, data_type,
img_filename)
img = get_transformed_img(impath)
# Save image to HDF5 file
images[i] = img
if i % 1000 == 0:
print("{}-th iteration is written".format(i))
if rank == 0:
print("{} is written".format(hdf5_filename))