def __init__(self, assignment: ClusterAssignment, split: str):
self.config = assignment.config
cluster_assignments = assignment.cluster_assignments[split]
self.cluster_to_image = self._to_cluster_to_image_map(
cluster_assignments)
self.dataset = build_dataset(self.config, split)
self.data_source = self.dataset.data_objs[0]
def save_cluster_assignment_as_dataset(cls, output_dir: str,
assignments: ClusterAssignment):
"""
Create a 'disk_filelist' dataset out of the cluster assignments:
- the inputs are the images
- the labels are the index of the cluster assigned to the image
"""
os.makedirs(output_dir, exist_ok=True)
for split in assignments.cluster_assignments.keys():
dataset = build_dataset(assignments.config, split)
image_paths = dataset.get_image_paths()
assert len(image_paths) == 1, "Multi-dataset not supported yet!"
image_paths = image_paths[0]
image_labels = []
for image_id in range(len(image_paths)):
image_labels.append(
assignments.cluster_assignments[split][image_id])
images_file_path = os.path.join(output_dir,
f"{split.lower()}_images.npy")
labels_file_path = os.path.join(output_dir,
f"{split.lower()}_labels.npy")
np.save(images_file_path, np.array(image_paths))
np.save(labels_file_path, np.array(image_labels))
def benchmark_data(cfg: AttrDict, split: str = "train"):
split = split.upper()
total_images = MAX_ITERS * cfg["DATA"][split]["BATCHSIZE_PER_REPLICA"]
timer = Timer()
dataset = build_dataset(cfg, split)
try:
device = torch.device("cuda" if cfg.MACHINE.DEVICE == "gpu" else "cpu")
except AttributeError:
device = torch.device("cuda")
# Gives sampler same seed for entire distributed group as per pytorch documentation.
sampler_seed = cfg.SEED_VALUE
dataloader = get_loader(
dataset=dataset,
dataset_config=cfg["DATA"][split],
num_dataloader_workers=cfg.DATA.NUM_DATALOADER_WORKERS,
pin_memory=False,
multi_processing_method=cfg.MULTI_PROCESSING_METHOD,
device=device,
sampler_seed=sampler_seed,
)
# Fairstore data sampler would require setting the start iter before it can start.
if hasattr(dataloader.sampler, "set_start_iter"):
dataloader.sampler.set_start_iter(0)
# initial warmup measured as warmup time
timer.reset()
data_iterator = iter(dataloader)
for i in range(10): # warmup
next(data_iterator)
if i == 0:
# the total number of seconds since the start/reset of the timer
warmup_time = timer.seconds()
logging.info(f"Warmup time {WARMUP_ITERS} batches: {warmup_time} seconds")
# measure the number of images per sec in 1000 iterations.
timer = Timer()
for _ in tqdm.trange(MAX_ITERS):
next(data_iterator)
time_elapsed = timer.seconds()
logging.info(
f"iters: {MAX_ITERS}; images: {total_images}; time: {time_elapsed} seconds; "
f"images/sec: {round(float(total_images / time_elapsed), 4)}; "
f"ms/img: {round(float(1000 * time_elapsed / total_images), 4)} ")
# run benchmark for a few more rounds to catch fluctuations
for round_idx in range(BENCHMARK_ROUNDS):
timer = Timer()
for _ in tqdm.trange(MAX_ITERS):
next(data_iterator)
time_elapsed = timer.seconds()
logging.info(
f"round: {round_idx}: iters: {MAX_ITERS}; images: {total_images}; "
f"time: {time_elapsed} seconds; "
f"images/sec: {round(float(total_images / time_elapsed), 4)}; "
f"ms/img: {round(float(1000 * time_elapsed / total_images), 4)} ")
del data_iterator
del dataloader
def get_image_paths(cfg: AttrDict, split: str) -> List[str]:
"""
Get the list of image path for the provided dataset and split
"""
dataset = build_dataset(cfg=cfg, split=split)
feature_image_paths = dataset.get_image_paths()
# due to multi-modality, we get image_paths as a nested list, one for each
# dataset. Check it's a list and extract images.
assert type(feature_image_paths) == list, "Image paths must be a list"
assert len(feature_image_paths) == 1, "Multi-modality not supported yet!"
return feature_image_paths[0]
def build_datasets(self):
"""
Get the datasets for the data splits we will use in the training. The
set_available_splits variable determines the splits used in the training.
"""
datasets, data_and_label_keys = {}, {}
for split in self.available_splits:
datasets[split] = build_dataset(self.config, split)
data_and_label_keys["input"] = self.config.DATA[split].INPUT_KEY_NAMES
data_and_label_keys["target"] = self.config.DATA[split].TARGET_KEY_NAMES
return datasets, data_and_label_keys
def get_data_features_and_images(cfg: AttrDict):
output_dir = get_checkpoint_folder(cfg)
split = cfg.RANKING.FEATURES.DATA_PARTITION
logging.info("Merging features...")
# merge the features across all nodes/gpus into one
feature_data = merge_features(output_dir, split.lower(),
cfg.RANKING.FEATURES.LAYER_NAME)
logging.info("Getting the image paths...")
# get the list of image Ids
dataset = build_dataset(cfg=cfg, split=split)
feature_image_paths = dataset.get_image_paths()
# due to multi-modality, we get image_paths as a nested list, one for each
# dataset. Check it's a list and extract images.
assert type(feature_image_paths) == list, "Image paths must be a list"
assert len(feature_image_paths) == 1, "Multi-modality not supported yet!"
return feature_data, feature_image_paths[0]
