Initial commit

ssd/utils/model_zoo.py

+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+import os
+import sys
+
+import torch
+
+from ssd.utils.dist_util import is_main_process, synchronize
+
+from torch.hub import download_url_to_file
+from torch.hub import urlparse
+from torch.hub import HASH_REGEX
+
+
+# very similar to https://github.com/pytorch/pytorch/blob/master/torch/utils/model_zoo.py
+# but with a few improvements and modifications
+def cache_url(url, model_dir=None, progress=True):
+    r"""Loads the Torch serialized object at the given URL.
+    If the object is already present in `model_dir`, it's deserialized and
+    returned. The filename part of the URL should follow the naming convention
+    ``filename-<sha256>.ext`` where ``<sha256>`` is the first eight or more
+    digits of the SHA256 hash of the contents of the file. The hash is used to
+    ensure unique names and to verify the contents of the file.
+    The default value of `model_dir` is ``$TORCH_HOME/models`` where
+    ``$TORCH_HOME`` defaults to ``~/.torch``. The default directory can be
+    overridden with the ``$TORCH_MODEL_ZOO`` environment variable.
+    Args:
+        url (string): URL of the object to download
+        model_dir (string, optional): directory in which to save the object
+        progress (bool, optional): whether or not to display a progress bar to stderr
+    Example:
+        >>> cached_file = maskrcnn_benchmark.utils.model_zoo.cache_url('https://s3.amazonaws.com/pytorch/models/resnet18-5c106cde.pth')
+    """
+    if model_dir is None:
+        torch_home = os.path.expanduser(os.getenv("TORCH_HOME", "~/.torch"))
+        model_dir = os.getenv("TORCH_MODEL_ZOO", os.path.join(torch_home, "models"))
+    if not os.path.exists(model_dir):
+        os.makedirs(model_dir)
+    parts = urlparse(url)
+    filename = os.path.basename(parts.path)
+    if filename == "model_final.pkl":
+        # workaround as pre-trained Caffe2 models from Detectron have all the same filename
+        # so make the full path the filename by replacing / with _
+        filename = parts.path.replace("/", "_")
+    cached_file = os.path.join(model_dir, filename)
+    if not os.path.exists(cached_file) and is_main_process():
+        sys.stderr.write('Downloading: "{}" to {}\n'.format(url, cached_file))
+        hash_prefix = HASH_REGEX.search(filename)
+        if hash_prefix is not None:
+            hash_prefix = hash_prefix.group(1)
+            # workaround: Caffe2 models don't have a hash, but follow the R-50 convention,
+            # which matches the hash PyTorch uses. So we skip the hash matching
+            # if the hash_prefix is less than 6 characters
+            if len(hash_prefix) < 6:
+                hash_prefix = None
+        download_url_to_file(url, cached_file, hash_prefix, progress=progress)
+    synchronize()
+    return cached_file
+
+
+def load_state_dict_from_url(url, map_location='cpu'):
+    cached_file = cache_url(url)
+    return torch.load(cached_file, map_location=map_location)

ssd/utils/nms.py

+import sys
+import warnings
+
+import torch
+import torchvision
+
+if torchvision.__version__ >= '0.0.0':
+    _nms = torchvision.ops.nms
+else:
+    warnings.warn('No NMS is available. Please upgrade torchvision to 0.3.0+')
+    sys.exit(-1)
+
+
+def nms(boxes, scores, nms_thresh):
+    """ Performs non-maximum suppression, run on GPU or CPU according to
+    boxes's device.
+    Args:
+        boxes(Tensor[N, 4]): boxes in (x1, y1, x2, y2) format, use absolute coordinates(or relative coordinates)
+        scores(Tensor[N]): scores
+        nms_thresh(float): thresh
+    Returns:
+        indices kept.
+    """
+    keep = _nms(boxes, scores, nms_thresh)
+    return keep
+
+
+def batched_nms(boxes, scores, idxs, iou_threshold):
+    """
+    Performs non-maximum suppression in a batched fashion.
+
+    Each index value correspond to a category, and NMS
+    will not be applied between elements of different categories.
+
+    Parameters
+    ----------
+    boxes : Tensor[N, 4]
+        boxes where NMS will be performed. They
+        are expected to be in (x1, y1, x2, y2) format
+    scores : Tensor[N]
+        scores for each one of the boxes
+    idxs : Tensor[N]
+        indices of the categories for each one of the boxes.
+    iou_threshold : float
+        discards all overlapping boxes
+        with IoU < iou_threshold
+
+    Returns
+    -------
+    keep : Tensor
+        int64 tensor with the indices of
+        the elements that have been kept by NMS, sorted
+        in decreasing order of scores
+    """
+    if boxes.numel() == 0:
+        return torch.empty((0,), dtype=torch.int64, device=boxes.device)
+    # strategy: in order to perform NMS independently per class.
+    # we add an offset to all the boxes. The offset is dependent
+    # only on the class idx, and is large enough so that boxes
+    # from different classes do not overlap
+    max_coordinate = boxes.max()
+    offsets = idxs.to(boxes) * (max_coordinate + 1)
+    boxes_for_nms = boxes + offsets[:, None]
+    keep = nms(boxes_for_nms, scores, iou_threshold)
+    return keep

ssd/utils/registry.py

+def _register_generic(module_dict, module_name, module):
+    assert module_name not in module_dict
+    module_dict[module_name] = module
+
+
+class Registry(dict):
+    """
+    A helper class for managing registering modules, it extends a dictionary
+    and provides a register functions.
+    Eg. creating a registry:
+        some_registry = Registry({"default": default_module})
+    There're two ways of registering new modules:
+    1): normal way is just calling register function:
+        def foo():
+            ...
+        some_registry.register("foo_module", foo)
+    2): used as decorator when declaring the module:
+        @some_registry.register("foo_module")
+        @some_registry.register("foo_module_nickname")
+        def foo():
+            ...
+    Access of module is just like using a dictionary, eg:
+        f = some_registry["foo_module"]
+    """
+
+    def __init__(self, *args, **kwargs):
+        super(Registry, self).__init__(*args, **kwargs)
+
+    def register(self, module_name, module=None):
+        # used as function call
+        if module is not None:
+            _register_generic(self, module_name, module)
+            return
+
+        # used as decorator
+        def register_fn(fn):
+            _register_generic(self, module_name, fn)
+            return fn
+
+        return register_fn

test.py

+import argparse
+import logging
+import os
+
+import torch
+import torch.utils.data
+
+from ssd.config import cfg
+from ssd.engine.inference import do_evaluation
+from ssd.modeling.detector import build_detection_model
+from ssd.utils import dist_util
+from ssd.utils.checkpoint import CheckPointer
+from ssd.utils.dist_util import synchronize
+from ssd.utils.logger import setup_logger
+
+
+def evaluation(cfg, ckpt, distributed):
+    logger = logging.getLogger("SSD.inference")
+
+    model = build_detection_model(cfg)
+    checkpointer = CheckPointer(model, save_dir=cfg.OUTPUT_DIR, logger=logger)
+    device = torch.device(cfg.MODEL.DEVICE)
+    model.to(device)
+    checkpointer.load(ckpt, use_latest=ckpt is None)
+    do_evaluation(cfg, model, distributed)
+
+
+def main():
+    parser = argparse.ArgumentParser(description='SSD Evaluation on VOC and COCO dataset.')
+    parser.add_argument(
+        "--config-file",
+        default="",
+        metavar="FILE",
+        help="path to config file",
+        type=str,
+    )
+    parser.add_argument("--local_rank", type=int, default=0)
+    parser.add_argument(
+        "--ckpt",
+        help="The path to the checkpoint for test, default is the latest checkpoint.",
+        default=None,
+        type=str,
+    )
+
+    parser.add_argument("--output_dir", default="eval_results", type=str, help="The directory to store evaluation results.")
+
+    parser.add_argument(
+        "opts",
+        help="Modify config options using the command-line",
+        default=None,
+        nargs=argparse.REMAINDER,
+    )
+    args = parser.parse_args()
+
+    num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
+    distributed = num_gpus > 1
+
+    if torch.cuda.is_available():
+        # This flag allows you to enable the inbuilt cudnn auto-tuner to
+        # find the best algorithm to use for your hardware.
+        torch.backends.cudnn.benchmark = True
+    if distributed:
+        torch.cuda.set_device(args.local_rank)
+        torch.distributed.init_process_group(backend="nccl", init_method="env://")
+        synchronize()
+
+    cfg.merge_from_file(args.config_file)
+    cfg.merge_from_list(args.opts)
+    cfg.freeze()
+
+    logger = setup_logger("SSD", dist_util.get_rank(), cfg.OUTPUT_DIR)
+    logger.info("Using {} GPUs".format(num_gpus))
+    logger.info(args)
+
+    logger.info("Loaded configuration file {}".format(args.config_file))
+    with open(args.config_file, "r") as cf:
+        config_str = "\n" + cf.read()
+        logger.info(config_str)
+    logger.info("Running with config:\n{}".format(cfg))
+    evaluation(cfg, ckpt=args.ckpt, distributed=distributed)
+
+
+if __name__ == '__main__':
+    main()

train.py

+import argparse
+import logging
+import os
+
+import torch
+import torch.distributed as dist
+
+from ssd.engine.inference import do_evaluation
+from ssd.config import cfg
+from ssd.data.build import make_data_loader
+from ssd.engine.trainer import do_train
+from ssd.modeling.detector import build_detection_model
+from ssd.solver.build import make_optimizer, make_lr_scheduler
+from ssd.utils import dist_util, mkdir
+from ssd.utils.checkpoint import CheckPointer
+from ssd.utils.dist_util import synchronize
+from ssd.utils.logger import setup_logger
+from ssd.utils.misc import str2bool
+
+
+def train(cfg, args):
+    logger = logging.getLogger('SSD.trainer')
+    model = build_detection_model(cfg)
+    device = torch.device(cfg.MODEL.DEVICE)
+    model.to(device)
+    if args.distributed:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank)
+
+    lr = cfg.SOLVER.LR * args.num_gpus  # scale by num gpus
+    optimizer = make_optimizer(cfg, model, lr)
+
+    milestones = [step // args.num_gpus for step in cfg.SOLVER.LR_STEPS]
+    scheduler = make_lr_scheduler(cfg, optimizer, milestones)
+
+    arguments = {"iteration": 0}
+    save_to_disk = dist_util.get_rank() == 0
+    checkpointer = CheckPointer(model, optimizer, scheduler, cfg.OUTPUT_DIR, save_to_disk, logger)
+    extra_checkpoint_data = checkpointer.load()
+    arguments.update(extra_checkpoint_data)
+
+    max_iter = cfg.SOLVER.MAX_ITER // args.num_gpus
+    train_loader = make_data_loader(cfg, is_train=True, distributed=args.distributed, max_iter=max_iter, start_iter=arguments['iteration'])
+
+    model = do_train(cfg, model, train_loader, optimizer, scheduler, checkpointer, device, arguments, args)
+    return model
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Single Shot MultiBox Detector Training With PyTorch')
+    parser.add_argument(
+        "--config-file",
+        default="",
+        metavar="FILE",
+        help="path to config file",
+        type=str,
+    )
+    parser.add_argument("--local_rank", type=int, default=0)
+    parser.add_argument('--local-rank', type=int, default=0)
+    parser.add_argument('--log_step', default=10, type=int, help='Print logs every log_step')
+    parser.add_argument('--save_step', default=2500, type=int, help='Save checkpoint every save_step')
+    parser.add_argument('--eval_step', default=2500, type=int, help='Evaluate dataset every eval_step, disabled when eval_step < 0')
+    parser.add_argument('--use_tensorboard', default=True, type=str2bool)
+    parser.add_argument(
+        "--skip-test",
+        dest="skip_test",
+        help="Do not test the final model",
+        action="store_true",
+    )
+    parser.add_argument(
+        "opts",
+        help="Modify config options using the command-line",
+        default=None,
+        nargs=argparse.REMAINDER,
+    )
+    args = parser.parse_args()
+    if 'LOCAL_RANK' not in os.environ:
+        os.environ['LOCAL_RANK'] = str(args.local_rank)
+    num_gpus = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
+    args.distributed = num_gpus > 1
+    args.num_gpus = num_gpus
+
+    if torch.cuda.is_available():
+        # This flag allows you to enable the inbuilt cudnn auto-tuner to
+        # find the best algorithm to use for your hardware.
+        torch.backends.cudnn.benchmark = True
+    if args.distributed:
+        torch.cuda.set_device(args.local_rank)
+        torch.distributed.init_process_group(backend="nccl", init_method="env://")
+        synchronize()
+
+    cfg.merge_from_file(args.config_file)
+    cfg.merge_from_list(args.opts)
+    cfg.freeze()
+
+    if cfg.OUTPUT_DIR:
+        mkdir(cfg.OUTPUT_DIR)
+
+    logger = setup_logger("SSD", dist_util.get_rank(), cfg.OUTPUT_DIR)
+    logger.info("Using {} GPUs".format(num_gpus))
+    logger.info(args)
+
+    logger.info("Loaded configuration file {}".format(args.config_file))
+    with open(args.config_file, "r") as cf:
+        config_str = "\n" + cf.read()
+        logger.info(config_str)
+    logger.info("Running with config:\n{}".format(cfg))
+
+    model = train(cfg, args)
+
+    if not args.skip_test:
+        logger.info('Start evaluating...')
+        torch.cuda.empty_cache()  # speed up evaluating after training finished
+        do_evaluation(cfg, model, distributed=args.distributed)
+
+
+if __name__ == '__main__':
+    main()