def sample_tiles_from_candidates(self, tiles_samplers):
"""
Apply a sampler over each satellite image's candidate tiles to generate a list of tiles (= regions of interest)
Args:
tiles_samplers(list[TilesSampler]|TilesSampler):
Returns:
in place. assign self.sampled_tiles
"""
sampled_tiles = []
if not isinstance(tiles_samplers, (list, tuple)):
tiles_samplers = [tiles_samplers]
self.tiles_samplers = tiles_samplers
LOGGER.info("Sampling tiles")
for tile_sample in tqdm(self.tiles_samplers, desc="Sampling tiles"):
sampled_tiles.extend(
tile_sample.sample_tiles_from_candidates(self.candidate_tiles))
LOGGER.info(
"Tiles sampled, now generate the dataset using Dataset.generate_tiles_dataset"
)
self.sampled_tiles = sampled_tiles
def items_dataset_from_path(path: Optional[str] = None) -> [SatelliteImage]:
"""
Get a list of Satellite Images items from path
Args:
path: folder where to look
Returns:
"""
assert path is not None, "Please set folder variable, likely ${TP_DATA}/raw/trainval/"
LOGGER.info("Looking in {}".format(path))
items = []
list_images = glob.glob(os.path.join(path, "*.jpg"))
def _parse_image(image_file):
image_id = os.path.splitext(os.path.basename(image_file))[0]
item = SatelliteImage.from_image_id_and_path(image_id, path)
# Read the when initialising to put data into cache
# LOGGER.info("Found item {}".format(item.image_id))
assert isinstance(item.image, np.ndarray)
assert isinstance(item.labels, list)
return item
dataset = Dataset(items=list_images)
dataset = dataset.map(_parse_image, n_jobs=8, desc="Parsing items")
dataset.sorted(key=lambda item: item.key)
LOGGER.info("Found {} items".format(len(dataset)))
return dataset
def list_items_from_path(cls, path=None):
"""
Get a list of Satellite Images items from path
Args:
path: folder where to look
Returns:
list(SatelliteImageItem):
"""
assert path is not None, "Please set folder variable, likely ${TP_ISAE_DATA}/raw/trainval/"
LOGGER.info("Looking in {}".format(path))
items = []
list_images = glob.glob(os.path.join(path, "*.jpg"))
for image_file in list_images:
image_id = os.path.splitext(os.path.basename(image_file))[0]
item = SatelliteImage.from_image_id_and_path(image_id, path=path)
# Read the when initialising to put data into cache
assert isinstance(item.image, np.ndarray)
assert isinstance(item.labels, list)
items.append(item)
items = list(sorted(items, key=lambda item: item.key))
return items
def generate_candidates_tiles(self, sliding_windows):
"""
Apply a sliding window over each satellite image to generate a list of tiles (= regions of interest) to sample from
Args:
sliding_windows(list[SlidingWindow]|SlidingWindow):
Returns:
in place (assign self.candidate_tiles)
"""
if not isinstance(sliding_windows, (list, tuple)):
sliding_windows = [sliding_windows]
self.sliding_windows = sliding_windows
sliding_windows = self.sliding_windows
items = self.items
LOGGER.info("Generating a pool of candidates tiles")
candidate_tiles = []
for sliding_window in tqdm(sliding_windows,
position=0,
desc="Applying slider"):
for item in tqdm(items, position=1, desc="On item"):
candidate_tiles.extend(sliding_window.get_tiles_for_item(item))
LOGGER.info(
"Candidates tiles generated ! Now sample them using Dataset.sample_tiles_from_candidates"
)
self.candidate_tiles = list(set(candidate_tiles))
self.found_labels = list_utils.get_labels_in_list(self.candidate_tiles)
# Initialise sampled tiles by default (copy candidate tiles)
self.sampled_tiles = self.candidate_tiles[:]
def download_eval_data(data_dir=None):
"""
Download the raw eval data to data dir and extracts
Args:
data_dir:
Returns:
"""
data_dir = data_dir or os.path.expandvars(os.environ.get("TP_DATA"))
LOGGER.info("Downloading evaluation data")
_download_data(archive="tp_isae_eval_data.tar.gz", data_dir=data_dir, check_dir="raw/eval")
LOGGER.info("Done. Your data is located here {}\n".format(os.path.join(data_dir, "raw", "eval")))
def predict_on_item(item, predictor=None, sliding_windows=None):
"""
Args:
item(SatelliteImage): the item on which to apply the prediction
predictor(Predictor): A Predictor object that encapsulates our model
sliding_windows(SlidingWindow): The sliding window used to generate candidates
Returns:
"""
if not isinstance(sliding_windows, (list, tuple)):
sliding_windows = [sliding_windows]
LOGGER.info("Generating tiles to predict")
tiles_to_predict = []
for sliding_window in tqdm(sliding_windows,
position=0,
desc="Applying slider"):
tiles_to_predict.extend(sliding_window.get_tiles_for_item(item))
tiles_to_predict = list(set(tiles_to_predict))
LOGGER.info("Generating predicting on item {} with {} tiles".format(
item.key, len(tiles_to_predict)))
image = item.image
tiles_results = []
if hasattr(predictor, "batch_size") and predictor.batch_size > 1:
batches = [
tiles_to_predict[i:i + predictor.batch_size]
for i in range(0, len(tiles_to_predict), predictor.batch_size)
]
for batch in tqdm(
batches,
desc="Calling .predict_on_batch() with batch_size {}".
format(predictor.batch_size)):
batch_data = [tile.get_data(image) for tile in batch]
batch_results = predictor.predict_on_batch(batch_data)
for i, tile in enumerate(batch):
tiles_results.append(
PredictionTile.from_labelled_tile_and_prediction(
tile, batch_results[i]))
else:
for tile in tqdm(tiles_to_predict,
desc="Calling .predict() with one tile"):
prediction = predictor.predict(tile.get_data(image))
tiles_results.append(
PredictionTile.from_labelled_tile_and_prediction(
tile, prediction))
return tiles_results
def predict_on_item(self, item):
"""
Args:
item(SatelliteImage): the item on which to apply the prediction
Returns:
"""
LOGGER.info("Generating tiles to predict")
item_dataset = Dataset(items=[item])
tiles = Dataset(items=[])
for sliding_window in tqdm(self.sliding_windows,
position=0,
desc="Applying slider"):
tiles = tiles.extend(item_dataset.flatmap(sliding_window))
tiles = tiles.apply(lambda items: list(set(items)))
LOGGER.info("Generating predicting on item {} with {} tiles".format(
item.key, len(tiles)))
image = item.image
def _batch(items):
return [
items[i:i + self.predictor.batch_size]
for i in range(0, len(items), self.predictor.batch_size)
]
batches = tiles.apply(_batch)
print(len(tiles))
print(len(batches))
def _predict(batch):
batch_data = list(map(lambda tile: tile.get_data(image), batch))
batch_results = self.predictor.predict_on_batch(batch_data)
batch_results = list(
map(
lambda tpl: PredictionTile.
from_labelled_tile_and_prediction(tpl[0], tpl[1]),
zip(batch, batch_results)))
return batch_results
tiles_results = batches.flatmap(_predict, desc="Predicting on batch")
return tiles_results
def generate_candidate_tiles_from_items(items_dataset: Dataset,
sliding_windows: [SlidingWindow],
n_jobs: int = 1) -> Dataset:
"""
High level helper function
Apply a sliding window over each satellite image
to generate a list of tiles (= regions of interest) to sample from
Args:
sliding_windows(list[SlidingWindow]|SlidingWindow):
items_dataset(Dataset):
n_jobs(int):
Returns:
dataset (Dataset):
"""
LOGGER.info("Generating a pool of candidates tiles")
tiles_dataset = Dataset(items=[])
if not isinstance(sliding_windows, (list, tuple)):
sliding_windows = [sliding_windows]
for sliding_window in sliding_windows:
LOGGER.info(sliding_window)
tiles_dataset = tiles_dataset.extend(
items_dataset.flatmap(sliding_window,
desc="Applying sliding window",
n_jobs=n_jobs))
tiles_dataset = tiles_dataset.apply(lambda items: list(set(items)))
LOGGER.info("State of dataset")
LOGGER.info(roi_list_utils.get_state(tiles_dataset.items))
return tiles_dataset
def _dump_tiles(item):
LOGGER.info("Dumping for item {}".format(item.key))
if output_dir is not None:
tiles_dumper = ImageItemTileDumper(item,
output_dir=output_dir,
save_format=save_format)
else:
tiles_dumper = NpArrayTileDumper(item)
tiles_dataset_ = tiles_dataset.filter(
lambda tile: tile.item_id == item.key, desc="Filtering")
tiles_dataset_ = tiles_dataset_.map(
tiles_dumper, desc="Saving tiles to {}".format(output_dir))
return tiles_dataset_.items
def sample_tiles_from_candidates(tiles_dataset: Dataset,
tiles_samplers: [TilesSampler]) -> Dataset:
"""
High level helper function
Apply a sampler over each satellite image's candidate tiles
to generate a list of tiles (= regions of interest)
Args:
tiles_samplers(list[TilesSampler]|TilesSampler):
tiles_dataset(Dataset)
Returns:
sampled_dataset(Dataset)
"""
sampled_dataset = Dataset(items=[])
LOGGER.info("Sampling tiles")
if not isinstance(tiles_samplers, (list, tuple)):
tiles_samplers = [tiles_samplers]
for tiles_sampler in tiles_samplers:
LOGGER.info(tiles_sampler)
sampled_dataset = sampled_dataset.extend(
tiles_dataset.apply(tiles_sampler))
LOGGER.info(
"Tiles sampled, now generate the dataset using Dataset.generate_tiles_dataset"
)
LOGGER.info(roi_list_utils.get_state(sampled_dataset.items))
return sampled_dataset
def download_train_data(data_dir=None):
"""
Download the raw training data to data dir and extracts
Args:
data_dir:
Returns:
"""
data_dir = data_dir or os.environ.get("TP_ISAE_DATA")
LOGGER.info("Downloading training data")
_download_data(archive="tp_isae_train_data.tar.gz", data_dir=data_dir)
LOGGER.info("Done. Your training data is located here {}".format(
os.path.join(data_dir, "raw", "trainval")))
def _download_test_ci_data(data_dir=None):
"""
Reserved for CI
Args:
data_dir:
Returns:
"""
data_dir = data_dir or os.environ.get("TP_ISAE_DATA")
LOGGER.info("Downloading test ci data")
_download_data(archive="tp_isae_test_ci.tar.gz", data_dir=data_dir)
LOGGER.info(
"Done. Your data is located here {}\nYour eval data is located here {}"
.format(os.path.join(data_dir, "raw"),
os.path.join(data_dir, "raw", "eval")))
def sample_tiles_from_candidates(self, candidate_tiles):
"""
Apply the sampling logic of this class to a list of `candidates`
Args:
candidate_tiles(list[Tiles]): List of regions of interest to apply the sampler on
Returns:
list[Tiles]: Sampled list
"""
LOGGER.info("Sampling")
if self.target_label is not None:
candidate_tiles = roi_list_utils.filter_tiles_by_label(candidate_tiles, self.target_label)
nb_tiles_max = self.nb_tiles_max or len(candidate_tiles)
return self._sample_n_tiles_from_list(candidate_tiles, nb_tiles_max)
def _download_data(archive="tp_isae_data.tar.gz", data_dir=None, check_dir=None):
"""
Args:
archive:
data_dir:
Returns:
"""
assert data_dir is not None, "please specify a download dir or better specify TP_DATA env variable"
if not os.path.exists(data_dir):
os.makedirs(data_dir)
LOGGER.info("Downloading data from {} to {}".format(archive, data_dir))
if not os.path.exists(os.path.join(data_dir, archive)):
# Download tar gz
LOGGER.info("Downloading {}".format("{}/{}").format(ROOT_URL, archive))
cmd = ["curl", "-X", "GET", "{}/{}".format(ROOT_URL, archive), "--output", os.path.join(data_dir, archive)]
subprocess.check_call(cmd)
if check_dir is None or not os.path.exists(os.path.join(data_dir, check_dir)):
# Untar it
LOGGER.info("Extracting tar gz")
cmd = ["tar", "-zxvf", os.path.join(data_dir, archive), "-C", data_dir]
subprocess.check_call(cmd)
def dump_dataset_tiles(tiles_dataset: Dataset,
items_dataset: Dataset,
output_dir=None,
remove_first=False,
save_format="jpg") -> Dataset:
"""
High level helper function
Actually generates training images from the dataset.sampled_tiles (= regions of interest)
The filestructure is compatible with keras.ImageDataGenerator.flow_from_directory() method
For more information on how to parse this, check this script:
https://gist.github.com/fchollet/0830affa1f7f19fd47b06d4cf89ed44d
In summary, this is our directory structure:
```markdown
output_dir/
aircrafts/
ac001.jpg
ac002.jpg
...
background/
bg001.jpg
bg002.jpg
...
```
Args:
items_dataset (Dataset):
tiles_dataset (Dataset):
output_dir(str): the output path
save_format: "jpg" the image format
remove_first(bool): erase output dir first?
Returns:
"""
LOGGER.info(
"Generating a dataset of tiles at location {}".format(output_dir))
if remove_first and output_dir is not None:
try:
shutil.rmtree(output_dir)
except FileNotFoundError:
pass
def _dump_tiles(item):
LOGGER.info("Dumping for item {}".format(item.key))
if output_dir is not None:
tiles_dumper = ImageItemTileDumper(item,
output_dir=output_dir,
save_format=save_format)
else:
tiles_dumper = NpArrayTileDumper(item)
tiles_dataset_ = tiles_dataset.filter(
lambda tile: tile.item_id == item.key, desc="Filtering")
tiles_dataset_ = tiles_dataset_.map(
tiles_dumper, desc="Saving tiles to {}".format(output_dir))
return tiles_dataset_.items
dumped_tiles = items_dataset.flatmap(_dump_tiles)
return dumped_tiles
def generate_tiles_dataset(self,
output_dir=None,
save_format="jpg",
remove_first=True):
"""
Actually generates training images from the dataset.sampled_tiles (= regions of interest)
The filestructure is compatible with keras.ImageDataGenerator.flow_from_directory() method
For more information on how to parse this, check this script:
https://gist.github.com/fchollet/0830affa1f7f19fd47b06d4cf89ed44d
In summary, this is our directory structure:
```markdown
output_dir/
aircrafts/
ac001.jpg
ac002.jpg
...
background/
bg001.jpg
bg002.jpg
...
```
Args:
output_dir(str): the output path
save_format: "jpg" the image format
remove_first(bool): erase output dir first?
Returns:
"""
LOGGER.info(
"Generating a dataset of tiles at location {}".format(output_dir))
for label in self.found_labels:
if remove_first:
shutil.rmtree(os.path.join(output_dir, label))
if not os.path.exists(os.path.join(output_dir, label)):
os.makedirs(os.path.join(output_dir, label))
def _generate_tiles(item, tiles):
image = item.image
tiles = list_utils.filter_tiles_by_item(tiles, item.key)
for tile in tiles:
tile_data = tile.get_data(image)
tile_label = tile.label
tile_basename = "{}_{}.{}".format(item.key, tile.key,
save_format)
io.imsave(os.path.join(output_dir, tile_label, tile_basename),
tile_data)
items = self.items
sampled_tiles = self.sampled_tiles
LOGGER.info("Dumping tiles to {}".format(output_dir))
for item in tqdm(items, desc="Saving tiles to {}".format(output_dir)):
_generate_tiles(item, sampled_tiles)
