def get_images_for_image_ids(self, image_ids):
""" returning a list because """
attention_map_file_paths = [
to_image_path_by_id(self.prefix, image_id, file_ending="bbx")
for image_id in image_ids
]
attention_maps = [
load_numpy_from(self.directory_path, image_file_path)
for image_file_path in attention_map_file_paths
]
attention_labels_file_paths = [
to_image_path_by_id(self.prefix, image_id, file_ending="lbx")
for image_id in image_ids
]
attention_labels = [
load_numpy_from(self.directory_path, image_file_path)
for image_file_path in attention_labels_file_paths
]
attention_ids_file_paths = [
to_image_path_by_id(self.prefix, image_id, file_ending="ibx")
for image_id in image_ids
]
attention_ids = [
load_numpy_from(self.directory_path, image_file_path)
for image_file_path in attention_ids_file_paths
]
""" we cannot simply cast to numpy array here b.c. there might be a different amount of attention maps per image """
""" what to do exactly (flatten, concatenate, sum) depends on the caller """
return attention_maps, attention_labels, attention_ids
def categorical_knearest_neighbors_from_model_dir(model_dir, categories, k=10):
word_sequence = load_json_from(model_dir, "word_sequence.json")
word_embeddings = load_numpy_from(model_dir, "word_embeddings.npy")
categories_with_neighbors = categorical_knearest_neighbors(
categories, word_sequence, word_embeddings, k)
store_json_to(categories_with_neighbors, model_dir,
"category_neighbors.json")
def load_attention_map_by_image_id_from_many(image_ids, directory_path, image_prefix, split_name=None):
"""
Returns a dict of {image_id : feature_map} for each given image_id.
"""
if split_name:
directory_path = "/".join([directory_path, split_name])
return dict([(image_id, load_numpy_from(to_image_path_by_id(image_prefix, image_id, directory_path, file_ending="bbx")))
for image_id in image_ids])
def get_images_for_image_ids(self, image_ids):
"""
@param captions: the list of captions as dicts of { "caption", "image_id", "id" }
@return: the image in the same order as captions
"""
feature_map_file_paths = [
to_image_path_by_id(self.prefix, image_id, file_ending="npy")
for image_id in image_ids
]
feature_maps = np.array([
load_numpy_from(self.directory_path, image_file_path)
for image_file_path in feature_map_file_paths
])
return feature_maps
def load_feature_maps_by_image_id_from_single(directory_path_or_file,
split_name=None,
flat=True):
lookup_filename = DEFAULT_FEATURES_FILE_NAME
if split_name and not flat:
directory_path_or_file = "/".join([directory_path_or_file, split_name])
if split_name and flat:
lookup_filename = "shatt_feature_maps_{}.json".format(split_name)
data = load_numpy_from(directory_path_or_file, lookup_filename)
image_ids = data["id"]
feature_maps = data["features"]
return dict(zip(image_ids, feature_maps))
def _get_image_for_image_id(self, image_id):
""" Overwritten to determine the file ending """
image_file_path = to_image_path_by_id(self.prefix,
image_id,
file_ending="bbx")
return load_numpy_from(self.directory_path, image_file_path)