Upload 440 files

.gitattributes

@@ -48,3 +48,5 @@ detectron2/detectron2/_C.cpython-311-x86_64-linux-gnu.so filter=lfs diff=lfs mer
 detectron2/detectron2/data/datasets/__pycache__/lvis_v0_5_categories.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 detectron2/detectron2/data/datasets/__pycache__/lvis_v1_categories.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 detectron2/_C.cpython-311-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
+detectron2/data/datasets/__pycache__/lvis_v0_5_categories.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
+detectron2/data/datasets/__pycache__/lvis_v1_categories.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text

detectron2/checkpoint/__init__.py

@@ -0,0 +1,10 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+# File:
+
+
+from . import catalog as _UNUSED  # register the handler
+from .detection_checkpoint import DetectionCheckpointer
+from fvcore.common.checkpoint import Checkpointer, PeriodicCheckpointer
+
+__all__ = ["Checkpointer", "PeriodicCheckpointer", "DetectionCheckpointer"]

detectron2/checkpoint/c2_model_loading.py

@@ -0,0 +1,406 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import copy
+import logging
+import re
+from typing import Dict, List
+import torch
+
+
+def convert_basic_c2_names(original_keys):
+    """
+    Apply some basic name conversion to names in C2 weights.
+    It only deals with typical backbone models.
+
+    Args:
+        original_keys (list[str]):
+    Returns:
+        list[str]: The same number of strings matching those in original_keys.
+    """
+    layer_keys = copy.deepcopy(original_keys)
+    layer_keys = [
+        {"pred_b": "linear_b", "pred_w": "linear_w"}.get(k, k) for k in layer_keys
+    ]  # some hard-coded mappings
+
+    layer_keys = [k.replace("_", ".") for k in layer_keys]
+    layer_keys = [re.sub("\\.b$", ".bias", k) for k in layer_keys]
+    layer_keys = [re.sub("\\.w$", ".weight", k) for k in layer_keys]
+    # Uniform both bn and gn names to "norm"
+    layer_keys = [re.sub("bn\\.s$", "norm.weight", k) for k in layer_keys]
+    layer_keys = [re.sub("bn\\.bias$", "norm.bias", k) for k in layer_keys]
+    layer_keys = [re.sub("bn\\.rm", "norm.running_mean", k) for k in layer_keys]
+    layer_keys = [re.sub("bn\\.running.mean$", "norm.running_mean", k) for k in layer_keys]
+    layer_keys = [re.sub("bn\\.riv$", "norm.running_var", k) for k in layer_keys]
+    layer_keys = [re.sub("bn\\.running.var$", "norm.running_var", k) for k in layer_keys]
+    layer_keys = [re.sub("bn\\.gamma$", "norm.weight", k) for k in layer_keys]
+    layer_keys = [re.sub("bn\\.beta$", "norm.bias", k) for k in layer_keys]
+    layer_keys = [re.sub("gn\\.s$", "norm.weight", k) for k in layer_keys]
+    layer_keys = [re.sub("gn\\.bias$", "norm.bias", k) for k in layer_keys]
+
+    # stem
+    layer_keys = [re.sub("^res\\.conv1\\.norm\\.", "conv1.norm.", k) for k in layer_keys]
+    # to avoid mis-matching with "conv1" in other components (e.g. detection head)
+    layer_keys = [re.sub("^conv1\\.", "stem.conv1.", k) for k in layer_keys]
+
+    # layer1-4 is used by torchvision, however we follow the C2 naming strategy (res2-5)
+    # layer_keys = [re.sub("^res2.", "layer1.", k) for k in layer_keys]
+    # layer_keys = [re.sub("^res3.", "layer2.", k) for k in layer_keys]
+    # layer_keys = [re.sub("^res4.", "layer3.", k) for k in layer_keys]
+    # layer_keys = [re.sub("^res5.", "layer4.", k) for k in layer_keys]
+
+    # blocks
+    layer_keys = [k.replace(".branch1.", ".shortcut.") for k in layer_keys]
+    layer_keys = [k.replace(".branch2a.", ".conv1.") for k in layer_keys]
+    layer_keys = [k.replace(".branch2b.", ".conv2.") for k in layer_keys]
+    layer_keys = [k.replace(".branch2c.", ".conv3.") for k in layer_keys]
+
+    # DensePose substitutions
+    layer_keys = [re.sub("^body.conv.fcn", "body_conv_fcn", k) for k in layer_keys]
+    layer_keys = [k.replace("AnnIndex.lowres", "ann_index_lowres") for k in layer_keys]
+    layer_keys = [k.replace("Index.UV.lowres", "index_uv_lowres") for k in layer_keys]
+    layer_keys = [k.replace("U.lowres", "u_lowres") for k in layer_keys]
+    layer_keys = [k.replace("V.lowres", "v_lowres") for k in layer_keys]
+    return layer_keys
+
+
+def convert_c2_detectron_names(weights):
+    """
+    Map Caffe2 Detectron weight names to Detectron2 names.
+
+    Args:
+        weights (dict): name -> tensor
+
+    Returns:
+        dict: detectron2 names -> tensor
+        dict: detectron2 names -> C2 names
+    """
+    logger = logging.getLogger(__name__)
+    logger.info("Renaming Caffe2 weights ......")
+    original_keys = sorted(weights.keys())
+    layer_keys = copy.deepcopy(original_keys)
+
+    layer_keys = convert_basic_c2_names(layer_keys)
+
+    # --------------------------------------------------------------------------
+    # RPN hidden representation conv
+    # --------------------------------------------------------------------------
+    # FPN case
+    # In the C2 model, the RPN hidden layer conv is defined for FPN level 2 and then
+    # shared for all other levels, hence the appearance of "fpn2"
+    layer_keys = [
+        k.replace("conv.rpn.fpn2", "proposal_generator.rpn_head.conv") for k in layer_keys
+    ]
+    # Non-FPN case
+    layer_keys = [k.replace("conv.rpn", "proposal_generator.rpn_head.conv") for k in layer_keys]
+
+    # --------------------------------------------------------------------------
+    # RPN box transformation conv
+    # --------------------------------------------------------------------------
+    # FPN case (see note above about "fpn2")
+    layer_keys = [
+        k.replace("rpn.bbox.pred.fpn2", "proposal_generator.rpn_head.anchor_deltas")
+        for k in layer_keys
+    ]
+    layer_keys = [
+        k.replace("rpn.cls.logits.fpn2", "proposal_generator.rpn_head.objectness_logits")
+        for k in layer_keys
+    ]
+    # Non-FPN case
+    layer_keys = [
+        k.replace("rpn.bbox.pred", "proposal_generator.rpn_head.anchor_deltas") for k in layer_keys
+    ]
+    layer_keys = [
+        k.replace("rpn.cls.logits", "proposal_generator.rpn_head.objectness_logits")
+        for k in layer_keys
+    ]
+
+    # --------------------------------------------------------------------------
+    # Fast R-CNN box head
+    # --------------------------------------------------------------------------
+    layer_keys = [re.sub("^bbox\\.pred", "bbox_pred", k) for k in layer_keys]
+    layer_keys = [re.sub("^cls\\.score", "cls_score", k) for k in layer_keys]
+    layer_keys = [re.sub("^fc6\\.", "box_head.fc1.", k) for k in layer_keys]
+    layer_keys = [re.sub("^fc7\\.", "box_head.fc2.", k) for k in layer_keys]
+    # 4conv1fc head tensor names: head_conv1_w, head_conv1_gn_s
+    layer_keys = [re.sub("^head\\.conv", "box_head.conv", k) for k in layer_keys]
+
+    # --------------------------------------------------------------------------
+    # FPN lateral and output convolutions
+    # --------------------------------------------------------------------------
+    def fpn_map(name):
+        """
+        Look for keys with the following patterns:
+        1) Starts with "fpn.inner."
+           Example: "fpn.inner.res2.2.sum.lateral.weight"
+           Meaning: These are lateral pathway convolutions
+        2) Starts with "fpn.res"
+           Example: "fpn.res2.2.sum.weight"
+           Meaning: These are FPN output convolutions
+        """
+        splits = name.split(".")
+        norm = ".norm" if "norm" in splits else ""
+        if name.startswith("fpn.inner."):
+            # splits example: ['fpn', 'inner', 'res2', '2', 'sum', 'lateral', 'weight']
+            stage = int(splits[2][len("res") :])
+            return "fpn_lateral{}{}.{}".format(stage, norm, splits[-1])
+        elif name.startswith("fpn.res"):
+            # splits example: ['fpn', 'res2', '2', 'sum', 'weight']
+            stage = int(splits[1][len("res") :])
+            return "fpn_output{}{}.{}".format(stage, norm, splits[-1])
+        return name
+
+    layer_keys = [fpn_map(k) for k in layer_keys]
+
+    # --------------------------------------------------------------------------
+    # Mask R-CNN mask head
+    # --------------------------------------------------------------------------
+    # roi_heads.StandardROIHeads case
+    layer_keys = [k.replace(".[mask].fcn", "mask_head.mask_fcn") for k in layer_keys]
+    layer_keys = [re.sub("^\\.mask\\.fcn", "mask_head.mask_fcn", k) for k in layer_keys]
+    layer_keys = [k.replace("mask.fcn.logits", "mask_head.predictor") for k in layer_keys]
+    # roi_heads.Res5ROIHeads case
+    layer_keys = [k.replace("conv5.mask", "mask_head.deconv") for k in layer_keys]
+
+    # --------------------------------------------------------------------------
+    # Keypoint R-CNN head
+    # --------------------------------------------------------------------------
+    # interestingly, the keypoint head convs have blob names that are simply "conv_fcnX"
+    layer_keys = [k.replace("conv.fcn", "roi_heads.keypoint_head.conv_fcn") for k in layer_keys]
+    layer_keys = [
+        k.replace("kps.score.lowres", "roi_heads.keypoint_head.score_lowres") for k in layer_keys
+    ]
+    layer_keys = [k.replace("kps.score.", "roi_heads.keypoint_head.score.") for k in layer_keys]
+
+    # --------------------------------------------------------------------------
+    # Done with replacements
+    # --------------------------------------------------------------------------
+    assert len(set(layer_keys)) == len(layer_keys)
+    assert len(original_keys) == len(layer_keys)
+
+    new_weights = {}
+    new_keys_to_original_keys = {}
+    for orig, renamed in zip(original_keys, layer_keys):
+        new_keys_to_original_keys[renamed] = orig
+        if renamed.startswith("bbox_pred.") or renamed.startswith("mask_head.predictor."):
+            # remove the meaningless prediction weight for background class
+            new_start_idx = 4 if renamed.startswith("bbox_pred.") else 1
+            new_weights[renamed] = weights[orig][new_start_idx:]
+            logger.info(
+                "Remove prediction weight for background class in {}. The shape changes from "
+                "{} to {}.".format(
+                    renamed, tuple(weights[orig].shape), tuple(new_weights[renamed].shape)
+                )
+            )
+        elif renamed.startswith("cls_score."):
+            # move weights of bg class from original index 0 to last index
+            logger.info(
+                "Move classification weights for background class in {} from index 0 to "
+                "index {}.".format(renamed, weights[orig].shape[0] - 1)
+            )
+            new_weights[renamed] = torch.cat([weights[orig][1:], weights[orig][:1]])
+        else:
+            new_weights[renamed] = weights[orig]
+
+    return new_weights, new_keys_to_original_keys
+
+
+# Note the current matching is not symmetric.
+# it assumes model_state_dict will have longer names.
+def align_and_update_state_dicts(model_state_dict, ckpt_state_dict, c2_conversion=True):
+    """
+    Match names between the two state-dict, and returns a new chkpt_state_dict with names
+    converted to match model_state_dict with heuristics. The returned dict can be later
+    loaded with fvcore checkpointer.
+    If `c2_conversion==True`, `ckpt_state_dict` is assumed to be a Caffe2
+    model and will be renamed at first.
+
+    Strategy: suppose that the models that we will create will have prefixes appended
+    to each of its keys, for example due to an extra level of nesting that the original
+    pre-trained weights from ImageNet won't contain. For example, model.state_dict()
+    might return backbone[0].body.res2.conv1.weight, while the pre-trained model contains
+    res2.conv1.weight. We thus want to match both parameters together.
+    For that, we look for each model weight, look among all loaded keys if there is one
+    that is a suffix of the current weight name, and use it if that's the case.
+    If multiple matches exist, take the one with longest size
+    of the corresponding name. For example, for the same model as before, the pretrained
+    weight file can contain both res2.conv1.weight, as well as conv1.weight. In this case,
+    we want to match backbone[0].body.conv1.weight to conv1.weight, and
+    backbone[0].body.res2.conv1.weight to res2.conv1.weight.
+    """
+    model_keys = sorted(model_state_dict.keys())
+    if c2_conversion:
+        ckpt_state_dict, original_keys = convert_c2_detectron_names(ckpt_state_dict)
+        # original_keys: the name in the original dict (before renaming)
+    else:
+        original_keys = {x: x for x in ckpt_state_dict.keys()}
+    ckpt_keys = sorted(ckpt_state_dict.keys())
+
+    def match(a, b):
+        # Matched ckpt_key should be a complete (starts with '.') suffix.
+        # For example, roi_heads.mesh_head.whatever_conv1 does not match conv1,
+        # but matches whatever_conv1 or mesh_head.whatever_conv1.
+        return a == b or a.endswith("." + b)
+
+    # get a matrix of string matches, where each (i, j) entry correspond to the size of the
+    # ckpt_key string, if it matches
+    match_matrix = [len(j) if match(i, j) else 0 for i in model_keys for j in ckpt_keys]
+    match_matrix = torch.as_tensor(match_matrix).view(len(model_keys), len(ckpt_keys))
+    # use the matched one with longest size in case of multiple matches
+    max_match_size, idxs = match_matrix.max(1)
+    # remove indices that correspond to no-match
+    idxs[max_match_size == 0] = -1
+
+    logger = logging.getLogger(__name__)
+    # matched_pairs (matched checkpoint key --> matched model key)
+    matched_keys = {}
+    result_state_dict = {}
+    for idx_model, idx_ckpt in enumerate(idxs.tolist()):
+        if idx_ckpt == -1:
+            continue
+        key_model = model_keys[idx_model]
+        key_ckpt = ckpt_keys[idx_ckpt]
+        value_ckpt = ckpt_state_dict[key_ckpt]
+        shape_in_model = model_state_dict[key_model].shape
+
+        if shape_in_model != value_ckpt.shape:
+            logger.warning(
+                "Shape of {} in checkpoint is {}, while shape of {} in model is {}.".format(
+                    key_ckpt, value_ckpt.shape, key_model, shape_in_model
+                )
+            )
+            logger.warning(
+                "{} will not be loaded. Please double check and see if this is desired.".format(
+                    key_ckpt
+                )
+            )
+            continue
+
+        assert key_model not in result_state_dict
+        result_state_dict[key_model] = value_ckpt
+        if key_ckpt in matched_keys:  # already added to matched_keys
+            logger.error(
+                "Ambiguity found for {} in checkpoint!"
+                "It matches at least two keys in the model ({} and {}).".format(
+                    key_ckpt, key_model, matched_keys[key_ckpt]
+                )
+            )
+            raise ValueError("Cannot match one checkpoint key to multiple keys in the model.")
+
+        matched_keys[key_ckpt] = key_model
+
+    # logging:
+    matched_model_keys = sorted(matched_keys.values())
+    if len(matched_model_keys) == 0:
+        logger.warning("No weights in checkpoint matched with model.")
+        return ckpt_state_dict
+    common_prefix = _longest_common_prefix(matched_model_keys)
+    rev_matched_keys = {v: k for k, v in matched_keys.items()}
+    original_keys = {k: original_keys[rev_matched_keys[k]] for k in matched_model_keys}
+
+    model_key_groups = _group_keys_by_module(matched_model_keys, original_keys)
+    table = []
+    memo = set()
+    for key_model in matched_model_keys:
+        if key_model in memo:
+            continue
+        if key_model in model_key_groups:
+            group = model_key_groups[key_model]
+            memo |= set(group)
+            shapes = [tuple(model_state_dict[k].shape) for k in group]
+            table.append(
+                (
+                    _longest_common_prefix([k[len(common_prefix) :] for k in group]) + "*",
+                    _group_str([original_keys[k] for k in group]),
+                    " ".join([str(x).replace(" ", "") for x in shapes]),
+                )
+            )
+        else:
+            key_checkpoint = original_keys[key_model]
+            shape = str(tuple(model_state_dict[key_model].shape))
+            table.append((key_model[len(common_prefix) :], key_checkpoint, shape))
+    submodule_str = common_prefix[:-1] if common_prefix else "model"
+    logger.info(
+        f"Following weights matched with submodule {submodule_str} - Total num: {len(table)}"
+    )
+
+    unmatched_ckpt_keys = [k for k in ckpt_keys if k not in set(matched_keys.keys())]
+    for k in unmatched_ckpt_keys:
+        result_state_dict[k] = ckpt_state_dict[k]
+    return result_state_dict
+
+
+def _group_keys_by_module(keys: List[str], original_names: Dict[str, str]):
+    """
+    Params in the same submodule are grouped together.
+
+    Args:
+        keys: names of all parameters
+        original_names: mapping from parameter name to their name in the checkpoint
+
+    Returns:
+        dict[name -> all other names in the same group]
+    """
+
+    def _submodule_name(key):
+        pos = key.rfind(".")
+        if pos < 0:
+            return None
+        prefix = key[: pos + 1]
+        return prefix
+
+    all_submodules = [_submodule_name(k) for k in keys]
+    all_submodules = [x for x in all_submodules if x]
+    all_submodules = sorted(all_submodules, key=len)
+
+    ret = {}
+    for prefix in all_submodules:
+        group = [k for k in keys if k.startswith(prefix)]
+        if len(group) <= 1:
+            continue
+        original_name_lcp = _longest_common_prefix_str([original_names[k] for k in group])
+        if len(original_name_lcp) == 0:
+            # don't group weights if original names don't share prefix
+            continue
+
+        for k in group:
+            if k in ret:
+                continue
+            ret[k] = group
+    return ret
+
+
+def _longest_common_prefix(names: List[str]) -> str:
+    """
+    ["abc.zfg", "abc.zef"] -> "abc."
+    """
+    names = [n.split(".") for n in names]
+    m1, m2 = min(names), max(names)
+    ret = [a for a, b in zip(m1, m2) if a == b]
+    ret = ".".join(ret) + "." if len(ret) else ""
+    return ret
+
+
+def _longest_common_prefix_str(names: List[str]) -> str:
+    m1, m2 = min(names), max(names)
+    lcp = []
+    for a, b in zip(m1, m2):
+        if a == b:
+            lcp.append(a)
+        else:
+            break
+    lcp = "".join(lcp)
+    return lcp
+
+
+def _group_str(names: List[str]) -> str:
+    """
+    Turn "common1", "common2", "common3" into "common{1,2,3}"
+    """
+    lcp = _longest_common_prefix_str(names)
+    rest = [x[len(lcp) :] for x in names]
+    rest = "{" + ",".join(rest) + "}"
+    ret = lcp + rest
+
+    # add some simplification for BN specifically
+    ret = ret.replace("bn_{beta,running_mean,running_var,gamma}", "bn_*")
+    ret = ret.replace("bn_beta,bn_running_mean,bn_running_var,bn_gamma", "bn_*")
+    return ret

detectron2/checkpoint/catalog.py

@@ -0,0 +1,115 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import logging
+
+from detectron2.utils.file_io import PathHandler, PathManager
+
+
+class ModelCatalog:
+    """
+    Store mappings from names to third-party models.
+    """
+
+    S3_C2_DETECTRON_PREFIX = "https://dl.fbaipublicfiles.com/detectron"
+
+    # MSRA models have STRIDE_IN_1X1=True. False otherwise.
+    # NOTE: all BN models here have fused BN into an affine layer.
+    # As a result, you should only load them to a model with "FrozenBN".
+    # Loading them to a model with regular BN or SyncBN is wrong.
+    # Even when loaded to FrozenBN, it is still different from affine by an epsilon,
+    # which should be negligible for training.
+    # NOTE: all models here uses PIXEL_STD=[1,1,1]
+    # NOTE: Most of the BN models here are no longer used. We use the
+    # re-converted pre-trained models under detectron2 model zoo instead.
+    C2_IMAGENET_MODELS = {
+        "MSRA/R-50": "ImageNetPretrained/MSRA/R-50.pkl",
+        "MSRA/R-101": "ImageNetPretrained/MSRA/R-101.pkl",
+        "FAIR/R-50-GN": "ImageNetPretrained/47261647/R-50-GN.pkl",
+        "FAIR/R-101-GN": "ImageNetPretrained/47592356/R-101-GN.pkl",
+        "FAIR/X-101-32x8d": "ImageNetPretrained/20171220/X-101-32x8d.pkl",
+        "FAIR/X-101-64x4d": "ImageNetPretrained/FBResNeXt/X-101-64x4d.pkl",
+        "FAIR/X-152-32x8d-IN5k": "ImageNetPretrained/25093814/X-152-32x8d-IN5k.pkl",
+    }
+
+    C2_DETECTRON_PATH_FORMAT = (
+        "{prefix}/{url}/output/train/{dataset}/{type}/model_final.pkl"  # noqa B950
+    )
+
+    C2_DATASET_COCO = "coco_2014_train%3Acoco_2014_valminusminival"
+    C2_DATASET_COCO_KEYPOINTS = "keypoints_coco_2014_train%3Akeypoints_coco_2014_valminusminival"
+
+    # format: {model_name} -> part of the url
+    C2_DETECTRON_MODELS = {
+        "35857197/e2e_faster_rcnn_R-50-C4_1x": "35857197/12_2017_baselines/e2e_faster_rcnn_R-50-C4_1x.yaml.01_33_49.iAX0mXvW",  # noqa B950
+        "35857345/e2e_faster_rcnn_R-50-FPN_1x": "35857345/12_2017_baselines/e2e_faster_rcnn_R-50-FPN_1x.yaml.01_36_30.cUF7QR7I",  # noqa B950
+        "35857890/e2e_faster_rcnn_R-101-FPN_1x": "35857890/12_2017_baselines/e2e_faster_rcnn_R-101-FPN_1x.yaml.01_38_50.sNxI7sX7",  # noqa B950
+        "36761737/e2e_faster_rcnn_X-101-32x8d-FPN_1x": "36761737/12_2017_baselines/e2e_faster_rcnn_X-101-32x8d-FPN_1x.yaml.06_31_39.5MIHi1fZ",  # noqa B950
+        "35858791/e2e_mask_rcnn_R-50-C4_1x": "35858791/12_2017_baselines/e2e_mask_rcnn_R-50-C4_1x.yaml.01_45_57.ZgkA7hPB",  # noqa B950
+        "35858933/e2e_mask_rcnn_R-50-FPN_1x": "35858933/12_2017_baselines/e2e_mask_rcnn_R-50-FPN_1x.yaml.01_48_14.DzEQe4wC",  # noqa B950
+        "35861795/e2e_mask_rcnn_R-101-FPN_1x": "35861795/12_2017_baselines/e2e_mask_rcnn_R-101-FPN_1x.yaml.02_31_37.KqyEK4tT",  # noqa B950
+        "36761843/e2e_mask_rcnn_X-101-32x8d-FPN_1x": "36761843/12_2017_baselines/e2e_mask_rcnn_X-101-32x8d-FPN_1x.yaml.06_35_59.RZotkLKI",  # noqa B950
+        "48616381/e2e_mask_rcnn_R-50-FPN_2x_gn": "GN/48616381/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn_0416.13_23_38.bTlTI97Q",  # noqa B950
+        "37697547/e2e_keypoint_rcnn_R-50-FPN_1x": "37697547/12_2017_baselines/e2e_keypoint_rcnn_R-50-FPN_1x.yaml.08_42_54.kdzV35ao",  # noqa B950
+        "35998355/rpn_R-50-C4_1x": "35998355/12_2017_baselines/rpn_R-50-C4_1x.yaml.08_00_43.njH5oD9L",  # noqa B950
+        "35998814/rpn_R-50-FPN_1x": "35998814/12_2017_baselines/rpn_R-50-FPN_1x.yaml.08_06_03.Axg0r179",  # noqa B950
+        "36225147/fast_R-50-FPN_1x": "36225147/12_2017_baselines/fast_rcnn_R-50-FPN_1x.yaml.08_39_09.L3obSdQ2",  # noqa B950
+    }
+
+    @staticmethod
+    def get(name):
+        if name.startswith("Caffe2Detectron/COCO"):
+            return ModelCatalog._get_c2_detectron_baseline(name)
+        if name.startswith("ImageNetPretrained/"):
+            return ModelCatalog._get_c2_imagenet_pretrained(name)
+        raise RuntimeError("model not present in the catalog: {}".format(name))
+
+    @staticmethod
+    def _get_c2_imagenet_pretrained(name):
+        prefix = ModelCatalog.S3_C2_DETECTRON_PREFIX
+        name = name[len("ImageNetPretrained/") :]
+        name = ModelCatalog.C2_IMAGENET_MODELS[name]
+        url = "/".join([prefix, name])
+        return url
+
+    @staticmethod
+    def _get_c2_detectron_baseline(name):
+        name = name[len("Caffe2Detectron/COCO/") :]
+        url = ModelCatalog.C2_DETECTRON_MODELS[name]
+        if "keypoint_rcnn" in name:
+            dataset = ModelCatalog.C2_DATASET_COCO_KEYPOINTS
+        else:
+            dataset = ModelCatalog.C2_DATASET_COCO
+
+        if "35998355/rpn_R-50-C4_1x" in name:
+            # this one model is somehow different from others ..
+            type = "rpn"
+        else:
+            type = "generalized_rcnn"
+
+        # Detectron C2 models are stored in the structure defined in `C2_DETECTRON_PATH_FORMAT`.
+        url = ModelCatalog.C2_DETECTRON_PATH_FORMAT.format(
+            prefix=ModelCatalog.S3_C2_DETECTRON_PREFIX, url=url, type=type, dataset=dataset
+        )
+        return url
+
+
+class ModelCatalogHandler(PathHandler):
+    """
+    Resolve URL like catalog://.
+    """
+
+    PREFIX = "catalog://"
+
+    def _get_supported_prefixes(self):
+        return [self.PREFIX]
+
+    def _get_local_path(self, path, **kwargs):
+        logger = logging.getLogger(__name__)
+        catalog_path = ModelCatalog.get(path[len(self.PREFIX) :])
+        logger.info("Catalog entry {} points to {}".format(path, catalog_path))
+        return PathManager.get_local_path(catalog_path, **kwargs)
+
+    def _open(self, path, mode="r", **kwargs):
+        return PathManager.open(self._get_local_path(path), mode, **kwargs)
+
+
+PathManager.register_handler(ModelCatalogHandler())

detectron2/checkpoint/detection_checkpoint.py

@@ -0,0 +1,143 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import logging
+import os
+import pickle
+from urllib.parse import parse_qs, urlparse
+import torch
+from fvcore.common.checkpoint import Checkpointer
+from torch.nn.parallel import DistributedDataParallel
+
+import detectron2.utils.comm as comm
+from detectron2.utils.file_io import PathManager
+
+from .c2_model_loading import align_and_update_state_dicts
+
+
+class DetectionCheckpointer(Checkpointer):
+    """
+    Same as :class:`Checkpointer`, but is able to:
+    1. handle models in detectron & detectron2 model zoo, and apply conversions for legacy models.
+    2. correctly load checkpoints that are only available on the master worker
+    """
+
+    def __init__(self, model, save_dir="", *, save_to_disk=None, **checkpointables):
+        is_main_process = comm.is_main_process()
+        super().__init__(
+            model,
+            save_dir,
+            save_to_disk=is_main_process if save_to_disk is None else save_to_disk,
+            **checkpointables,
+        )
+        self.path_manager = PathManager
+        self._parsed_url_during_load = None
+
+    def load(self, path, *args, **kwargs):
+        assert self._parsed_url_during_load is None
+        need_sync = False
+        logger = logging.getLogger(__name__)
+        logger.info("[DetectionCheckpointer] Loading from {} ...".format(path))
+
+        if path and isinstance(self.model, DistributedDataParallel):
+            path = self.path_manager.get_local_path(path)
+            has_file = os.path.isfile(path)
+            all_has_file = comm.all_gather(has_file)
+            if not all_has_file[0]:
+                raise OSError(f"File {path} not found on main worker.")
+            if not all(all_has_file):
+                logger.warning(
+                    f"Not all workers can read checkpoint {path}. "
+                    "Training may fail to fully resume."
+                )
+                # TODO: broadcast the checkpoint file contents from main
+                # worker, and load from it instead.
+                need_sync = True
+            if not has_file:
+                path = None  # don't load if not readable
+
+        if path:
+            parsed_url = urlparse(path)
+            self._parsed_url_during_load = parsed_url
+            path = parsed_url._replace(query="").geturl()  # remove query from filename
+            path = self.path_manager.get_local_path(path)
+        ret = super().load(path, *args, **kwargs)
+
+        if need_sync:
+            logger.info("Broadcasting model states from main worker ...")
+            self.model._sync_params_and_buffers()
+        self._parsed_url_during_load = None  # reset to None
+        return ret
+
+    def _load_file(self, filename):
+        if filename.endswith(".pkl"):
+            with PathManager.open(filename, "rb") as f:
+                data = pickle.load(f, encoding="latin1")
+            if "model" in data and "__author__" in data:
+                # file is in Detectron2 model zoo format
+                self.logger.info("Reading a file from '{}'".format(data["__author__"]))
+                return data
+            else:
+                # assume file is from Caffe2 / Detectron1 model zoo
+                if "blobs" in data:
+                    # Detection models have "blobs", but ImageNet models don't
+                    data = data["blobs"]
+                data = {k: v for k, v in data.items() if not k.endswith("_momentum")}
+                return {"model": data, "__author__": "Caffe2", "matching_heuristics": True}
+        elif filename.endswith(".pyth"):
+            # assume file is from pycls; no one else seems to use the ".pyth" extension
+            with PathManager.open(filename, "rb") as f:
+                data = torch.load(f)
+            assert (
+                "model_state" in data
+            ), f"Cannot load .pyth file {filename}; pycls checkpoints must contain 'model_state'."
+            model_state = {
+                k: v
+                for k, v in data["model_state"].items()
+                if not k.endswith("num_batches_tracked")
+            }
+            return {"model": model_state, "__author__": "pycls", "matching_heuristics": True}
+
+        loaded = self._torch_load(filename)
+        if "model" not in loaded:
+            loaded = {"model": loaded}
+        assert self._parsed_url_during_load is not None, "`_load_file` must be called inside `load`"
+        parsed_url = self._parsed_url_during_load
+        queries = parse_qs(parsed_url.query)
+        if queries.pop("matching_heuristics", "False") == ["True"]:
+            loaded["matching_heuristics"] = True
+        if len(queries) > 0:
+            raise ValueError(
+                f"Unsupported query remaining: f{queries}, orginal filename: {parsed_url.geturl()}"
+            )
+        return loaded
+
+    def _torch_load(self, f):
+        return super()._load_file(f)
+
+    def _load_model(self, checkpoint):
+        if checkpoint.get("matching_heuristics", False):
+            self._convert_ndarray_to_tensor(checkpoint["model"])
+            # convert weights by name-matching heuristics
+            checkpoint["model"] = align_and_update_state_dicts(
+                self.model.state_dict(),
+                checkpoint["model"],
+                c2_conversion=checkpoint.get("__author__", None) == "Caffe2",
+            )
+        # for non-caffe2 models, use standard ways to load it
+        incompatible = super()._load_model(checkpoint)
+
+        model_buffers = dict(self.model.named_buffers(recurse=False))
+        for k in ["pixel_mean", "pixel_std"]:
+            # Ignore missing key message about pixel_mean/std.
+            # Though they may be missing in old checkpoints, they will be correctly
+            # initialized from config anyway.
+            if k in model_buffers:
+                try:
+                    incompatible.missing_keys.remove(k)
+                except ValueError:
+                    pass
+        for k in incompatible.unexpected_keys[:]:
+            # Ignore unexpected keys about cell anchors. They exist in old checkpoints
+            # but now they are non-persistent buffers and will not be in new checkpoints.
+            if "anchor_generator.cell_anchors" in k:
+                incompatible.unexpected_keys.remove(k)
+        return incompatible

detectron2/config/__init__.py

@@ -0,0 +1,24 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+from .compat import downgrade_config, upgrade_config
+from .config import CfgNode, get_cfg, global_cfg, set_global_cfg, configurable
+from .instantiate import instantiate
+from .lazy import LazyCall, LazyConfig
+
+__all__ = [
+    "CfgNode",
+    "get_cfg",
+    "global_cfg",
+    "set_global_cfg",
+    "downgrade_config",
+    "upgrade_config",
+    "configurable",
+    "instantiate",
+    "LazyCall",
+    "LazyConfig",
+]
+
+
+from detectron2.utils.env import fixup_module_metadata
+
+fixup_module_metadata(__name__, globals(), __all__)
+del fixup_module_metadata

detectron2/config/compat.py

@@ -0,0 +1,229 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+"""
+Backward compatibility of configs.
+
+Instructions to bump version:
++ It's not needed to bump version if new keys are added.
+  It's only needed when backward-incompatible changes happen
+  (i.e., some existing keys disappear, or the meaning of a key changes)
++ To bump version, do the following:
+    1. Increment _C.VERSION in defaults.py
+    2. Add a converter in this file.
+
+      Each ConverterVX has a function "upgrade" which in-place upgrades config from X-1 to X,
+      and a function "downgrade" which in-place downgrades config from X to X-1
+
+      In each function, VERSION is left unchanged.
+
+      Each converter assumes that its input has the relevant keys
+      (i.e., the input is not a partial config).
+    3. Run the tests (test_config.py) to make sure the upgrade & downgrade
+       functions are consistent.
+"""
+
+import logging
+from typing import List, Optional, Tuple
+
+from .config import CfgNode as CN
+from .defaults import _C
+
+__all__ = ["upgrade_config", "downgrade_config"]
+
+
+def upgrade_config(cfg: CN, to_version: Optional[int] = None) -> CN:
+    """
+    Upgrade a config from its current version to a newer version.
+
+    Args:
+        cfg (CfgNode):
+        to_version (int): defaults to the latest version.
+    """
+    cfg = cfg.clone()
+    if to_version is None:
+        to_version = _C.VERSION
+
+    assert cfg.VERSION <= to_version, "Cannot upgrade from v{} to v{}!".format(
+        cfg.VERSION, to_version
+    )
+    for k in range(cfg.VERSION, to_version):
+        converter = globals()["ConverterV" + str(k + 1)]
+        converter.upgrade(cfg)
+        cfg.VERSION = k + 1
+    return cfg
+
+
+def downgrade_config(cfg: CN, to_version: int) -> CN:
+    """
+    Downgrade a config from its current version to an older version.
+
+    Args:
+        cfg (CfgNode):
+        to_version (int):
+
+    Note:
+        A general downgrade of arbitrary configs is not always possible due to the
+        different functionalities in different versions.
+        The purpose of downgrade is only to recover the defaults in old versions,
+        allowing it to load an old partial yaml config.
+        Therefore, the implementation only needs to fill in the default values
+        in the old version when a general downgrade is not possible.
+    """
+    cfg = cfg.clone()
+    assert cfg.VERSION >= to_version, "Cannot downgrade from v{} to v{}!".format(
+        cfg.VERSION, to_version
+    )
+    for k in range(cfg.VERSION, to_version, -1):
+        converter = globals()["ConverterV" + str(k)]
+        converter.downgrade(cfg)
+        cfg.VERSION = k - 1
+    return cfg
+
+
+def guess_version(cfg: CN, filename: str) -> int:
+    """
+    Guess the version of a partial config where the VERSION field is not specified.
+    Returns the version, or the latest if cannot make a guess.
+
+    This makes it easier for users to migrate.
+    """
+    logger = logging.getLogger(__name__)
+
+    def _has(name: str) -> bool:
+        cur = cfg
+        for n in name.split("."):
+            if n not in cur:
+                return False
+            cur = cur[n]
+        return True
+
+    # Most users' partial configs have "MODEL.WEIGHT", so guess on it
+    ret = None
+    if _has("MODEL.WEIGHT") or _has("TEST.AUG_ON"):
+        ret = 1
+
+    if ret is not None:
+        logger.warning("Config '{}' has no VERSION. Assuming it to be v{}.".format(filename, ret))
+    else:
+        ret = _C.VERSION
+        logger.warning(
+            "Config '{}' has no VERSION. Assuming it to be compatible with latest v{}.".format(
+                filename, ret
+            )
+        )
+    return ret
+
+
+def _rename(cfg: CN, old: str, new: str) -> None:
+    old_keys = old.split(".")
+    new_keys = new.split(".")
+
+    def _set(key_seq: List[str], val: str) -> None:
+        cur = cfg
+        for k in key_seq[:-1]:
+            if k not in cur:
+                cur[k] = CN()
+            cur = cur[k]
+        cur[key_seq[-1]] = val
+
+    def _get(key_seq: List[str]) -> CN:
+        cur = cfg
+        for k in key_seq:
+            cur = cur[k]
+        return cur
+
+    def _del(key_seq: List[str]) -> None:
+        cur = cfg
+        for k in key_seq[:-1]:
+            cur = cur[k]
+        del cur[key_seq[-1]]
+        if len(cur) == 0 and len(key_seq) > 1:
+            _del(key_seq[:-1])
+
+    _set(new_keys, _get(old_keys))
+    _del(old_keys)
+
+
+class _RenameConverter:
+    """
+    A converter that handles simple rename.
+    """
+
+    RENAME: List[Tuple[str, str]] = []  # list of tuples of (old name, new name)
+
+    @classmethod
+    def upgrade(cls, cfg: CN) -> None:
+        for old, new in cls.RENAME:
+            _rename(cfg, old, new)
+
+    @classmethod
+    def downgrade(cls, cfg: CN) -> None:
+        for old, new in cls.RENAME[::-1]:
+            _rename(cfg, new, old)
+
+
+class ConverterV1(_RenameConverter):
+    RENAME = [("MODEL.RPN_HEAD.NAME", "MODEL.RPN.HEAD_NAME")]
+
+
+class ConverterV2(_RenameConverter):
+    """
+    A large bulk of rename, before public release.
+    """
+
+    RENAME = [
+        ("MODEL.WEIGHT", "MODEL.WEIGHTS"),
+        ("MODEL.PANOPTIC_FPN.SEMANTIC_LOSS_SCALE", "MODEL.SEM_SEG_HEAD.LOSS_WEIGHT"),
+        ("MODEL.PANOPTIC_FPN.RPN_LOSS_SCALE", "MODEL.RPN.LOSS_WEIGHT"),
+        ("MODEL.PANOPTIC_FPN.INSTANCE_LOSS_SCALE", "MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT"),
+        ("MODEL.PANOPTIC_FPN.COMBINE_ON", "MODEL.PANOPTIC_FPN.COMBINE.ENABLED"),
+        (
+            "MODEL.PANOPTIC_FPN.COMBINE_OVERLAP_THRESHOLD",
+            "MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH",
+        ),
+        (
+            "MODEL.PANOPTIC_FPN.COMBINE_STUFF_AREA_LIMIT",
+            "MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT",
+        ),
+        (
+            "MODEL.PANOPTIC_FPN.COMBINE_INSTANCES_CONFIDENCE_THRESHOLD",
+            "MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH",
+        ),
+        ("MODEL.ROI_HEADS.SCORE_THRESH", "MODEL.ROI_HEADS.SCORE_THRESH_TEST"),
+        ("MODEL.ROI_HEADS.NMS", "MODEL.ROI_HEADS.NMS_THRESH_TEST"),
+        ("MODEL.RETINANET.INFERENCE_SCORE_THRESHOLD", "MODEL.RETINANET.SCORE_THRESH_TEST"),
+        ("MODEL.RETINANET.INFERENCE_TOPK_CANDIDATES", "MODEL.RETINANET.TOPK_CANDIDATES_TEST"),
+        ("MODEL.RETINANET.INFERENCE_NMS_THRESHOLD", "MODEL.RETINANET.NMS_THRESH_TEST"),
+        ("TEST.DETECTIONS_PER_IMG", "TEST.DETECTIONS_PER_IMAGE"),
+        ("TEST.AUG_ON", "TEST.AUG.ENABLED"),
+        ("TEST.AUG_MIN_SIZES", "TEST.AUG.MIN_SIZES"),
+        ("TEST.AUG_MAX_SIZE", "TEST.AUG.MAX_SIZE"),
+        ("TEST.AUG_FLIP", "TEST.AUG.FLIP"),
+    ]
+
+    @classmethod
+    def upgrade(cls, cfg: CN) -> None:
+        super().upgrade(cfg)
+
+        if cfg.MODEL.META_ARCHITECTURE == "RetinaNet":
+            _rename(
+                cfg, "MODEL.RETINANET.ANCHOR_ASPECT_RATIOS", "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS"
+            )
+            _rename(cfg, "MODEL.RETINANET.ANCHOR_SIZES", "MODEL.ANCHOR_GENERATOR.SIZES")
+            del cfg["MODEL"]["RPN"]["ANCHOR_SIZES"]
+            del cfg["MODEL"]["RPN"]["ANCHOR_ASPECT_RATIOS"]
+        else:
+            _rename(cfg, "MODEL.RPN.ANCHOR_ASPECT_RATIOS", "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS")
+            _rename(cfg, "MODEL.RPN.ANCHOR_SIZES", "MODEL.ANCHOR_GENERATOR.SIZES")
+            del cfg["MODEL"]["RETINANET"]["ANCHOR_SIZES"]
+            del cfg["MODEL"]["RETINANET"]["ANCHOR_ASPECT_RATIOS"]
+        del cfg["MODEL"]["RETINANET"]["ANCHOR_STRIDES"]
+
+    @classmethod
+    def downgrade(cls, cfg: CN) -> None:
+        super().downgrade(cfg)
+
+        _rename(cfg, "MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS", "MODEL.RPN.ANCHOR_ASPECT_RATIOS")
+        _rename(cfg, "MODEL.ANCHOR_GENERATOR.SIZES", "MODEL.RPN.ANCHOR_SIZES")
+        cfg.MODEL.RETINANET.ANCHOR_ASPECT_RATIOS = cfg.MODEL.RPN.ANCHOR_ASPECT_RATIOS
+        cfg.MODEL.RETINANET.ANCHOR_SIZES = cfg.MODEL.RPN.ANCHOR_SIZES
+        cfg.MODEL.RETINANET.ANCHOR_STRIDES = []  # this is not used anywhere in any version

detectron2/config/config.py

@@ -0,0 +1,265 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import functools
+import inspect
+import logging
+from fvcore.common.config import CfgNode as _CfgNode
+
+from detectron2.utils.file_io import PathManager
+
+
+class CfgNode(_CfgNode):
+    """
+    The same as `fvcore.common.config.CfgNode`, but different in:
+
+    1. Use unsafe yaml loading by default.
+       Note that this may lead to arbitrary code execution: you must not
+       load a config file from untrusted sources before manually inspecting
+       the content of the file.
+    2. Support config versioning.
+       When attempting to merge an old config, it will convert the old config automatically.
+
+    .. automethod:: clone
+    .. automethod:: freeze
+    .. automethod:: defrost
+    .. automethod:: is_frozen
+    .. automethod:: load_yaml_with_base
+    .. automethod:: merge_from_list
+    .. automethod:: merge_from_other_cfg
+    """
+
+    @classmethod
+    def _open_cfg(cls, filename):
+        return PathManager.open(filename, "r")
+
+    # Note that the default value of allow_unsafe is changed to True
+    def merge_from_file(self, cfg_filename: str, allow_unsafe: bool = True) -> None:
+        """
+        Load content from the given config file and merge it into self.
+
+        Args:
+            cfg_filename: config filename
+            allow_unsafe: allow unsafe yaml syntax
+        """
+        assert PathManager.isfile(cfg_filename), f"Config file '{cfg_filename}' does not exist!"
+        loaded_cfg = self.load_yaml_with_base(cfg_filename, allow_unsafe=allow_unsafe)
+        loaded_cfg = type(self)(loaded_cfg)
+
+        # defaults.py needs to import CfgNode
+        from .defaults import _C
+
+        latest_ver = _C.VERSION
+        assert (
+            latest_ver == self.VERSION
+        ), "CfgNode.merge_from_file is only allowed on a config object of latest version!"
+
+        logger = logging.getLogger(__name__)
+
+        loaded_ver = loaded_cfg.get("VERSION", None)
+        if loaded_ver is None:
+            from .compat import guess_version
+
+            loaded_ver = guess_version(loaded_cfg, cfg_filename)
+        assert loaded_ver <= self.VERSION, "Cannot merge a v{} config into a v{} config.".format(
+            loaded_ver, self.VERSION
+        )
+
+        if loaded_ver == self.VERSION:
+            self.merge_from_other_cfg(loaded_cfg)
+        else:
+            # compat.py needs to import CfgNode
+            from .compat import upgrade_config, downgrade_config
+
+            logger.warning(
+                "Loading an old v{} config file '{}' by automatically upgrading to v{}. "
+                "See docs/CHANGELOG.md for instructions to update your files.".format(
+                    loaded_ver, cfg_filename, self.VERSION
+                )
+            )
+            # To convert, first obtain a full config at an old version
+            old_self = downgrade_config(self, to_version=loaded_ver)
+            old_self.merge_from_other_cfg(loaded_cfg)
+            new_config = upgrade_config(old_self)
+            self.clear()
+            self.update(new_config)
+
+    def dump(self, *args, **kwargs):
+        """
+        Returns:
+            str: a yaml string representation of the config
+        """
+        # to make it show up in docs
+        return super().dump(*args, **kwargs)
+
+
+global_cfg = CfgNode()
+
+
+def get_cfg() -> CfgNode:
+    """
+    Get a copy of the default config.
+
+    Returns:
+        a detectron2 CfgNode instance.
+    """
+    from .defaults import _C
+
+    return _C.clone()
+
+
+def set_global_cfg(cfg: CfgNode) -> None:
+    """
+    Let the global config point to the given cfg.
+
+    Assume that the given "cfg" has the key "KEY", after calling
+    `set_global_cfg(cfg)`, the key can be accessed by:
+    ::
+        from detectron2.config import global_cfg
+        print(global_cfg.KEY)
+
+    By using a hacky global config, you can access these configs anywhere,
+    without having to pass the config object or the values deep into the code.
+    This is a hacky feature introduced for quick prototyping / research exploration.
+    """
+    global global_cfg
+    global_cfg.clear()
+    global_cfg.update(cfg)
+
+
+def configurable(init_func=None, *, from_config=None):
+    """
+    Decorate a function or a class's __init__ method so that it can be called
+    with a :class:`CfgNode` object using a :func:`from_config` function that translates
+    :class:`CfgNode` to arguments.
+
+    Examples:
+    ::
+        # Usage 1: Decorator on __init__:
+        class A:
+            @configurable
+            def __init__(self, a, b=2, c=3):
+                pass
+
+            @classmethod
+            def from_config(cls, cfg):   # 'cfg' must be the first argument
+                # Returns kwargs to be passed to __init__
+                return {"a": cfg.A, "b": cfg.B}
+
+        a1 = A(a=1, b=2)  # regular construction
+        a2 = A(cfg)       # construct with a cfg
+        a3 = A(cfg, b=3, c=4)  # construct with extra overwrite
+
+        # Usage 2: Decorator on any function. Needs an extra from_config argument:
+        @configurable(from_config=lambda cfg: {"a: cfg.A, "b": cfg.B})
+        def a_func(a, b=2, c=3):
+            pass
+
+        a1 = a_func(a=1, b=2)  # regular call
+        a2 = a_func(cfg)       # call with a cfg
+        a3 = a_func(cfg, b=3, c=4)  # call with extra overwrite
+
+    Args:
+        init_func (callable): a class's ``__init__`` method in usage 1. The
+            class must have a ``from_config`` classmethod which takes `cfg` as
+            the first argument.
+        from_config (callable): the from_config function in usage 2. It must take `cfg`
+            as its first argument.
+    """
+
+    if init_func is not None:
+        assert (
+            inspect.isfunction(init_func)
+            and from_config is None
+            and init_func.__name__ == "__init__"
+        ), "Incorrect use of @configurable. Check API documentation for examples."
+
+        @functools.wraps(init_func)
+        def wrapped(self, *args, **kwargs):
+            try:
+                from_config_func = type(self).from_config
+            except AttributeError as e:
+                raise AttributeError(
+                    "Class with @configurable must have a 'from_config' classmethod."
+                ) from e
+            if not inspect.ismethod(from_config_func):
+                raise TypeError("Class with @configurable must have a 'from_config' classmethod.")
+
+            if _called_with_cfg(*args, **kwargs):
+                explicit_args = _get_args_from_config(from_config_func, *args, **kwargs)
+                init_func(self, **explicit_args)
+            else:
+                init_func(self, *args, **kwargs)
+
+        return wrapped
+
+    else:
+        if from_config is None:
+            return configurable  # @configurable() is made equivalent to @configurable
+        assert inspect.isfunction(
+            from_config
+        ), "from_config argument of configurable must be a function!"
+
+        def wrapper(orig_func):
+            @functools.wraps(orig_func)
+            def wrapped(*args, **kwargs):
+                if _called_with_cfg(*args, **kwargs):
+                    explicit_args = _get_args_from_config(from_config, *args, **kwargs)
+                    return orig_func(**explicit_args)
+                else:
+                    return orig_func(*args, **kwargs)
+
+            wrapped.from_config = from_config
+            return wrapped
+
+        return wrapper
+
+
+def _get_args_from_config(from_config_func, *args, **kwargs):
+    """
+    Use `from_config` to obtain explicit arguments.
+
+    Returns:
+        dict: arguments to be used for cls.__init__
+    """
+    signature = inspect.signature(from_config_func)
+    if list(signature.parameters.keys())[0] != "cfg":
+        if inspect.isfunction(from_config_func):
+            name = from_config_func.__name__
+        else:
+            name = f"{from_config_func.__self__}.from_config"
+        raise TypeError(f"{name} must take 'cfg' as the first argument!")
+    support_var_arg = any(
+        param.kind in [param.VAR_POSITIONAL, param.VAR_KEYWORD]
+        for param in signature.parameters.values()
+    )
+    if support_var_arg:  # forward all arguments to from_config, if from_config accepts them
+        ret = from_config_func(*args, **kwargs)
+    else:
+        # forward supported arguments to from_config
+        supported_arg_names = set(signature.parameters.keys())
+        extra_kwargs = {}
+        for name in list(kwargs.keys()):
+            if name not in supported_arg_names:
+                extra_kwargs[name] = kwargs.pop(name)
+        ret = from_config_func(*args, **kwargs)
+        # forward the other arguments to __init__
+        ret.update(extra_kwargs)
+    return ret
+
+
+def _called_with_cfg(*args, **kwargs):
+    """
+    Returns:
+        bool: whether the arguments contain CfgNode and should be considered
+            forwarded to from_config.
+    """
+    from omegaconf import DictConfig
+
+    if len(args) and isinstance(args[0], (_CfgNode, DictConfig)):
+        return True
+    if isinstance(kwargs.pop("cfg", None), (_CfgNode, DictConfig)):
+        return True
+    # `from_config`'s first argument is forced to be "cfg".
+    # So the above check covers all cases.
+    return False

detectron2/config/defaults.py

@@ -0,0 +1,656 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+from .config import CfgNode as CN
+
+# NOTE: given the new config system
+# (https://detectron2.readthedocs.io/en/latest/tutorials/lazyconfigs.html),
+# we will stop adding new functionalities to default CfgNode.
+
+# -----------------------------------------------------------------------------
+# Convention about Training / Test specific parameters
+# -----------------------------------------------------------------------------
+# Whenever an argument can be either used for training or for testing, the
+# corresponding name will be post-fixed by a _TRAIN for a training parameter,
+# or _TEST for a test-specific parameter.
+# For example, the number of images during training will be
+# IMAGES_PER_BATCH_TRAIN, while the number of images for testing will be
+# IMAGES_PER_BATCH_TEST
+
+# -----------------------------------------------------------------------------
+# Config definition
+# -----------------------------------------------------------------------------
+
+_C = CN()
+
+# The version number, to upgrade from old configs to new ones if any
+# changes happen. It's recommended to keep a VERSION in your config file.
+_C.VERSION = 2
+
+_C.MODEL = CN()
+_C.MODEL.LOAD_PROPOSALS = False
+_C.MODEL.MASK_ON = False
+_C.MODEL.KEYPOINT_ON = False
+_C.MODEL.DEVICE = "cuda"
+_C.MODEL.META_ARCHITECTURE = "GeneralizedRCNN"
+
+# Path (a file path, or URL like detectron2://.., https://..) to a checkpoint file
+# to be loaded to the model. You can find available models in the model zoo.
+_C.MODEL.WEIGHTS = ""
+
+# Values to be used for image normalization (BGR order, since INPUT.FORMAT defaults to BGR).
+# To train on images of different number of channels, just set different mean & std.
+# Default values are the mean pixel value from ImageNet: [103.53, 116.28, 123.675]
+_C.MODEL.PIXEL_MEAN = [103.530, 116.280, 123.675]
+# When using pre-trained models in Detectron1 or any MSRA models,
+# std has been absorbed into its conv1 weights, so the std needs to be set 1.
+# Otherwise, you can use [57.375, 57.120, 58.395] (ImageNet std)
+_C.MODEL.PIXEL_STD = [1.0, 1.0, 1.0]
+
+
+# -----------------------------------------------------------------------------
+# INPUT
+# -----------------------------------------------------------------------------
+_C.INPUT = CN()
+# By default, {MIN,MAX}_SIZE options are used in transforms.ResizeShortestEdge.
+# Please refer to ResizeShortestEdge for detailed definition.
+# Size of the smallest side of the image during training
+_C.INPUT.MIN_SIZE_TRAIN = (800,)
+# Sample size of smallest side by choice or random selection from range give by
+# INPUT.MIN_SIZE_TRAIN
+_C.INPUT.MIN_SIZE_TRAIN_SAMPLING = "choice"
+# Maximum size of the side of the image during training
+_C.INPUT.MAX_SIZE_TRAIN = 1333
+# Size of the smallest side of the image during testing. Set to zero to disable resize in testing.
+_C.INPUT.MIN_SIZE_TEST = 800
+# Maximum size of the side of the image during testing
+_C.INPUT.MAX_SIZE_TEST = 1333
+# Mode for flipping images used in data augmentation during training
+# choose one of ["horizontal, "vertical", "none"]
+_C.INPUT.RANDOM_FLIP = "horizontal"
+
+# `True` if cropping is used for data augmentation during training
+_C.INPUT.CROP = CN({"ENABLED": False})
+# Cropping type. See documentation of `detectron2.data.transforms.RandomCrop` for explanation.
+_C.INPUT.CROP.TYPE = "relative_range"
+# Size of crop in range (0, 1] if CROP.TYPE is "relative" or "relative_range" and in number of
+# pixels if CROP.TYPE is "absolute"
+_C.INPUT.CROP.SIZE = [0.9, 0.9]
+
+
+# Whether the model needs RGB, YUV, HSV etc.
+# Should be one of the modes defined here, as we use PIL to read the image:
+# https://pillow.readthedocs.io/en/stable/handbook/concepts.html#concept-modes
+# with BGR being the one exception. One can set image format to BGR, we will
+# internally use RGB for conversion and flip the channels over
+_C.INPUT.FORMAT = "BGR"
+# The ground truth mask format that the model will use.
+# Mask R-CNN supports either "polygon" or "bitmask" as ground truth.
+_C.INPUT.MASK_FORMAT = "polygon"  # alternative: "bitmask"
+
+
+# -----------------------------------------------------------------------------
+# Dataset
+# -----------------------------------------------------------------------------
+_C.DATASETS = CN()
+# List of the dataset names for training. Must be registered in DatasetCatalog
+# Samples from these datasets will be merged and used as one dataset.
+_C.DATASETS.TRAIN = ()
+# List of the pre-computed proposal files for training, which must be consistent
+# with datasets listed in DATASETS.TRAIN.
+_C.DATASETS.PROPOSAL_FILES_TRAIN = ()
+# Number of top scoring precomputed proposals to keep for training
+_C.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TRAIN = 2000
+# List of the dataset names for testing. Must be registered in DatasetCatalog
+_C.DATASETS.TEST = ()
+# List of the pre-computed proposal files for test, which must be consistent
+# with datasets listed in DATASETS.TEST.
+_C.DATASETS.PROPOSAL_FILES_TEST = ()
+# Number of top scoring precomputed proposals to keep for test
+_C.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TEST = 1000
+
+# -----------------------------------------------------------------------------
+# DataLoader
+# -----------------------------------------------------------------------------
+_C.DATALOADER = CN()
+# Number of data loading threads
+_C.DATALOADER.NUM_WORKERS = 4
+# If True, each batch should contain only images for which the aspect ratio
+# is compatible. This groups portrait images together, and landscape images
+# are not batched with portrait images.
+_C.DATALOADER.ASPECT_RATIO_GROUPING = True
+# Options: TrainingSampler, RepeatFactorTrainingSampler
+_C.DATALOADER.SAMPLER_TRAIN = "TrainingSampler"
+# Repeat threshold for RepeatFactorTrainingSampler
+_C.DATALOADER.REPEAT_THRESHOLD = 0.0
+# if True, take square root when computing repeating factor
+_C.DATALOADER.REPEAT_SQRT = True
+# Tf True, when working on datasets that have instance annotations, the
+# training dataloader will filter out images without associated annotations
+_C.DATALOADER.FILTER_EMPTY_ANNOTATIONS = True
+
+# ---------------------------------------------------------------------------- #
+# Backbone options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.BACKBONE = CN()
+
+_C.MODEL.BACKBONE.NAME = "build_resnet_backbone"
+# Freeze the first several stages so they are not trained.
+# There are 5 stages in ResNet. The first is a convolution, and the following
+# stages are each group of residual blocks.
+_C.MODEL.BACKBONE.FREEZE_AT = 2
+
+
+# ---------------------------------------------------------------------------- #
+# FPN options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.FPN = CN()
+# Names of the input feature maps to be used by FPN
+# They must have contiguous power of 2 strides
+# e.g., ["res2", "res3", "res4", "res5"]
+_C.MODEL.FPN.IN_FEATURES = []
+_C.MODEL.FPN.OUT_CHANNELS = 256
+
+# Options: "" (no norm), "GN"
+_C.MODEL.FPN.NORM = ""
+
+# Types for fusing the FPN top-down and lateral features. Can be either "sum" or "avg"
+_C.MODEL.FPN.FUSE_TYPE = "sum"
+
+
+# ---------------------------------------------------------------------------- #
+# Proposal generator options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.PROPOSAL_GENERATOR = CN()
+# Current proposal generators include "RPN", "RRPN" and "PrecomputedProposals"
+_C.MODEL.PROPOSAL_GENERATOR.NAME = "RPN"
+# Proposal height and width both need to be greater than MIN_SIZE
+# (a the scale used during training or inference)
+_C.MODEL.PROPOSAL_GENERATOR.MIN_SIZE = 0
+
+
+# ---------------------------------------------------------------------------- #
+# Anchor generator options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ANCHOR_GENERATOR = CN()
+# The generator can be any name in the ANCHOR_GENERATOR registry
+_C.MODEL.ANCHOR_GENERATOR.NAME = "DefaultAnchorGenerator"
+# Anchor sizes (i.e. sqrt of area) in absolute pixels w.r.t. the network input.
+# Format: list[list[float]]. SIZES[i] specifies the list of sizes to use for
+# IN_FEATURES[i]; len(SIZES) must be equal to len(IN_FEATURES) or 1.
+# When len(SIZES) == 1, SIZES[0] is used for all IN_FEATURES.
+_C.MODEL.ANCHOR_GENERATOR.SIZES = [[32, 64, 128, 256, 512]]
+# Anchor aspect ratios. For each area given in `SIZES`, anchors with different aspect
+# ratios are generated by an anchor generator.
+# Format: list[list[float]]. ASPECT_RATIOS[i] specifies the list of aspect ratios (H/W)
+# to use for IN_FEATURES[i]; len(ASPECT_RATIOS) == len(IN_FEATURES) must be true,
+# or len(ASPECT_RATIOS) == 1 is true and aspect ratio list ASPECT_RATIOS[0] is used
+# for all IN_FEATURES.
+_C.MODEL.ANCHOR_GENERATOR.ASPECT_RATIOS = [[0.5, 1.0, 2.0]]
+# Anchor angles.
+# list[list[float]], the angle in degrees, for each input feature map.
+# ANGLES[i] specifies the list of angles for IN_FEATURES[i].
+_C.MODEL.ANCHOR_GENERATOR.ANGLES = [[-90, 0, 90]]
+# Relative offset between the center of the first anchor and the top-left corner of the image
+# Value has to be in [0, 1). Recommend to use 0.5, which means half stride.
+# The value is not expected to affect model accuracy.
+_C.MODEL.ANCHOR_GENERATOR.OFFSET = 0.0
+
+# ---------------------------------------------------------------------------- #
+# RPN options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.RPN = CN()
+_C.MODEL.RPN.HEAD_NAME = "StandardRPNHead"  # used by RPN_HEAD_REGISTRY
+
+# Names of the input feature maps to be used by RPN
+# e.g., ["p2", "p3", "p4", "p5", "p6"] for FPN
+_C.MODEL.RPN.IN_FEATURES = ["res4"]
+# Remove RPN anchors that go outside the image by BOUNDARY_THRESH pixels
+# Set to -1 or a large value, e.g. 100000, to disable pruning anchors
+_C.MODEL.RPN.BOUNDARY_THRESH = -1
+# IOU overlap ratios [BG_IOU_THRESHOLD, FG_IOU_THRESHOLD]
+# Minimum overlap required between an anchor and ground-truth box for the
+# (anchor, gt box) pair to be a positive example (IoU >= FG_IOU_THRESHOLD
+# ==> positive RPN example: 1)
+# Maximum overlap allowed between an anchor and ground-truth box for the
+# (anchor, gt box) pair to be a negative examples (IoU < BG_IOU_THRESHOLD
+# ==> negative RPN example: 0)
+# Anchors with overlap in between (BG_IOU_THRESHOLD <= IoU < FG_IOU_THRESHOLD)
+# are ignored (-1)
+_C.MODEL.RPN.IOU_THRESHOLDS = [0.3, 0.7]
+_C.MODEL.RPN.IOU_LABELS = [0, -1, 1]
+# Number of regions per image used to train RPN
+_C.MODEL.RPN.BATCH_SIZE_PER_IMAGE = 256
+# Target fraction of foreground (positive) examples per RPN minibatch
+_C.MODEL.RPN.POSITIVE_FRACTION = 0.5
+# Options are: "smooth_l1", "giou", "diou", "ciou"
+_C.MODEL.RPN.BBOX_REG_LOSS_TYPE = "smooth_l1"
+_C.MODEL.RPN.BBOX_REG_LOSS_WEIGHT = 1.0
+# Weights on (dx, dy, dw, dh) for normalizing RPN anchor regression targets
+_C.MODEL.RPN.BBOX_REG_WEIGHTS = (1.0, 1.0, 1.0, 1.0)
+# The transition point from L1 to L2 loss. Set to 0.0 to make the loss simply L1.
+_C.MODEL.RPN.SMOOTH_L1_BETA = 0.0
+_C.MODEL.RPN.LOSS_WEIGHT = 1.0
+# Number of top scoring RPN proposals to keep before applying NMS
+# When FPN is used, this is *per FPN level* (not total)
+_C.MODEL.RPN.PRE_NMS_TOPK_TRAIN = 12000
+_C.MODEL.RPN.PRE_NMS_TOPK_TEST = 6000
+# Number of top scoring RPN proposals to keep after applying NMS
+# When FPN is used, this limit is applied per level and then again to the union
+# of proposals from all levels
+# NOTE: When FPN is used, the meaning of this config is different from Detectron1.
+# It means per-batch topk in Detectron1, but per-image topk here.
+# See the "find_top_rpn_proposals" function for details.
+_C.MODEL.RPN.POST_NMS_TOPK_TRAIN = 2000
+_C.MODEL.RPN.POST_NMS_TOPK_TEST = 1000
+# NMS threshold used on RPN proposals
+_C.MODEL.RPN.NMS_THRESH = 0.7
+# Set this to -1 to use the same number of output channels as input channels.
+_C.MODEL.RPN.CONV_DIMS = [-1]
+
+# ---------------------------------------------------------------------------- #
+# ROI HEADS options
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_HEADS = CN()
+_C.MODEL.ROI_HEADS.NAME = "Res5ROIHeads"
+# Number of foreground classes
+_C.MODEL.ROI_HEADS.NUM_CLASSES = 80
+# Names of the input feature maps to be used by ROI heads
+# Currently all heads (box, mask, ...) use the same input feature map list
+# e.g., ["p2", "p3", "p4", "p5"] is commonly used for FPN
+_C.MODEL.ROI_HEADS.IN_FEATURES = ["res4"]
+# IOU overlap ratios [IOU_THRESHOLD]
+# Overlap threshold for an RoI to be considered background (if < IOU_THRESHOLD)
+# Overlap threshold for an RoI to be considered foreground (if >= IOU_THRESHOLD)
+_C.MODEL.ROI_HEADS.IOU_THRESHOLDS = [0.5]
+_C.MODEL.ROI_HEADS.IOU_LABELS = [0, 1]
+# RoI minibatch size *per image* (number of regions of interest [ROIs]) during training
+# Total number of RoIs per training minibatch =
+#   ROI_HEADS.BATCH_SIZE_PER_IMAGE * SOLVER.IMS_PER_BATCH
+# E.g., a common configuration is: 512 * 16 = 8192
+_C.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 512
+# Target fraction of RoI minibatch that is labeled foreground (i.e. class > 0)
+_C.MODEL.ROI_HEADS.POSITIVE_FRACTION = 0.25
+
+# Only used on test mode
+
+# Minimum score threshold (assuming scores in a [0, 1] range); a value chosen to
+# balance obtaining high recall with not having too many low precision
+# detections that will slow down inference post processing steps (like NMS)
+# A default threshold of 0.0 increases AP by ~0.2-0.3 but significantly slows down
+# inference.
+_C.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.05
+# Overlap threshold used for non-maximum suppression (suppress boxes with
+# IoU >= this threshold)
+_C.MODEL.ROI_HEADS.NMS_THRESH_TEST = 0.5
+# If True, augment proposals with ground-truth boxes before sampling proposals to
+# train ROI heads.
+_C.MODEL.ROI_HEADS.PROPOSAL_APPEND_GT = True
+
+# ---------------------------------------------------------------------------- #
+# Box Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_BOX_HEAD = CN()
+# C4 don't use head name option
+# Options for non-C4 models: FastRCNNConvFCHead,
+_C.MODEL.ROI_BOX_HEAD.NAME = ""
+# Options are: "smooth_l1", "giou", "diou", "ciou"
+_C.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_TYPE = "smooth_l1"
+# The final scaling coefficient on the box regression loss, used to balance the magnitude of its
+# gradients with other losses in the model. See also `MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT`.
+_C.MODEL.ROI_BOX_HEAD.BBOX_REG_LOSS_WEIGHT = 1.0
+# Default weights on (dx, dy, dw, dh) for normalizing bbox regression targets
+# These are empirically chosen to approximately lead to unit variance targets
+_C.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS = (10.0, 10.0, 5.0, 5.0)
+# The transition point from L1 to L2 loss. Set to 0.0 to make the loss simply L1.
+_C.MODEL.ROI_BOX_HEAD.SMOOTH_L1_BETA = 0.0
+_C.MODEL.ROI_BOX_HEAD.POOLER_RESOLUTION = 14
+_C.MODEL.ROI_BOX_HEAD.POOLER_SAMPLING_RATIO = 0
+# Type of pooling operation applied to the incoming feature map for each RoI
+_C.MODEL.ROI_BOX_HEAD.POOLER_TYPE = "ROIAlignV2"
+
+_C.MODEL.ROI_BOX_HEAD.NUM_FC = 0
+# Hidden layer dimension for FC layers in the RoI box head
+_C.MODEL.ROI_BOX_HEAD.FC_DIM = 1024
+_C.MODEL.ROI_BOX_HEAD.NUM_CONV = 0
+# Channel dimension for Conv layers in the RoI box head
+_C.MODEL.ROI_BOX_HEAD.CONV_DIM = 256
+# Normalization method for the convolution layers.
+# Options: "" (no norm), "GN", "SyncBN".
+_C.MODEL.ROI_BOX_HEAD.NORM = ""
+# Whether to use class agnostic for bbox regression
+_C.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG = False
+# If true, RoI heads use bounding boxes predicted by the box head rather than proposal boxes.
+_C.MODEL.ROI_BOX_HEAD.TRAIN_ON_PRED_BOXES = False
+
+# Federated loss can be used to improve the training of LVIS
+_C.MODEL.ROI_BOX_HEAD.USE_FED_LOSS = False
+# Sigmoid cross entrophy is used with federated loss
+_C.MODEL.ROI_BOX_HEAD.USE_SIGMOID_CE = False
+# The power value applied to image_count when calcualting frequency weight
+_C.MODEL.ROI_BOX_HEAD.FED_LOSS_FREQ_WEIGHT_POWER = 0.5
+# Number of classes to keep in total
+_C.MODEL.ROI_BOX_HEAD.FED_LOSS_NUM_CLASSES = 50
+
+# ---------------------------------------------------------------------------- #
+# Cascaded Box Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_BOX_CASCADE_HEAD = CN()
+# The number of cascade stages is implicitly defined by the length of the following two configs.
+_C.MODEL.ROI_BOX_CASCADE_HEAD.BBOX_REG_WEIGHTS = (
+    (10.0, 10.0, 5.0, 5.0),
+    (20.0, 20.0, 10.0, 10.0),
+    (30.0, 30.0, 15.0, 15.0),
+)
+_C.MODEL.ROI_BOX_CASCADE_HEAD.IOUS = (0.5, 0.6, 0.7)
+
+
+# ---------------------------------------------------------------------------- #
+# Mask Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_MASK_HEAD = CN()
+_C.MODEL.ROI_MASK_HEAD.NAME = "MaskRCNNConvUpsampleHead"
+_C.MODEL.ROI_MASK_HEAD.POOLER_RESOLUTION = 14
+_C.MODEL.ROI_MASK_HEAD.POOLER_SAMPLING_RATIO = 0
+_C.MODEL.ROI_MASK_HEAD.NUM_CONV = 0  # The number of convs in the mask head
+_C.MODEL.ROI_MASK_HEAD.CONV_DIM = 256
+# Normalization method for the convolution layers.
+# Options: "" (no norm), "GN", "SyncBN".
+_C.MODEL.ROI_MASK_HEAD.NORM = ""
+# Whether to use class agnostic for mask prediction
+_C.MODEL.ROI_MASK_HEAD.CLS_AGNOSTIC_MASK = False
+# Type of pooling operation applied to the incoming feature map for each RoI
+_C.MODEL.ROI_MASK_HEAD.POOLER_TYPE = "ROIAlignV2"
+
+
+# ---------------------------------------------------------------------------- #
+# Keypoint Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.ROI_KEYPOINT_HEAD = CN()
+_C.MODEL.ROI_KEYPOINT_HEAD.NAME = "KRCNNConvDeconvUpsampleHead"
+_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_RESOLUTION = 14
+_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_SAMPLING_RATIO = 0
+_C.MODEL.ROI_KEYPOINT_HEAD.CONV_DIMS = tuple(512 for _ in range(8))
+_C.MODEL.ROI_KEYPOINT_HEAD.NUM_KEYPOINTS = 17  # 17 is the number of keypoints in COCO.
+
+# Images with too few (or no) keypoints are excluded from training.
+_C.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE = 1
+# Normalize by the total number of visible keypoints in the minibatch if True.
+# Otherwise, normalize by the total number of keypoints that could ever exist
+# in the minibatch.
+# The keypoint softmax loss is only calculated on visible keypoints.
+# Since the number of visible keypoints can vary significantly between
+# minibatches, this has the effect of up-weighting the importance of
+# minibatches with few visible keypoints. (Imagine the extreme case of
+# only one visible keypoint versus N: in the case of N, each one
+# contributes 1/N to the gradient compared to the single keypoint
+# determining the gradient direction). Instead, we can normalize the
+# loss by the total number of keypoints, if it were the case that all
+# keypoints were visible in a full minibatch. (Returning to the example,
+# this means that the one visible keypoint contributes as much as each
+# of the N keypoints.)
+_C.MODEL.ROI_KEYPOINT_HEAD.NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS = True
+# Multi-task loss weight to use for keypoints
+# Recommended values:
+#   - use 1.0 if NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS is True
+#   - use 4.0 if NORMALIZE_LOSS_BY_VISIBLE_KEYPOINTS is False
+_C.MODEL.ROI_KEYPOINT_HEAD.LOSS_WEIGHT = 1.0
+# Type of pooling operation applied to the incoming feature map for each RoI
+_C.MODEL.ROI_KEYPOINT_HEAD.POOLER_TYPE = "ROIAlignV2"
+
+# ---------------------------------------------------------------------------- #
+# Semantic Segmentation Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.SEM_SEG_HEAD = CN()
+_C.MODEL.SEM_SEG_HEAD.NAME = "SemSegFPNHead"
+_C.MODEL.SEM_SEG_HEAD.IN_FEATURES = ["p2", "p3", "p4", "p5"]
+# Label in the semantic segmentation ground truth that is ignored, i.e., no loss is calculated for
+# the correposnding pixel.
+_C.MODEL.SEM_SEG_HEAD.IGNORE_VALUE = 255
+# Number of classes in the semantic segmentation head
+_C.MODEL.SEM_SEG_HEAD.NUM_CLASSES = 54
+# Number of channels in the 3x3 convs inside semantic-FPN heads.
+_C.MODEL.SEM_SEG_HEAD.CONVS_DIM = 128
+# Outputs from semantic-FPN heads are up-scaled to the COMMON_STRIDE stride.
+_C.MODEL.SEM_SEG_HEAD.COMMON_STRIDE = 4
+# Normalization method for the convolution layers. Options: "" (no norm), "GN".
+_C.MODEL.SEM_SEG_HEAD.NORM = "GN"
+_C.MODEL.SEM_SEG_HEAD.LOSS_WEIGHT = 1.0
+
+_C.MODEL.PANOPTIC_FPN = CN()
+# Scaling of all losses from instance detection / segmentation head.
+_C.MODEL.PANOPTIC_FPN.INSTANCE_LOSS_WEIGHT = 1.0
+
+# options when combining instance & semantic segmentation outputs
+_C.MODEL.PANOPTIC_FPN.COMBINE = CN({"ENABLED": True})  # "COMBINE.ENABLED" is deprecated & not used
+_C.MODEL.PANOPTIC_FPN.COMBINE.OVERLAP_THRESH = 0.5
+_C.MODEL.PANOPTIC_FPN.COMBINE.STUFF_AREA_LIMIT = 4096
+_C.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = 0.5
+
+
+# ---------------------------------------------------------------------------- #
+# RetinaNet Head
+# ---------------------------------------------------------------------------- #
+_C.MODEL.RETINANET = CN()
+
+# This is the number of foreground classes.
+_C.MODEL.RETINANET.NUM_CLASSES = 80
+
+_C.MODEL.RETINANET.IN_FEATURES = ["p3", "p4", "p5", "p6", "p7"]
+
+# Convolutions to use in the cls and bbox tower
+# NOTE: this doesn't include the last conv for logits
+_C.MODEL.RETINANET.NUM_CONVS = 4
+
+# IoU overlap ratio [bg, fg] for labeling anchors.
+# Anchors with < bg are labeled negative (0)
+# Anchors  with >= bg and < fg are ignored (-1)
+# Anchors with >= fg are labeled positive (1)
+_C.MODEL.RETINANET.IOU_THRESHOLDS = [0.4, 0.5]
+_C.MODEL.RETINANET.IOU_LABELS = [0, -1, 1]
+
+# Prior prob for rare case (i.e. foreground) at the beginning of training.
+# This is used to set the bias for the logits layer of the classifier subnet.
+# This improves training stability in the case of heavy class imbalance.
+_C.MODEL.RETINANET.PRIOR_PROB = 0.01
+
+# Inference cls score threshold, only anchors with score > INFERENCE_TH are
+# considered for inference (to improve speed)
+_C.MODEL.RETINANET.SCORE_THRESH_TEST = 0.05
+# Select topk candidates before NMS
+_C.MODEL.RETINANET.TOPK_CANDIDATES_TEST = 1000
+_C.MODEL.RETINANET.NMS_THRESH_TEST = 0.5
+
+# Weights on (dx, dy, dw, dh) for normalizing Retinanet anchor regression targets
+_C.MODEL.RETINANET.BBOX_REG_WEIGHTS = (1.0, 1.0, 1.0, 1.0)
+
+# Loss parameters
+_C.MODEL.RETINANET.FOCAL_LOSS_GAMMA = 2.0
+_C.MODEL.RETINANET.FOCAL_LOSS_ALPHA = 0.25
+_C.MODEL.RETINANET.SMOOTH_L1_LOSS_BETA = 0.1
+# Options are: "smooth_l1", "giou", "diou", "ciou"
+_C.MODEL.RETINANET.BBOX_REG_LOSS_TYPE = "smooth_l1"
+
+# One of BN, SyncBN, FrozenBN, GN
+# Only supports GN until unshared norm is implemented
+_C.MODEL.RETINANET.NORM = ""
+
+
+# ---------------------------------------------------------------------------- #
+# ResNe[X]t options (ResNets = {ResNet, ResNeXt}
+# Note that parts of a resnet may be used for both the backbone and the head
+# These options apply to both
+# ---------------------------------------------------------------------------- #
+_C.MODEL.RESNETS = CN()
+
+_C.MODEL.RESNETS.DEPTH = 50
+_C.MODEL.RESNETS.OUT_FEATURES = ["res4"]  # res4 for C4 backbone, res2..5 for FPN backbone
+
+# Number of groups to use; 1 ==> ResNet; > 1 ==> ResNeXt
+_C.MODEL.RESNETS.NUM_GROUPS = 1
+
+# Options: FrozenBN, GN, "SyncBN", "BN"
+_C.MODEL.RESNETS.NORM = "FrozenBN"
+
+# Baseline width of each group.
+# Scaling this parameters will scale the width of all bottleneck layers.
+_C.MODEL.RESNETS.WIDTH_PER_GROUP = 64
+
+# Place the stride 2 conv on the 1x1 filter
+# Use True only for the original MSRA ResNet; use False for C2 and Torch models
+_C.MODEL.RESNETS.STRIDE_IN_1X1 = True
+
+# Apply dilation in stage "res5"
+_C.MODEL.RESNETS.RES5_DILATION = 1
+
+# Output width of res2. Scaling this parameters will scale the width of all 1x1 convs in ResNet
+# For R18 and R34, this needs to be set to 64
+_C.MODEL.RESNETS.RES2_OUT_CHANNELS = 256
+_C.MODEL.RESNETS.STEM_OUT_CHANNELS = 64
+
+# Apply Deformable Convolution in stages
+# Specify if apply deform_conv on Res2, Res3, Res4, Res5
+_C.MODEL.RESNETS.DEFORM_ON_PER_STAGE = [False, False, False, False]
+# Use True to use modulated deform_conv (DeformableV2, https://arxiv.org/abs/1811.11168);
+# Use False for DeformableV1.
+_C.MODEL.RESNETS.DEFORM_MODULATED = False
+# Number of groups in deformable conv.
+_C.MODEL.RESNETS.DEFORM_NUM_GROUPS = 1
+
+
+# ---------------------------------------------------------------------------- #
+# Solver
+# ---------------------------------------------------------------------------- #
+_C.SOLVER = CN()
+
+# Options: WarmupMultiStepLR, WarmupCosineLR.
+# See detectron2/solver/build.py for definition.
+_C.SOLVER.LR_SCHEDULER_NAME = "WarmupMultiStepLR"
+
+_C.SOLVER.MAX_ITER = 40000
+
+_C.SOLVER.BASE_LR = 0.001
+# The end lr, only used by WarmupCosineLR
+_C.SOLVER.BASE_LR_END = 0.0
+
+_C.SOLVER.MOMENTUM = 0.9
+
+_C.SOLVER.NESTEROV = False
+
+_C.SOLVER.WEIGHT_DECAY = 0.0001
+# The weight decay that's applied to parameters of normalization layers
+# (typically the affine transformation)
+_C.SOLVER.WEIGHT_DECAY_NORM = 0.0
+
+_C.SOLVER.GAMMA = 0.1
+# The iteration number to decrease learning rate by GAMMA.
+_C.SOLVER.STEPS = (30000,)
+# Number of decays in WarmupStepWithFixedGammaLR schedule
+_C.SOLVER.NUM_DECAYS = 3
+
+_C.SOLVER.WARMUP_FACTOR = 1.0 / 1000
+_C.SOLVER.WARMUP_ITERS = 1000
+_C.SOLVER.WARMUP_METHOD = "linear"
+# Whether to rescale the interval for the learning schedule after warmup
+_C.SOLVER.RESCALE_INTERVAL = False
+
+# Save a checkpoint after every this number of iterations
+_C.SOLVER.CHECKPOINT_PERIOD = 5000
+
+# Number of images per batch across all machines. This is also the number
+# of training images per step (i.e. per iteration). If we use 16 GPUs
+# and IMS_PER_BATCH = 32, each GPU will see 2 images per batch.
+# May be adjusted automatically if REFERENCE_WORLD_SIZE is set.
+_C.SOLVER.IMS_PER_BATCH = 16
+
+# The reference number of workers (GPUs) this config is meant to train with.
+# It takes no effect when set to 0.
+# With a non-zero value, it will be used by DefaultTrainer to compute a desired
+# per-worker batch size, and then scale the other related configs (total batch size,
+# learning rate, etc) to match the per-worker batch size.
+# See documentation of `DefaultTrainer.auto_scale_workers` for details:
+_C.SOLVER.REFERENCE_WORLD_SIZE = 0
+
+# Detectron v1 (and previous detection code) used a 2x higher LR and 0 WD for
+# biases. This is not useful (at least for recent models). You should avoid
+# changing these and they exist only to reproduce Detectron v1 training if
+# desired.
+_C.SOLVER.BIAS_LR_FACTOR = 1.0
+_C.SOLVER.WEIGHT_DECAY_BIAS = None  # None means following WEIGHT_DECAY
+
+# Gradient clipping
+_C.SOLVER.CLIP_GRADIENTS = CN({"ENABLED": False})
+# Type of gradient clipping, currently 2 values are supported:
+# - "value": the absolute values of elements of each gradients are clipped
+# - "norm": the norm of the gradient for each parameter is clipped thus
+#   affecting all elements in the parameter
+_C.SOLVER.CLIP_GRADIENTS.CLIP_TYPE = "value"
+# Maximum absolute value used for clipping gradients
+_C.SOLVER.CLIP_GRADIENTS.CLIP_VALUE = 1.0
+# Floating point number p for L-p norm to be used with the "norm"
+# gradient clipping type; for L-inf, please specify .inf
+_C.SOLVER.CLIP_GRADIENTS.NORM_TYPE = 2.0
+
+# Enable automatic mixed precision for training
+# Note that this does not change model's inference behavior.
+# To use AMP in inference, run inference under autocast()
+_C.SOLVER.AMP = CN({"ENABLED": False})
+
+# ---------------------------------------------------------------------------- #
+# Specific test options
+# ---------------------------------------------------------------------------- #
+_C.TEST = CN()
+# For end-to-end tests to verify the expected accuracy.
+# Each item is [task, metric, value, tolerance]
+# e.g.: [['bbox', 'AP', 38.5, 0.2]]
+_C.TEST.EXPECTED_RESULTS = []
+# The period (in terms of steps) to evaluate the model during training.
+# Set to 0 to disable.
+_C.TEST.EVAL_PERIOD = 0
+# The sigmas used to calculate keypoint OKS. See http://cocodataset.org/#keypoints-eval
+# When empty, it will use the defaults in COCO.
+# Otherwise it should be a list[float] with the same length as ROI_KEYPOINT_HEAD.NUM_KEYPOINTS.
+_C.TEST.KEYPOINT_OKS_SIGMAS = []
+# Maximum number of detections to return per image during inference (100 is
+# based on the limit established for the COCO dataset).
+_C.TEST.DETECTIONS_PER_IMAGE = 100
+
+_C.TEST.AUG = CN({"ENABLED": False})
+_C.TEST.AUG.MIN_SIZES = (400, 500, 600, 700, 800, 900, 1000, 1100, 1200)
+_C.TEST.AUG.MAX_SIZE = 4000
+_C.TEST.AUG.FLIP = True
+
+_C.TEST.PRECISE_BN = CN({"ENABLED": False})
+_C.TEST.PRECISE_BN.NUM_ITER = 200
+
+# ---------------------------------------------------------------------------- #
+# Misc options
+# ---------------------------------------------------------------------------- #
+# Directory where output files are written
+_C.OUTPUT_DIR = "./output"
+# Set seed to negative to fully randomize everything.
+# Set seed to positive to use a fixed seed. Note that a fixed seed increases
+# reproducibility but does not guarantee fully deterministic behavior.
+# Disabling all parallelism further increases reproducibility.
+_C.SEED = -1
+# Benchmark different cudnn algorithms.
+# If input images have very different sizes, this option will have large overhead
+# for about 10k iterations. It usually hurts total time, but can benefit for certain models.
+# If input images have the same or similar sizes, benchmark is often helpful.
+_C.CUDNN_BENCHMARK = False
+# Option to set PyTorch matmul and CuDNN's float32 precision. When set to non-empty string,
+# the corresponding precision ("highest", "high" or "medium") will be used. The highest
+# precision will effectively disable tf32.
+_C.FLOAT32_PRECISION = ""
+# The period (in terms of steps) for minibatch visualization at train time.
+# Set to 0 to disable.
+_C.VIS_PERIOD = 0
+
+# global config is for quick hack purposes.
+# You can set them in command line or config files,
+# and access it with:
+#
+# from detectron2.config import global_cfg
+# print(global_cfg.HACK)
+#
+# Do not commit any configs into it.
+_C.GLOBAL = CN()
+_C.GLOBAL.HACK = 1.0

detectron2/config/instantiate.py

@@ -0,0 +1,88 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import collections.abc as abc
+import dataclasses
+import logging
+from typing import Any
+
+from detectron2.utils.registry import _convert_target_to_string, locate
+
+__all__ = ["dump_dataclass", "instantiate"]
+
+
+def dump_dataclass(obj: Any):
+    """
+    Dump a dataclass recursively into a dict that can be later instantiated.
+
+    Args:
+        obj: a dataclass object
+
+    Returns:
+        dict
+    """
+    assert dataclasses.is_dataclass(obj) and not isinstance(
+        obj, type
+    ), "dump_dataclass() requires an instance of a dataclass."
+    ret = {"_target_": _convert_target_to_string(type(obj))}
+    for f in dataclasses.fields(obj):
+        v = getattr(obj, f.name)
+        if dataclasses.is_dataclass(v):
+            v = dump_dataclass(v)
+        if isinstance(v, (list, tuple)):
+            v = [dump_dataclass(x) if dataclasses.is_dataclass(x) else x for x in v]
+        ret[f.name] = v
+    return ret
+
+
+def instantiate(cfg):
+    """
+    Recursively instantiate objects defined in dictionaries by
+    "_target_" and arguments.
+
+    Args:
+        cfg: a dict-like object with "_target_" that defines the caller, and
+            other keys that define the arguments
+
+    Returns:
+        object instantiated by cfg
+    """
+    from omegaconf import ListConfig, DictConfig, OmegaConf
+
+    if isinstance(cfg, ListConfig):
+        lst = [instantiate(x) for x in cfg]
+        return ListConfig(lst, flags={"allow_objects": True})
+    if isinstance(cfg, list):
+        # Specialize for list, because many classes take
+        # list[objects] as arguments, such as ResNet, DatasetMapper
+        return [instantiate(x) for x in cfg]
+
+    # If input is a DictConfig backed by dataclasses (i.e. omegaconf's structured config),
+    # instantiate it to the actual dataclass.
+    if isinstance(cfg, DictConfig) and dataclasses.is_dataclass(cfg._metadata.object_type):
+        return OmegaConf.to_object(cfg)
+
+    if isinstance(cfg, abc.Mapping) and "_target_" in cfg:
+        # conceptually equivalent to hydra.utils.instantiate(cfg) with _convert_=all,
+        # but faster: https://github.com/facebookresearch/hydra/issues/1200
+        cfg = {k: instantiate(v) for k, v in cfg.items()}
+        cls = cfg.pop("_target_")
+        cls = instantiate(cls)
+
+        if isinstance(cls, str):
+            cls_name = cls
+            cls = locate(cls_name)
+            assert cls is not None, cls_name
+        else:
+            try:
+                cls_name = cls.__module__ + "." + cls.__qualname__
+            except Exception:
+                # target could be anything, so the above could fail
+                cls_name = str(cls)
+        assert callable(cls), f"_target_ {cls} does not define a callable object"
+        try:
+            return cls(**cfg)
+        except TypeError:
+            logger = logging.getLogger(__name__)
+            logger.error(f"Error when instantiating {cls_name}!")
+            raise
+    return cfg  # return as-is if don't know what to do

detectron2/config/lazy.py

@@ -0,0 +1,436 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import ast
+import builtins
+import collections.abc as abc
+import importlib
+import inspect
+import logging
+import os
+import uuid
+from contextlib import contextmanager
+from copy import deepcopy
+from dataclasses import is_dataclass
+from typing import List, Tuple, Union
+import cloudpickle
+import yaml
+from omegaconf import DictConfig, ListConfig, OmegaConf, SCMode
+
+from detectron2.utils.file_io import PathManager
+from detectron2.utils.registry import _convert_target_to_string
+
+__all__ = ["LazyCall", "LazyConfig"]
+
+
+class LazyCall:
+    """
+    Wrap a callable so that when it's called, the call will not be executed,
+    but returns a dict that describes the call.
+
+    LazyCall object has to be called with only keyword arguments. Positional
+    arguments are not yet supported.
+
+    Examples:
+    ::
+        from detectron2.config import instantiate, LazyCall
+
+        layer_cfg = LazyCall(nn.Conv2d)(in_channels=32, out_channels=32)
+        layer_cfg.out_channels = 64   # can edit it afterwards
+        layer = instantiate(layer_cfg)
+    """
+
+    def __init__(self, target):
+        if not (callable(target) or isinstance(target, (str, abc.Mapping))):
+            raise TypeError(
+                f"target of LazyCall must be a callable or defines a callable! Got {target}"
+            )
+        self._target = target
+
+    def __call__(self, **kwargs):
+        if is_dataclass(self._target):
+            # omegaconf object cannot hold dataclass type
+            # https://github.com/omry/omegaconf/issues/784
+            target = _convert_target_to_string(self._target)
+        else:
+            target = self._target
+        kwargs["_target_"] = target
+
+        return DictConfig(content=kwargs, flags={"allow_objects": True})
+
+
+def _visit_dict_config(cfg, func):
+    """
+    Apply func recursively to all DictConfig in cfg.
+    """
+    if isinstance(cfg, DictConfig):
+        func(cfg)
+        for v in cfg.values():
+            _visit_dict_config(v, func)
+    elif isinstance(cfg, ListConfig):
+        for v in cfg:
+            _visit_dict_config(v, func)
+
+
+def _validate_py_syntax(filename):
+    # see also https://github.com/open-mmlab/mmcv/blob/master/mmcv/utils/config.py
+    with PathManager.open(filename, "r") as f:
+        content = f.read()
+    try:
+        ast.parse(content)
+    except SyntaxError as e:
+        raise SyntaxError(f"Config file {filename} has syntax error!") from e
+
+
+def _cast_to_config(obj):
+    # if given a dict, return DictConfig instead
+    if isinstance(obj, dict):
+        return DictConfig(obj, flags={"allow_objects": True})
+    return obj
+
+
+_CFG_PACKAGE_NAME = "detectron2._cfg_loader"
+"""
+A namespace to put all imported config into.
+"""
+
+
+def _random_package_name(filename):
+    # generate a random package name when loading config files
+    return _CFG_PACKAGE_NAME + str(uuid.uuid4())[:4] + "." + os.path.basename(filename)
+
+
+@contextmanager
+def _patch_import():
+    """
+    Enhance relative import statements in config files, so that they:
+    1. locate files purely based on relative location, regardless of packages.
+       e.g. you can import file without having __init__
+    2. do not cache modules globally; modifications of module states has no side effect
+    3. support other storage system through PathManager, so config files can be in the cloud
+    4. imported dict are turned into omegaconf.DictConfig automatically
+    """
+    old_import = builtins.__import__
+
+    def find_relative_file(original_file, relative_import_path, level):
+        # NOTE: "from . import x" is not handled. Because then it's unclear
+        # if such import should produce `x` as a python module or DictConfig.
+        # This can be discussed further if needed.
+        relative_import_err = """
+Relative import of directories is not allowed within config files.
+Within a config file, relative import can only import other config files.
+""".replace(
+            "\n", " "
+        )
+        if not len(relative_import_path):
+            raise ImportError(relative_import_err)
+
+        cur_file = os.path.dirname(original_file)
+        for _ in range(level - 1):
+            cur_file = os.path.dirname(cur_file)
+        cur_name = relative_import_path.lstrip(".")
+        for part in cur_name.split("."):
+            cur_file = os.path.join(cur_file, part)
+        if not cur_file.endswith(".py"):
+            cur_file += ".py"
+        if not PathManager.isfile(cur_file):
+            cur_file_no_suffix = cur_file[: -len(".py")]
+            if PathManager.isdir(cur_file_no_suffix):
+                raise ImportError(f"Cannot import from {cur_file_no_suffix}." + relative_import_err)
+            else:
+                raise ImportError(
+                    f"Cannot import name {relative_import_path} from "
+                    f"{original_file}: {cur_file} does not exist."
+                )
+        return cur_file
+
+    def new_import(name, globals=None, locals=None, fromlist=(), level=0):
+        if (
+            # Only deal with relative imports inside config files
+            level != 0
+            and globals is not None
+            and (globals.get("__package__", "") or "").startswith(_CFG_PACKAGE_NAME)
+        ):
+            cur_file = find_relative_file(globals["__file__"], name, level)
+            _validate_py_syntax(cur_file)
+            spec = importlib.machinery.ModuleSpec(
+                _random_package_name(cur_file), None, origin=cur_file
+            )
+            module = importlib.util.module_from_spec(spec)
+            module.__file__ = cur_file
+            with PathManager.open(cur_file) as f:
+                content = f.read()
+            exec(compile(content, cur_file, "exec"), module.__dict__)
+            for name in fromlist:  # turn imported dict into DictConfig automatically
+                val = _cast_to_config(module.__dict__[name])
+                module.__dict__[name] = val
+            return module
+        return old_import(name, globals, locals, fromlist=fromlist, level=level)
+
+    builtins.__import__ = new_import
+    yield new_import
+    builtins.__import__ = old_import
+
+
+class LazyConfig:
+    """
+    Provide methods to save, load, and overrides an omegaconf config object
+    which may contain definition of lazily-constructed objects.
+    """
+
+    @staticmethod
+    def load_rel(filename: str, keys: Union[None, str, Tuple[str, ...]] = None):
+        """
+        Similar to :meth:`load()`, but load path relative to the caller's
+        source file.
+
+        This has the same functionality as a relative import, except that this method
+        accepts filename as a string, so more characters are allowed in the filename.
+        """
+        caller_frame = inspect.stack()[1]
+        caller_fname = caller_frame[0].f_code.co_filename
+        assert caller_fname != "<string>", "load_rel Unable to find caller"
+        caller_dir = os.path.dirname(caller_fname)
+        filename = os.path.join(caller_dir, filename)
+        return LazyConfig.load(filename, keys)
+
+    @staticmethod
+    def load(filename: str, keys: Union[None, str, Tuple[str, ...]] = None):
+        """
+        Load a config file.
+
+        Args:
+            filename: absolute path or relative path w.r.t. the current working directory
+            keys: keys to load and return. If not given, return all keys
+                (whose values are config objects) in a dict.
+        """
+        has_keys = keys is not None
+        filename = filename.replace("/./", "/")  # redundant
+        if os.path.splitext(filename)[1] not in [".py", ".yaml", ".yml"]:
+            raise ValueError(f"Config file {filename} has to be a python or yaml file.")
+        if filename.endswith(".py"):
+            _validate_py_syntax(filename)
+
+            with _patch_import():
+                # Record the filename
+                module_namespace = {
+                    "__file__": filename,
+                    "__package__": _random_package_name(filename),
+                }
+                with PathManager.open(filename) as f:
+                    content = f.read()
+                # Compile first with filename to:
+                # 1. make filename appears in stacktrace
+                # 2. make load_rel able to find its parent's (possibly remote) location
+                exec(compile(content, filename, "exec"), module_namespace)
+
+            ret = module_namespace
+        else:
+            with PathManager.open(filename) as f:
+                obj = yaml.unsafe_load(f)
+            ret = OmegaConf.create(obj, flags={"allow_objects": True})
+
+        if has_keys:
+            if isinstance(keys, str):
+                return _cast_to_config(ret[keys])
+            else:
+                return tuple(_cast_to_config(ret[a]) for a in keys)
+        else:
+            if filename.endswith(".py"):
+                # when not specified, only load those that are config objects
+                ret = DictConfig(
+                    {
+                        name: _cast_to_config(value)
+                        for name, value in ret.items()
+                        if isinstance(value, (DictConfig, ListConfig, dict))
+                        and not name.startswith("_")
+                    },
+                    flags={"allow_objects": True},
+                )
+            return ret
+
+    @staticmethod
+    def save(cfg, filename: str):
+        """
+        Save a config object to a yaml file.
+        Note that when the config dictionary contains complex objects (e.g. lambda),
+        it can't be saved to yaml. In that case we will print an error and
+        attempt to save to a pkl file instead.
+
+        Args:
+            cfg: an omegaconf config object
+            filename: yaml file name to save the config file
+        """
+        logger = logging.getLogger(__name__)
+        try:
+            cfg = deepcopy(cfg)
+        except Exception:
+            pass
+        else:
+            # if it's deep-copyable, then...
+            def _replace_type_by_name(x):
+                if "_target_" in x and callable(x._target_):
+                    try:
+                        x._target_ = _convert_target_to_string(x._target_)
+                    except AttributeError:
+                        pass
+
+            # not necessary, but makes yaml looks nicer
+            _visit_dict_config(cfg, _replace_type_by_name)
+
+        save_pkl = False
+        try:
+            dict = OmegaConf.to_container(
+                cfg,
+                # Do not resolve interpolation when saving, i.e. do not turn ${a} into
+                # actual values when saving.
+                resolve=False,
+                # Save structures (dataclasses) in a format that can be instantiated later.
+                # Without this option, the type information of the dataclass will be erased.
+                structured_config_mode=SCMode.INSTANTIATE,
+            )
+            dumped = yaml.dump(dict, default_flow_style=None, allow_unicode=True, width=9999)
+            with PathManager.open(filename, "w") as f:
+                f.write(dumped)
+
+            try:
+                _ = yaml.unsafe_load(dumped)  # test that it is loadable
+            except Exception:
+                logger.warning(
+                    "The config contains objects that cannot serialize to a valid yaml. "
+                    f"{filename} is human-readable but cannot be loaded."
+                )
+                save_pkl = True
+        except Exception:
+            logger.exception("Unable to serialize the config to yaml. Error:")
+            save_pkl = True
+
+        if save_pkl:
+            new_filename = filename + ".pkl"
+            try:
+                # retry by pickle
+                with PathManager.open(new_filename, "wb") as f:
+                    cloudpickle.dump(cfg, f)
+                logger.warning(f"Config is saved using cloudpickle at {new_filename}.")
+            except Exception:
+                pass
+
+    @staticmethod
+    def apply_overrides(cfg, overrides: List[str]):
+        """
+        In-place override contents of cfg.
+
+        Args:
+            cfg: an omegaconf config object
+            overrides: list of strings in the format of "a=b" to override configs.
+                See https://hydra.cc/docs/next/advanced/override_grammar/basic/
+                for syntax.
+
+        Returns:
+            the cfg object
+        """
+
+        def safe_update(cfg, key, value):
+            parts = key.split(".")
+            for idx in range(1, len(parts)):
+                prefix = ".".join(parts[:idx])
+                v = OmegaConf.select(cfg, prefix, default=None)
+                if v is None:
+                    break
+                if not OmegaConf.is_config(v):
+                    raise KeyError(
+                        f"Trying to update key {key}, but {prefix} "
+                        f"is not a config, but has type {type(v)}."
+                    )
+            OmegaConf.update(cfg, key, value, merge=True)
+
+        try:
+            from hydra.core.override_parser.overrides_parser import OverridesParser
+
+            has_hydra = True
+        except ImportError:
+            has_hydra = False
+
+        if has_hydra:
+            parser = OverridesParser.create()
+            overrides = parser.parse_overrides(overrides)
+            for o in overrides:
+                key = o.key_or_group
+                value = o.value()
+                if o.is_delete():
+                    # TODO support this
+                    raise NotImplementedError("deletion is not yet a supported override")
+                safe_update(cfg, key, value)
+        else:
+            # Fallback. Does not support all the features and error checking like hydra.
+            for o in overrides:
+                key, value = o.split("=")
+                try:
+                    value = ast.literal_eval(value)
+                except NameError:
+                    pass
+                safe_update(cfg, key, value)
+        return cfg
+
+    @staticmethod
+    def to_py(cfg, prefix: str = "cfg."):
+        """
+        Try to convert a config object into Python-like psuedo code.
+
+        Note that perfect conversion is not always possible. So the returned
+        results are mainly meant to be human-readable, and not meant to be executed.
+
+        Args:
+            cfg: an omegaconf config object
+            prefix: root name for the resulting code (default: "cfg.")
+
+
+        Returns:
+            str of formatted Python code
+        """
+        import black
+
+        cfg = OmegaConf.to_container(cfg, resolve=True)
+
+        def _to_str(obj, prefix=None, inside_call=False):
+            if prefix is None:
+                prefix = []
+            if isinstance(obj, abc.Mapping) and "_target_" in obj:
+                # Dict representing a function call
+                target = _convert_target_to_string(obj.pop("_target_"))
+                args = []
+                for k, v in sorted(obj.items()):
+                    args.append(f"{k}={_to_str(v, inside_call=True)}")
+                args = ", ".join(args)
+                call = f"{target}({args})"
+                return "".join(prefix) + call
+            elif isinstance(obj, abc.Mapping) and not inside_call:
+                # Dict that is not inside a call is a list of top-level config objects that we
+                # render as one object per line with dot separated prefixes
+                key_list = []
+                for k, v in sorted(obj.items()):
+                    if isinstance(v, abc.Mapping) and "_target_" not in v:
+                        key_list.append(_to_str(v, prefix=prefix + [k + "."]))
+                    else:
+                        key = "".join(prefix) + k
+                        key_list.append(f"{key}={_to_str(v)}")
+                return "\n".join(key_list)
+            elif isinstance(obj, abc.Mapping):
+                # Dict that is inside a call is rendered as a regular dict
+                return (
+                    "{"
+                    + ",".join(
+                        f"{repr(k)}: {_to_str(v, inside_call=inside_call)}"
+                        for k, v in sorted(obj.items())
+                    )
+                    + "}"
+                )
+            elif isinstance(obj, list):
+                return "[" + ",".join(_to_str(x, inside_call=inside_call) for x in obj) + "]"
+            else:
+                return repr(obj)
+
+        py_str = _to_str(cfg, prefix=[prefix])
+        try:
+            return black.format_str(py_str, mode=black.Mode())
+        except black.InvalidInput:
+            return py_str

detectron2/data/__init__.py

@@ -0,0 +1,19 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+from . import transforms  # isort:skip
+
+from .build import (
+    build_batch_data_loader,
+    build_detection_test_loader,
+    build_detection_train_loader,
+    get_detection_dataset_dicts,
+    load_proposals_into_dataset,
+    print_instances_class_histogram,
+)
+from .catalog import DatasetCatalog, MetadataCatalog, Metadata
+from .common import DatasetFromList, MapDataset, ToIterableDataset
+from .dataset_mapper import DatasetMapper
+
+# ensure the builtin datasets are registered
+from . import datasets, samplers  # isort:skip
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

detectron2/data/benchmark.py

@@ -0,0 +1,225 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import logging
+import numpy as np
+from itertools import count
+from typing import List, Tuple
+import torch
+import tqdm
+from fvcore.common.timer import Timer
+
+from detectron2.utils import comm
+
+from .build import build_batch_data_loader
+from .common import DatasetFromList, MapDataset
+from .samplers import TrainingSampler
+
+logger = logging.getLogger(__name__)
+
+
+class _EmptyMapDataset(torch.utils.data.Dataset):
+    """
+    Map anything to emptiness.
+    """
+
+    def __init__(self, dataset):
+        self.ds = dataset
+
+    def __len__(self):
+        return len(self.ds)
+
+    def __getitem__(self, idx):
+        _ = self.ds[idx]
+        return [0]
+
+
+def iter_benchmark(
+    iterator, num_iter: int, warmup: int = 5, max_time_seconds: float = 60
+) -> Tuple[float, List[float]]:
+    """
+    Benchmark an iterator/iterable for `num_iter` iterations with an extra
+    `warmup` iterations of warmup.
+    End early if `max_time_seconds` time is spent on iterations.
+
+    Returns:
+        float: average time (seconds) per iteration
+        list[float]: time spent on each iteration. Sometimes useful for further analysis.
+    """
+    num_iter, warmup = int(num_iter), int(warmup)
+
+    iterator = iter(iterator)
+    for _ in range(warmup):
+        next(iterator)
+    timer = Timer()
+    all_times = []
+    for curr_iter in tqdm.trange(num_iter):
+        start = timer.seconds()
+        if start > max_time_seconds:
+            num_iter = curr_iter
+            break
+        next(iterator)
+        all_times.append(timer.seconds() - start)
+    avg = timer.seconds() / num_iter
+    return avg, all_times
+
+
+class DataLoaderBenchmark:
+    """
+    Some common benchmarks that help understand perf bottleneck of a standard dataloader
+    made of dataset, mapper and sampler.
+    """
+
+    def __init__(
+        self,
+        dataset,
+        *,
+        mapper,
+        sampler=None,
+        total_batch_size,
+        num_workers=0,
+        max_time_seconds: int = 90,
+    ):
+        """
+        Args:
+            max_time_seconds (int): maximum time to spent for each benchmark
+            other args: same as in `build.py:build_detection_train_loader`
+        """
+        if isinstance(dataset, list):
+            dataset = DatasetFromList(dataset, copy=False, serialize=True)
+        if sampler is None:
+            sampler = TrainingSampler(len(dataset))
+
+        self.dataset = dataset
+        self.mapper = mapper
+        self.sampler = sampler
+        self.total_batch_size = total_batch_size
+        self.num_workers = num_workers
+        self.per_gpu_batch_size = self.total_batch_size // comm.get_world_size()
+
+        self.max_time_seconds = max_time_seconds
+
+    def _benchmark(self, iterator, num_iter, warmup, msg=None):
+        avg, all_times = iter_benchmark(iterator, num_iter, warmup, self.max_time_seconds)
+        if msg is not None:
+            self._log_time(msg, avg, all_times)
+        return avg, all_times
+
+    def _log_time(self, msg, avg, all_times, distributed=False):
+        percentiles = [np.percentile(all_times, k, interpolation="nearest") for k in [1, 5, 95, 99]]
+        if not distributed:
+            logger.info(
+                f"{msg}: avg={1.0/avg:.1f} it/s, "
+                f"p1={percentiles[0]:.2g}s, p5={percentiles[1]:.2g}s, "
+                f"p95={percentiles[2]:.2g}s, p99={percentiles[3]:.2g}s."
+            )
+            return
+        avg_per_gpu = comm.all_gather(avg)
+        percentiles_per_gpu = comm.all_gather(percentiles)
+        if comm.get_rank() > 0:
+            return
+        for idx, avg, percentiles in zip(count(), avg_per_gpu, percentiles_per_gpu):
+            logger.info(
+                f"GPU{idx} {msg}: avg={1.0/avg:.1f} it/s, "
+                f"p1={percentiles[0]:.2g}s, p5={percentiles[1]:.2g}s, "
+                f"p95={percentiles[2]:.2g}s, p99={percentiles[3]:.2g}s."
+            )
+
+    def benchmark_dataset(self, num_iter, warmup=5):
+        """
+        Benchmark the speed of taking raw samples from the dataset.
+        """
+
+        def loader():
+            while True:
+                for k in self.sampler:
+                    yield self.dataset[k]
+
+        self._benchmark(loader(), num_iter, warmup, "Dataset Alone")
+
+    def benchmark_mapper(self, num_iter, warmup=5):
+        """
+        Benchmark the speed of taking raw samples from the dataset and map
+        them in a single process.
+        """
+
+        def loader():
+            while True:
+                for k in self.sampler:
+                    yield self.mapper(self.dataset[k])
+
+        self._benchmark(loader(), num_iter, warmup, "Single Process Mapper (sec/sample)")
+
+    def benchmark_workers(self, num_iter, warmup=10):
+        """
+        Benchmark the dataloader by tuning num_workers to [0, 1, self.num_workers].
+        """
+        candidates = [0, 1]
+        if self.num_workers not in candidates:
+            candidates.append(self.num_workers)
+
+        dataset = MapDataset(self.dataset, self.mapper)
+        for n in candidates:
+            loader = build_batch_data_loader(
+                dataset,
+                self.sampler,
+                self.total_batch_size,
+                num_workers=n,
+            )
+            self._benchmark(
+                iter(loader),
+                num_iter * max(n, 1),
+                warmup * max(n, 1),
+                f"DataLoader ({n} workers, bs={self.per_gpu_batch_size})",
+            )
+            del loader
+
+    def benchmark_IPC(self, num_iter, warmup=10):
+        """
+        Benchmark the dataloader where each worker outputs nothing. This
+        eliminates the IPC overhead compared to the regular dataloader.
+
+        PyTorch multiprocessing's IPC only optimizes for torch tensors.
+        Large numpy arrays or other data structure may incur large IPC overhead.
+        """
+        n = self.num_workers
+        dataset = _EmptyMapDataset(MapDataset(self.dataset, self.mapper))
+        loader = build_batch_data_loader(
+            dataset, self.sampler, self.total_batch_size, num_workers=n
+        )
+        self._benchmark(
+            iter(loader),
+            num_iter * max(n, 1),
+            warmup * max(n, 1),
+            f"DataLoader ({n} workers, bs={self.per_gpu_batch_size}) w/o comm",
+        )
+
+    def benchmark_distributed(self, num_iter, warmup=10):
+        """
+        Benchmark the dataloader in each distributed worker, and log results of
+        all workers. This helps understand the final performance as well as
+        the variances among workers.
+
+        It also prints startup time (first iter) of the dataloader.
+        """
+        gpu = comm.get_world_size()
+        dataset = MapDataset(self.dataset, self.mapper)
+        n = self.num_workers
+        loader = build_batch_data_loader(
+            dataset, self.sampler, self.total_batch_size, num_workers=n
+        )
+
+        timer = Timer()
+        loader = iter(loader)
+        next(loader)
+        startup_time = timer.seconds()
+        logger.info("Dataloader startup time: {:.2f} seconds".format(startup_time))
+
+        comm.synchronize()
+
+        avg, all_times = self._benchmark(loader, num_iter * max(n, 1), warmup * max(n, 1))
+        del loader
+        self._log_time(
+            f"DataLoader ({gpu} GPUs x {n} workers, total bs={self.total_batch_size})",
+            avg,
+            all_times,
+            True,
+        )

detectron2/data/build.py

@@ -0,0 +1,694 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import itertools
+import logging
+import numpy as np
+import operator
+import pickle
+from collections import OrderedDict, defaultdict
+from typing import Any, Callable, Dict, List, Optional, Union
+import torch
+import torch.utils.data as torchdata
+from tabulate import tabulate
+from termcolor import colored
+
+from detectron2.config import configurable
+from detectron2.structures import BoxMode
+from detectron2.utils.comm import get_world_size
+from detectron2.utils.env import seed_all_rng
+from detectron2.utils.file_io import PathManager
+from detectron2.utils.logger import _log_api_usage, log_first_n
+
+from .catalog import DatasetCatalog, MetadataCatalog
+from .common import AspectRatioGroupedDataset, DatasetFromList, MapDataset, ToIterableDataset
+from .dataset_mapper import DatasetMapper
+from .detection_utils import check_metadata_consistency
+from .samplers import (
+    InferenceSampler,
+    RandomSubsetTrainingSampler,
+    RepeatFactorTrainingSampler,
+    TrainingSampler,
+)
+
+"""
+This file contains the default logic to build a dataloader for training or testing.
+"""
+
+__all__ = [
+    "build_batch_data_loader",
+    "build_detection_train_loader",
+    "build_detection_test_loader",
+    "get_detection_dataset_dicts",
+    "load_proposals_into_dataset",
+    "print_instances_class_histogram",
+]
+
+
+def filter_images_with_only_crowd_annotations(dataset_dicts):
+    """
+    Filter out images with none annotations or only crowd annotations
+    (i.e., images without non-crowd annotations).
+    A common training-time preprocessing on COCO dataset.
+
+    Args:
+        dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
+
+    Returns:
+        list[dict]: the same format, but filtered.
+    """
+    num_before = len(dataset_dicts)
+
+    def valid(anns):
+        for ann in anns:
+            if ann.get("iscrowd", 0) == 0:
+                return True
+        return False
+
+    dataset_dicts = [x for x in dataset_dicts if valid(x["annotations"])]
+    num_after = len(dataset_dicts)
+    logger = logging.getLogger(__name__)
+    logger.info(
+        "Removed {} images with no usable annotations. {} images left.".format(
+            num_before - num_after, num_after
+        )
+    )
+    return dataset_dicts
+
+
+def filter_images_with_few_keypoints(dataset_dicts, min_keypoints_per_image):
+    """
+    Filter out images with too few number of keypoints.
+
+    Args:
+        dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
+
+    Returns:
+        list[dict]: the same format as dataset_dicts, but filtered.
+    """
+    num_before = len(dataset_dicts)
+
+    def visible_keypoints_in_image(dic):
+        # Each keypoints field has the format [x1, y1, v1, ...], where v is visibility
+        annotations = dic["annotations"]
+        return sum(
+            (np.array(ann["keypoints"][2::3]) > 0).sum()
+            for ann in annotations
+            if "keypoints" in ann
+        )
+
+    dataset_dicts = [
+        x for x in dataset_dicts if visible_keypoints_in_image(x) >= min_keypoints_per_image
+    ]
+    num_after = len(dataset_dicts)
+    logger = logging.getLogger(__name__)
+    logger.info(
+        "Removed {} images with fewer than {} keypoints.".format(
+            num_before - num_after, min_keypoints_per_image
+        )
+    )
+    return dataset_dicts
+
+
+def load_proposals_into_dataset(dataset_dicts, proposal_file):
+    """
+    Load precomputed object proposals into the dataset.
+
+    The proposal file should be a pickled dict with the following keys:
+
+    - "ids": list[int] or list[str], the image ids
+    - "boxes": list[np.ndarray], each is an Nx4 array of boxes corresponding to the image id
+    - "objectness_logits": list[np.ndarray], each is an N sized array of objectness scores
+      corresponding to the boxes.
+    - "bbox_mode": the BoxMode of the boxes array. Defaults to ``BoxMode.XYXY_ABS``.
+
+    Args:
+        dataset_dicts (list[dict]): annotations in Detectron2 Dataset format.
+        proposal_file (str): file path of pre-computed proposals, in pkl format.
+
+    Returns:
+        list[dict]: the same format as dataset_dicts, but added proposal field.
+    """
+    logger = logging.getLogger(__name__)
+    logger.info("Loading proposals from: {}".format(proposal_file))
+
+    with PathManager.open(proposal_file, "rb") as f:
+        proposals = pickle.load(f, encoding="latin1")
+
+    # Rename the key names in D1 proposal files
+    rename_keys = {"indexes": "ids", "scores": "objectness_logits"}
+    for key in rename_keys:
+        if key in proposals:
+            proposals[rename_keys[key]] = proposals.pop(key)
+
+    # Fetch the indexes of all proposals that are in the dataset
+    # Convert image_id to str since they could be int.
+    img_ids = set({str(record["image_id"]) for record in dataset_dicts})
+    id_to_index = {str(id): i for i, id in enumerate(proposals["ids"]) if str(id) in img_ids}
+
+    # Assuming default bbox_mode of precomputed proposals are 'XYXY_ABS'
+    bbox_mode = BoxMode(proposals["bbox_mode"]) if "bbox_mode" in proposals else BoxMode.XYXY_ABS
+
+    for record in dataset_dicts:
+        # Get the index of the proposal
+        i = id_to_index[str(record["image_id"])]
+
+        boxes = proposals["boxes"][i]
+        objectness_logits = proposals["objectness_logits"][i]
+        # Sort the proposals in descending order of the scores
+        inds = objectness_logits.argsort()[::-1]
+        record["proposal_boxes"] = boxes[inds]
+        record["proposal_objectness_logits"] = objectness_logits[inds]
+        record["proposal_bbox_mode"] = bbox_mode
+
+    return dataset_dicts
+
+
+def print_instances_class_histogram(dataset_dicts, class_names):
+    """
+    Args:
+        dataset_dicts (list[dict]): list of dataset dicts.
+        class_names (list[str]): list of class names (zero-indexed).
+    """
+    num_classes = len(class_names)
+    hist_bins = np.arange(num_classes + 1)
+    histogram = np.zeros((num_classes,), dtype=int)
+    for entry in dataset_dicts:
+        annos = entry["annotations"]
+        classes = np.asarray(
+            [x["category_id"] for x in annos if not x.get("iscrowd", 0)], dtype=int
+        )
+        if len(classes):
+            assert classes.min() >= 0, f"Got an invalid category_id={classes.min()}"
+            assert (
+                classes.max() < num_classes
+            ), f"Got an invalid category_id={classes.max()} for a dataset of {num_classes} classes"
+        histogram += np.histogram(classes, bins=hist_bins)[0]
+
+    N_COLS = min(6, len(class_names) * 2)
+
+    def short_name(x):
+        # make long class names shorter. useful for lvis
+        if len(x) > 13:
+            return x[:11] + ".."
+        return x
+
+    data = list(
+        itertools.chain(*[[short_name(class_names[i]), int(v)] for i, v in enumerate(histogram)])
+    )
+    total_num_instances = sum(data[1::2])
+    data.extend([None] * (N_COLS - (len(data) % N_COLS)))
+    if num_classes > 1:
+        data.extend(["total", total_num_instances])
+    data = itertools.zip_longest(*[data[i::N_COLS] for i in range(N_COLS)])
+    table = tabulate(
+        data,
+        headers=["category", "#instances"] * (N_COLS // 2),
+        tablefmt="pipe",
+        numalign="left",
+        stralign="center",
+    )
+    log_first_n(
+        logging.INFO,
+        "Distribution of instances among all {} categories:\n".format(num_classes)
+        + colored(table, "cyan"),
+        key="message",
+    )
+
+
+def get_detection_dataset_dicts(
+    names,
+    filter_empty=True,
+    min_keypoints=0,
+    proposal_files=None,
+    check_consistency=True,
+):
+    """
+    Load and prepare dataset dicts for instance detection/segmentation and semantic segmentation.
+
+    Args:
+        names (str or list[str]): a dataset name or a list of dataset names
+        filter_empty (bool): whether to filter out images without instance annotations
+        min_keypoints (int): filter out images with fewer keypoints than
+            `min_keypoints`. Set to 0 to do nothing.
+        proposal_files (list[str]): if given, a list of object proposal files
+            that match each dataset in `names`.
+        check_consistency (bool): whether to check if datasets have consistent metadata.
+
+    Returns:
+        list[dict]: a list of dicts following the standard dataset dict format.
+    """
+    if isinstance(names, str):
+        names = [names]
+    assert len(names), names
+
+    available_datasets = DatasetCatalog.keys()
+    names_set = set(names)
+    if not names_set.issubset(available_datasets):
+        logger = logging.getLogger(__name__)
+        logger.warning(
+            "The following dataset names are not registered in the DatasetCatalog: "
+            f"{names_set - available_datasets}. "
+            f"Available datasets are {available_datasets}"
+        )
+
+    dataset_dicts = [DatasetCatalog.get(dataset_name) for dataset_name in names]
+
+    if isinstance(dataset_dicts[0], torchdata.Dataset):
+        if len(dataset_dicts) > 1:
+            # ConcatDataset does not work for iterable style dataset.
+            # We could support concat for iterable as well, but it's often
+            # not a good idea to concat iterables anyway.
+            return torchdata.ConcatDataset(dataset_dicts)
+        return dataset_dicts[0]
+
+    for dataset_name, dicts in zip(names, dataset_dicts):
+        assert len(dicts), "Dataset '{}' is empty!".format(dataset_name)
+
+    if proposal_files is not None:
+        assert len(names) == len(proposal_files)
+        # load precomputed proposals from proposal files
+        dataset_dicts = [
+            load_proposals_into_dataset(dataset_i_dicts, proposal_file)
+            for dataset_i_dicts, proposal_file in zip(dataset_dicts, proposal_files)
+        ]
+
+    dataset_dicts = list(itertools.chain.from_iterable(dataset_dicts))
+
+    has_instances = "annotations" in dataset_dicts[0]
+    if filter_empty and has_instances:
+        dataset_dicts = filter_images_with_only_crowd_annotations(dataset_dicts)
+    if min_keypoints > 0 and has_instances:
+        dataset_dicts = filter_images_with_few_keypoints(dataset_dicts, min_keypoints)
+
+    if check_consistency and has_instances:
+        try:
+            class_names = MetadataCatalog.get(names[0]).thing_classes
+            check_metadata_consistency("thing_classes", names)
+            print_instances_class_histogram(dataset_dicts, class_names)
+        except AttributeError:  # class names are not available for this dataset
+            pass
+
+    assert len(dataset_dicts), "No valid data found in {}.".format(",".join(names))
+    return dataset_dicts
+
+
+def build_batch_data_loader(
+    dataset,
+    sampler,
+    total_batch_size,
+    *,
+    aspect_ratio_grouping=False,
+    num_workers=0,
+    collate_fn=None,
+    drop_last: bool = True,
+    single_gpu_batch_size=None,
+    prefetch_factor=2,
+    persistent_workers=False,
+    pin_memory=False,
+    seed=None,
+    **kwargs,
+):
+    """
+    Build a batched dataloader. The main differences from `torch.utils.data.DataLoader` are:
+    1. support aspect ratio grouping options
+    2. use no "batch collation", because this is common for detection training
+
+    Args:
+        dataset (torch.utils.data.Dataset): a pytorch map-style or iterable dataset.
+        sampler (torch.utils.data.sampler.Sampler or None): a sampler that produces indices.
+            Must be provided iff. ``dataset`` is a map-style dataset.
+        total_batch_size, aspect_ratio_grouping, num_workers, collate_fn: see
+            :func:`build_detection_train_loader`.
+        single_gpu_batch_size: You can specify either `single_gpu_batch_size` or `total_batch_size`.
+            `single_gpu_batch_size` specifies the batch size that will be used for each gpu/process.
+            `total_batch_size` allows you to specify the total aggregate batch size across gpus.
+            It is an error to supply a value for both.
+        drop_last (bool): if ``True``, the dataloader will drop incomplete batches.
+
+    Returns:
+        iterable[list]. Length of each list is the batch size of the current
+            GPU. Each element in the list comes from the dataset.
+    """
+    if single_gpu_batch_size:
+        if total_batch_size:
+            raise ValueError(
+                """total_batch_size and single_gpu_batch_size are mutually incompatible.
+                Please specify only one. """
+            )
+        batch_size = single_gpu_batch_size
+    else:
+        world_size = get_world_size()
+        assert (
+            total_batch_size > 0 and total_batch_size % world_size == 0
+        ), "Total batch size ({}) must be divisible by the number of gpus ({}).".format(
+            total_batch_size, world_size
+        )
+        batch_size = total_batch_size // world_size
+    logger = logging.getLogger(__name__)
+    logger.info("Making batched data loader with batch_size=%d", batch_size)
+
+    if isinstance(dataset, torchdata.IterableDataset):
+        assert sampler is None, "sampler must be None if dataset is IterableDataset"
+    else:
+        dataset = ToIterableDataset(dataset, sampler, shard_chunk_size=batch_size)
+
+    generator = None
+    if seed is not None:
+        generator = torch.Generator()
+        generator.manual_seed(seed)
+
+    if aspect_ratio_grouping:
+        assert drop_last, "Aspect ratio grouping will drop incomplete batches."
+        data_loader = torchdata.DataLoader(
+            dataset,
+            num_workers=num_workers,
+            collate_fn=operator.itemgetter(0),  # don't batch, but yield individual elements
+            worker_init_fn=worker_init_reset_seed,
+            prefetch_factor=prefetch_factor if num_workers > 0 else None,
+            persistent_workers=persistent_workers,
+            pin_memory=pin_memory,
+            generator=generator,
+            **kwargs,
+        )  # yield individual mapped dict
+        data_loader = AspectRatioGroupedDataset(data_loader, batch_size)
+        if collate_fn is None:
+            return data_loader
+        return MapDataset(data_loader, collate_fn)
+    else:
+        return torchdata.DataLoader(
+            dataset,
+            batch_size=batch_size,
+            drop_last=drop_last,
+            num_workers=num_workers,
+            collate_fn=trivial_batch_collator if collate_fn is None else collate_fn,
+            worker_init_fn=worker_init_reset_seed,
+            prefetch_factor=prefetch_factor if num_workers > 0 else None,
+            persistent_workers=persistent_workers,
+            pin_memory=pin_memory,
+            generator=generator,
+            **kwargs,
+        )
+
+
+def _get_train_datasets_repeat_factors(cfg) -> Dict[str, float]:
+    repeat_factors = cfg.DATASETS.TRAIN_REPEAT_FACTOR
+    assert all(len(tup) == 2 for tup in repeat_factors)
+    name_to_weight = defaultdict(lambda: 1, dict(repeat_factors))
+    # The sampling weights map should only contain datasets in train config
+    unrecognized = set(name_to_weight.keys()) - set(cfg.DATASETS.TRAIN)
+    assert not unrecognized, f"unrecognized datasets: {unrecognized}"
+    logger = logging.getLogger(__name__)
+    logger.info(f"Found repeat factors: {list(name_to_weight.items())}")
+
+    # pyre-fixme[7]: Expected `Dict[str, float]` but got `DefaultDict[typing.Any, int]`.
+    return name_to_weight
+
+
+def _build_weighted_sampler(cfg, enable_category_balance=False):
+    dataset_repeat_factors = _get_train_datasets_repeat_factors(cfg)
+    # OrderedDict to guarantee order of values() consistent with repeat factors
+    dataset_name_to_dicts = OrderedDict(
+        {
+            name: get_detection_dataset_dicts(
+                [name],
+                filter_empty=cfg.DATALOADER.FILTER_EMPTY_ANNOTATIONS,
+                min_keypoints=(
+                    cfg.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE
+                    if cfg.MODEL.KEYPOINT_ON
+                    else 0
+                ),
+                proposal_files=(
+                    cfg.DATASETS.PROPOSAL_FILES_TRAIN if cfg.MODEL.LOAD_PROPOSALS else None
+                ),
+            )
+            for name in cfg.DATASETS.TRAIN
+        }
+    )
+    # Repeat factor for every sample in the dataset
+    repeat_factors = [
+        [dataset_repeat_factors[dsname]] * len(dataset_name_to_dicts[dsname])
+        for dsname in cfg.DATASETS.TRAIN
+    ]
+
+    repeat_factors = list(itertools.chain.from_iterable(repeat_factors))
+
+    repeat_factors = torch.tensor(repeat_factors)
+    logger = logging.getLogger(__name__)
+    if enable_category_balance:
+        """
+        1. Calculate repeat factors using category frequency for each dataset and then merge them.
+        2. Element wise dot producting the dataset frequency repeat factors with
+            the category frequency repeat factors gives the final repeat factors.
+        """
+        category_repeat_factors = [
+            RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
+                dataset_dict, cfg.DATALOADER.REPEAT_THRESHOLD, sqrt=cfg.DATALOADER.REPEAT_SQRT
+            )
+            for dataset_dict in dataset_name_to_dicts.values()
+        ]
+        # flatten the category repeat factors from all datasets
+        category_repeat_factors = list(itertools.chain.from_iterable(category_repeat_factors))
+        category_repeat_factors = torch.tensor(category_repeat_factors)
+        repeat_factors = torch.mul(category_repeat_factors, repeat_factors)
+        repeat_factors = repeat_factors / torch.min(repeat_factors)
+        logger.info(
+            "Using WeightedCategoryTrainingSampler with repeat_factors={}".format(
+                cfg.DATASETS.TRAIN_REPEAT_FACTOR
+            )
+        )
+    else:
+        logger.info(
+            "Using WeightedTrainingSampler with repeat_factors={}".format(
+                cfg.DATASETS.TRAIN_REPEAT_FACTOR
+            )
+        )
+
+    sampler = RepeatFactorTrainingSampler(repeat_factors)
+    return sampler
+
+
+def _train_loader_from_config(cfg, mapper=None, *, dataset=None, sampler=None):
+    if dataset is None:
+        dataset = get_detection_dataset_dicts(
+            cfg.DATASETS.TRAIN,
+            filter_empty=cfg.DATALOADER.FILTER_EMPTY_ANNOTATIONS,
+            min_keypoints=(
+                cfg.MODEL.ROI_KEYPOINT_HEAD.MIN_KEYPOINTS_PER_IMAGE if cfg.MODEL.KEYPOINT_ON else 0
+            ),
+            proposal_files=cfg.DATASETS.PROPOSAL_FILES_TRAIN if cfg.MODEL.LOAD_PROPOSALS else None,
+        )
+        _log_api_usage("dataset." + cfg.DATASETS.TRAIN[0])
+
+    if mapper is None:
+        mapper = DatasetMapper(cfg, True)
+
+    if sampler is None:
+        sampler_name = cfg.DATALOADER.SAMPLER_TRAIN
+        logger = logging.getLogger(__name__)
+        if isinstance(dataset, torchdata.IterableDataset):
+            logger.info("Not using any sampler since the dataset is IterableDataset.")
+            sampler = None
+        else:
+            logger.info("Using training sampler {}".format(sampler_name))
+            if sampler_name == "TrainingSampler":
+                sampler = TrainingSampler(len(dataset), seed=cfg.SEED)
+            elif sampler_name == "RepeatFactorTrainingSampler":
+                repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
+                    dataset, cfg.DATALOADER.REPEAT_THRESHOLD, sqrt=cfg.DATALOADER.REPEAT_SQRT
+                )
+                sampler = RepeatFactorTrainingSampler(repeat_factors, seed=cfg.SEED)
+            elif sampler_name == "RandomSubsetTrainingSampler":
+                sampler = RandomSubsetTrainingSampler(
+                    len(dataset), cfg.DATALOADER.RANDOM_SUBSET_RATIO
+                )
+            elif sampler_name == "WeightedTrainingSampler":
+                sampler = _build_weighted_sampler(cfg)
+            elif sampler_name == "WeightedCategoryTrainingSampler":
+                sampler = _build_weighted_sampler(cfg, enable_category_balance=True)
+            else:
+                raise ValueError("Unknown training sampler: {}".format(sampler_name))
+
+    return {
+        "dataset": dataset,
+        "sampler": sampler,
+        "mapper": mapper,
+        "total_batch_size": cfg.SOLVER.IMS_PER_BATCH,
+        "aspect_ratio_grouping": cfg.DATALOADER.ASPECT_RATIO_GROUPING,
+        "num_workers": cfg.DATALOADER.NUM_WORKERS,
+    }
+
+
+@configurable(from_config=_train_loader_from_config)
+def build_detection_train_loader(
+    dataset,
+    *,
+    mapper,
+    sampler=None,
+    total_batch_size,
+    aspect_ratio_grouping=True,
+    num_workers=0,
+    collate_fn=None,
+    **kwargs,
+):
+    """
+    Build a dataloader for object detection with some default features.
+
+    Args:
+        dataset (list or torch.utils.data.Dataset): a list of dataset dicts,
+            or a pytorch dataset (either map-style or iterable). It can be obtained
+            by using :func:`DatasetCatalog.get` or :func:`get_detection_dataset_dicts`.
+        mapper (callable): a callable which takes a sample (dict) from dataset and
+            returns the format to be consumed by the model.
+            When using cfg, the default choice is ``DatasetMapper(cfg, is_train=True)``.
+        sampler (torch.utils.data.sampler.Sampler or None): a sampler that produces
+            indices to be applied on ``dataset``.
+            If ``dataset`` is map-style, the default sampler is a :class:`TrainingSampler`,
+            which coordinates an infinite random shuffle sequence across all workers.
+            Sampler must be None if ``dataset`` is iterable.
+        total_batch_size (int): total batch size across all workers.
+        aspect_ratio_grouping (bool): whether to group images with similar
+            aspect ratio for efficiency. When enabled, it requires each
+            element in dataset be a dict with keys "width" and "height".
+        num_workers (int): number of parallel data loading workers
+        collate_fn: a function that determines how to do batching, same as the argument of
+            `torch.utils.data.DataLoader`. Defaults to do no collation and return a list of
+            data. No collation is OK for small batch size and simple data structures.
+            If your batch size is large and each sample contains too many small tensors,
+            it's more efficient to collate them in data loader.
+
+    Returns:
+        torch.utils.data.DataLoader:
+            a dataloader. Each output from it is a ``list[mapped_element]`` of length
+            ``total_batch_size / num_workers``, where ``mapped_element`` is produced
+            by the ``mapper``.
+    """
+    if isinstance(dataset, list):
+        dataset = DatasetFromList(dataset, copy=False)
+    if mapper is not None:
+        dataset = MapDataset(dataset, mapper)
+
+    if isinstance(dataset, torchdata.IterableDataset):
+        assert sampler is None, "sampler must be None if dataset is IterableDataset"
+    else:
+        if sampler is None:
+            sampler = TrainingSampler(len(dataset))
+        assert isinstance(sampler, torchdata.Sampler), f"Expect a Sampler but got {type(sampler)}"
+    return build_batch_data_loader(
+        dataset,
+        sampler,
+        total_batch_size,
+        aspect_ratio_grouping=aspect_ratio_grouping,
+        num_workers=num_workers,
+        collate_fn=collate_fn,
+        **kwargs,
+    )
+
+
+def _test_loader_from_config(cfg, dataset_name, mapper=None):
+    """
+    Uses the given `dataset_name` argument (instead of the names in cfg), because the
+    standard practice is to evaluate each test set individually (not combining them).
+    """
+    if isinstance(dataset_name, str):
+        dataset_name = [dataset_name]
+
+    dataset = get_detection_dataset_dicts(
+        dataset_name,
+        filter_empty=False,
+        proposal_files=(
+            [
+                cfg.DATASETS.PROPOSAL_FILES_TEST[list(cfg.DATASETS.TEST).index(x)]
+                for x in dataset_name
+            ]
+            if cfg.MODEL.LOAD_PROPOSALS
+            else None
+        ),
+    )
+    if mapper is None:
+        mapper = DatasetMapper(cfg, False)
+    return {
+        "dataset": dataset,
+        "mapper": mapper,
+        "num_workers": cfg.DATALOADER.NUM_WORKERS,
+        "sampler": (
+            InferenceSampler(len(dataset))
+            if not isinstance(dataset, torchdata.IterableDataset)
+            else None
+        ),
+    }
+
+
+@configurable(from_config=_test_loader_from_config)
+def build_detection_test_loader(
+    dataset: Union[List[Any], torchdata.Dataset],
+    *,
+    mapper: Callable[[Dict[str, Any]], Any],
+    sampler: Optional[torchdata.Sampler] = None,
+    batch_size: int = 1,
+    num_workers: int = 0,
+    collate_fn: Optional[Callable[[List[Any]], Any]] = None,
+) -> torchdata.DataLoader:
+    """
+    Similar to `build_detection_train_loader`, with default batch size = 1,
+    and sampler = :class:`InferenceSampler`. This sampler coordinates all workers
+    to produce the exact set of all samples.
+
+    Args:
+        dataset: a list of dataset dicts,
+            or a pytorch dataset (either map-style or iterable). They can be obtained
+            by using :func:`DatasetCatalog.get` or :func:`get_detection_dataset_dicts`.
+        mapper: a callable which takes a sample (dict) from dataset
+           and returns the format to be consumed by the model.
+           When using cfg, the default choice is ``DatasetMapper(cfg, is_train=False)``.
+        sampler: a sampler that produces
+            indices to be applied on ``dataset``. Default to :class:`InferenceSampler`,
+            which splits the dataset across all workers. Sampler must be None
+            if `dataset` is iterable.
+        batch_size: the batch size of the data loader to be created.
+            Default to 1 image per worker since this is the standard when reporting
+            inference time in papers.
+        num_workers: number of parallel data loading workers
+        collate_fn: same as the argument of `torch.utils.data.DataLoader`.
+            Defaults to do no collation and return a list of data.
+
+    Returns:
+        DataLoader: a torch DataLoader, that loads the given detection
+        dataset, with test-time transformation and batching.
+
+    Examples:
+    ::
+        data_loader = build_detection_test_loader(
+            DatasetRegistry.get("my_test"),
+            mapper=DatasetMapper(...))
+
+        # or, instantiate with a CfgNode:
+        data_loader = build_detection_test_loader(cfg, "my_test")
+    """
+    if isinstance(dataset, list):
+        dataset = DatasetFromList(dataset, copy=False)
+    if mapper is not None:
+        dataset = MapDataset(dataset, mapper)
+    if isinstance(dataset, torchdata.IterableDataset):
+        assert sampler is None, "sampler must be None if dataset is IterableDataset"
+    else:
+        if sampler is None:
+            sampler = InferenceSampler(len(dataset))
+    return torchdata.DataLoader(
+        dataset,
+        batch_size=batch_size,
+        sampler=sampler,
+        drop_last=False,
+        num_workers=num_workers,
+        collate_fn=trivial_batch_collator if collate_fn is None else collate_fn,
+    )
+
+
+def trivial_batch_collator(batch):
+    """
+    A batch collator that does nothing.
+    """
+    return batch
+
+
+def worker_init_reset_seed(worker_id):
+    initial_seed = torch.initial_seed() % 2**31
+    seed_all_rng(initial_seed + worker_id)

detectron2/data/catalog.py

@@ -0,0 +1,236 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import copy
+import logging
+import types
+from collections import UserDict
+from typing import List
+
+from detectron2.utils.logger import log_first_n
+
+__all__ = ["DatasetCatalog", "MetadataCatalog", "Metadata"]
+
+
+class _DatasetCatalog(UserDict):
+    """
+    A global dictionary that stores information about the datasets and how to obtain them.
+
+    It contains a mapping from strings
+    (which are names that identify a dataset, e.g. "coco_2014_train")
+    to a function which parses the dataset and returns the samples in the
+    format of `list[dict]`.
+
+    The returned dicts should be in Detectron2 Dataset format (See DATASETS.md for details)
+    if used with the data loader functionalities in `data/build.py,data/detection_transform.py`.
+
+    The purpose of having this catalog is to make it easy to choose
+    different datasets, by just using the strings in the config.
+    """
+
+    def register(self, name, func):
+        """
+        Args:
+            name (str): the name that identifies a dataset, e.g. "coco_2014_train".
+            func (callable): a callable which takes no arguments and returns a list of dicts.
+                It must return the same results if called multiple times.
+        """
+        assert callable(func), "You must register a function with `DatasetCatalog.register`!"
+        assert name not in self, "Dataset '{}' is already registered!".format(name)
+        self[name] = func
+
+    def get(self, name):
+        """
+        Call the registered function and return its results.
+
+        Args:
+            name (str): the name that identifies a dataset, e.g. "coco_2014_train".
+
+        Returns:
+            list[dict]: dataset annotations.
+        """
+        try:
+            f = self[name]
+        except KeyError as e:
+            raise KeyError(
+                "Dataset '{}' is not registered! Available datasets are: {}".format(
+                    name, ", ".join(list(self.keys()))
+                )
+            ) from e
+        return f()
+
+    def list(self) -> List[str]:
+        """
+        List all registered datasets.
+
+        Returns:
+            list[str]
+        """
+        return list(self.keys())
+
+    def remove(self, name):
+        """
+        Alias of ``pop``.
+        """
+        self.pop(name)
+
+    def __str__(self):
+        return "DatasetCatalog(registered datasets: {})".format(", ".join(self.keys()))
+
+    __repr__ = __str__
+
+
+DatasetCatalog = _DatasetCatalog()
+DatasetCatalog.__doc__ = (
+    _DatasetCatalog.__doc__
+    + """
+    .. automethod:: detectron2.data.catalog.DatasetCatalog.register
+    .. automethod:: detectron2.data.catalog.DatasetCatalog.get
+"""
+)
+
+
+class Metadata(types.SimpleNamespace):
+    """
+    A class that supports simple attribute setter/getter.
+    It is intended for storing metadata of a dataset and make it accessible globally.
+
+    Examples:
+    ::
+        # somewhere when you load the data:
+        MetadataCatalog.get("mydataset").thing_classes = ["person", "dog"]
+
+        # somewhere when you print statistics or visualize:
+        classes = MetadataCatalog.get("mydataset").thing_classes
+    """
+
+    # the name of the dataset
+    # set default to N/A so that `self.name` in the errors will not trigger getattr again
+    name: str = "N/A"
+
+    _RENAMED = {
+        "class_names": "thing_classes",
+        "dataset_id_to_contiguous_id": "thing_dataset_id_to_contiguous_id",
+        "stuff_class_names": "stuff_classes",
+    }
+
+    def __getattr__(self, key):
+        if key in self._RENAMED:
+            log_first_n(
+                logging.WARNING,
+                "Metadata '{}' was renamed to '{}'!".format(key, self._RENAMED[key]),
+                n=10,
+            )
+            return getattr(self, self._RENAMED[key])
+
+        # "name" exists in every metadata
+        if len(self.__dict__) > 1:
+            raise AttributeError(
+                "Attribute '{}' does not exist in the metadata of dataset '{}'. Available "
+                "keys are {}.".format(key, self.name, str(self.__dict__.keys()))
+            )
+        else:
+            raise AttributeError(
+                f"Attribute '{key}' does not exist in the metadata of dataset '{self.name}': "
+                "metadata is empty."
+            )
+
+    def __setattr__(self, key, val):
+        if key in self._RENAMED:
+            log_first_n(
+                logging.WARNING,
+                "Metadata '{}' was renamed to '{}'!".format(key, self._RENAMED[key]),
+                n=10,
+            )
+            setattr(self, self._RENAMED[key], val)
+
+        # Ensure that metadata of the same name stays consistent
+        try:
+            oldval = getattr(self, key)
+            assert oldval == val, (
+                "Attribute '{}' in the metadata of '{}' cannot be set "
+                "to a different value!\n{} != {}".format(key, self.name, oldval, val)
+            )
+        except AttributeError:
+            super().__setattr__(key, val)
+
+    def as_dict(self):
+        """
+        Returns all the metadata as a dict.
+        Note that modifications to the returned dict will not reflect on the Metadata object.
+        """
+        return copy.copy(self.__dict__)
+
+    def set(self, **kwargs):
+        """
+        Set multiple metadata with kwargs.
+        """
+        for k, v in kwargs.items():
+            setattr(self, k, v)
+        return self
+
+    def get(self, key, default=None):
+        """
+        Access an attribute and return its value if exists.
+        Otherwise return default.
+        """
+        try:
+            return getattr(self, key)
+        except AttributeError:
+            return default
+
+
+class _MetadataCatalog(UserDict):
+    """
+    MetadataCatalog is a global dictionary that provides access to
+    :class:`Metadata` of a given dataset.
+
+    The metadata associated with a certain name is a singleton: once created, the
+    metadata will stay alive and will be returned by future calls to ``get(name)``.
+
+    It's like global variables, so don't abuse it.
+    It's meant for storing knowledge that's constant and shared across the execution
+    of the program, e.g.: the class names in COCO.
+    """
+
+    def get(self, name):
+        """
+        Args:
+            name (str): name of a dataset (e.g. coco_2014_train).
+
+        Returns:
+            Metadata: The :class:`Metadata` instance associated with this name,
+            or create an empty one if none is available.
+        """
+        assert len(name)
+        r = super().get(name, None)
+        if r is None:
+            r = self[name] = Metadata(name=name)
+        return r
+
+    def list(self):
+        """
+        List all registered metadata.
+
+        Returns:
+            list[str]: keys (names of datasets) of all registered metadata
+        """
+        return list(self.keys())
+
+    def remove(self, name):
+        """
+        Alias of ``pop``.
+        """
+        self.pop(name)
+
+    def __str__(self):
+        return "MetadataCatalog(registered metadata: {})".format(", ".join(self.keys()))
+
+    __repr__ = __str__
+
+
+MetadataCatalog = _MetadataCatalog()
+MetadataCatalog.__doc__ = (
+    _MetadataCatalog.__doc__
+    + """
+    .. automethod:: detectron2.data.catalog.MetadataCatalog.get
+"""
+)

detectron2/data/common.py

@@ -0,0 +1,339 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import contextlib
+import copy
+import itertools
+import logging
+import numpy as np
+import pickle
+import random
+from typing import Callable, Union
+import torch
+import torch.utils.data as data
+from torch.utils.data.sampler import Sampler
+
+from detectron2.utils.serialize import PicklableWrapper
+
+__all__ = ["MapDataset", "DatasetFromList", "AspectRatioGroupedDataset", "ToIterableDataset"]
+
+logger = logging.getLogger(__name__)
+
+
+# copied from: https://docs.python.org/3/library/itertools.html#recipes
+def _roundrobin(*iterables):
+    "roundrobin('ABC', 'D', 'EF') --> A D E B F C"
+    # Recipe credited to George Sakkis
+    num_active = len(iterables)
+    nexts = itertools.cycle(iter(it).__next__ for it in iterables)
+    while num_active:
+        try:
+            for next in nexts:
+                yield next()
+        except StopIteration:
+            # Remove the iterator we just exhausted from the cycle.
+            num_active -= 1
+            nexts = itertools.cycle(itertools.islice(nexts, num_active))
+
+
+def _shard_iterator_dataloader_worker(iterable, chunk_size=1):
+    # Shard the iterable if we're currently inside pytorch dataloader worker.
+    worker_info = data.get_worker_info()
+    if worker_info is None or worker_info.num_workers == 1:
+        # do nothing
+        yield from iterable
+    else:
+        # worker0: 0, 1, ..., chunk_size-1, num_workers*chunk_size, num_workers*chunk_size+1, ...
+        # worker1: chunk_size, chunk_size+1, ...
+        # worker2: 2*chunk_size, 2*chunk_size+1, ...
+        # ...
+        yield from _roundrobin(
+            *[
+                itertools.islice(
+                    iterable,
+                    worker_info.id * chunk_size + chunk_i,
+                    None,
+                    worker_info.num_workers * chunk_size,
+                )
+                for chunk_i in range(chunk_size)
+            ]
+        )
+
+
+class _MapIterableDataset(data.IterableDataset):
+    """
+    Map a function over elements in an IterableDataset.
+
+    Similar to pytorch's MapIterDataPipe, but support filtering when map_func
+    returns None.
+
+    This class is not public-facing. Will be called by `MapDataset`.
+    """
+
+    def __init__(self, dataset, map_func):
+        self._dataset = dataset
+        self._map_func = PicklableWrapper(map_func)  # wrap so that a lambda will work
+
+    def __len__(self):
+        return len(self._dataset)
+
+    def __iter__(self):
+        for x in map(self._map_func, self._dataset):
+            if x is not None:
+                yield x
+
+
+class MapDataset(data.Dataset):
+    """
+    Map a function over the elements in a dataset.
+    """
+
+    def __init__(self, dataset, map_func):
+        """
+        Args:
+            dataset: a dataset where map function is applied. Can be either
+                map-style or iterable dataset. When given an iterable dataset,
+                the returned object will also be an iterable dataset.
+            map_func: a callable which maps the element in dataset. map_func can
+                return None to skip the data (e.g. in case of errors).
+                How None is handled depends on the style of `dataset`.
+                If `dataset` is map-style, it randomly tries other elements.
+                If `dataset` is iterable, it skips the data and tries the next.
+        """
+        self._dataset = dataset
+        self._map_func = PicklableWrapper(map_func)  # wrap so that a lambda will work
+
+        self._rng = random.Random(42)
+        self._fallback_candidates = set(range(len(dataset)))
+
+    def __new__(cls, dataset, map_func):
+        is_iterable = isinstance(dataset, data.IterableDataset)
+        if is_iterable:
+            return _MapIterableDataset(dataset, map_func)
+        else:
+            return super().__new__(cls)
+
+    def __getnewargs__(self):
+        return self._dataset, self._map_func
+
+    def __len__(self):
+        return len(self._dataset)
+
+    def __getitem__(self, idx):
+        retry_count = 0
+        cur_idx = int(idx)
+
+        while True:
+            data = self._map_func(self._dataset[cur_idx])
+            if data is not None:
+                self._fallback_candidates.add(cur_idx)
+                return data
+
+            # _map_func fails for this idx, use a random new index from the pool
+            retry_count += 1
+            self._fallback_candidates.discard(cur_idx)
+            cur_idx = self._rng.sample(self._fallback_candidates, k=1)[0]
+
+            if retry_count >= 3:
+                logger = logging.getLogger(__name__)
+                logger.warning(
+                    "Failed to apply `_map_func` for idx: {}, retry count: {}".format(
+                        idx, retry_count
+                    )
+                )
+
+
+class _TorchSerializedList:
+    """
+    A list-like object whose items are serialized and stored in a torch tensor. When
+    launching a process that uses TorchSerializedList with "fork" start method,
+    the subprocess can read the same buffer without triggering copy-on-access. When
+    launching a process that uses TorchSerializedList with "spawn/forkserver" start
+    method, the list will be pickled by a special ForkingPickler registered by PyTorch
+    that moves data to shared memory. In both cases, this allows parent and child
+    processes to share RAM for the list data, hence avoids the issue in
+    https://github.com/pytorch/pytorch/issues/13246.
+
+    See also https://ppwwyyxx.com/blog/2022/Demystify-RAM-Usage-in-Multiprocess-DataLoader/
+    on how it works.
+    """
+
+    def __init__(self, lst: list):
+        self._lst = lst
+
+        def _serialize(data):
+            buffer = pickle.dumps(data, protocol=-1)
+            return np.frombuffer(buffer, dtype=np.uint8)
+
+        logger.info(
+            "Serializing {} elements to byte tensors and concatenating them all ...".format(
+                len(self._lst)
+            )
+        )
+        self._lst = [_serialize(x) for x in self._lst]
+        self._addr = np.asarray([len(x) for x in self._lst], dtype=np.int64)
+        self._addr = torch.from_numpy(np.cumsum(self._addr))
+        self._lst = torch.from_numpy(np.concatenate(self._lst))
+        logger.info("Serialized dataset takes {:.2f} MiB".format(len(self._lst) / 1024**2))
+
+    def __len__(self):
+        return len(self._addr)
+
+    def __getitem__(self, idx):
+        start_addr = 0 if idx == 0 else self._addr[idx - 1].item()
+        end_addr = self._addr[idx].item()
+        bytes = memoryview(self._lst[start_addr:end_addr].numpy())
+
+        # @lint-ignore PYTHONPICKLEISBAD
+        return pickle.loads(bytes)
+
+
+_DEFAULT_DATASET_FROM_LIST_SERIALIZE_METHOD = _TorchSerializedList
+
+
+@contextlib.contextmanager
+def set_default_dataset_from_list_serialize_method(new):
+    """
+    Context manager for using custom serialize function when creating DatasetFromList
+    """
+
+    global _DEFAULT_DATASET_FROM_LIST_SERIALIZE_METHOD
+    orig = _DEFAULT_DATASET_FROM_LIST_SERIALIZE_METHOD
+    _DEFAULT_DATASET_FROM_LIST_SERIALIZE_METHOD = new
+    yield
+    _DEFAULT_DATASET_FROM_LIST_SERIALIZE_METHOD = orig
+
+
+class DatasetFromList(data.Dataset):
+    """
+    Wrap a list to a torch Dataset. It produces elements of the list as data.
+    """
+
+    def __init__(
+        self,
+        lst: list,
+        copy: bool = True,
+        serialize: Union[bool, Callable] = True,
+    ):
+        """
+        Args:
+            lst (list): a list which contains elements to produce.
+            copy (bool): whether to deepcopy the element when producing it,
+                so that the result can be modified in place without affecting the
+                source in the list.
+            serialize (bool or callable): whether to serialize the stroage to other
+                backend. If `True`, the default serialize method will be used, if given
+                a callable, the callable will be used as serialize method.
+        """
+        self._lst = lst
+        self._copy = copy
+        if not isinstance(serialize, (bool, Callable)):
+            raise TypeError(f"Unsupported type for argument `serailzie`: {serialize}")
+        self._serialize = serialize is not False
+
+        if self._serialize:
+            serialize_method = (
+                serialize
+                if isinstance(serialize, Callable)
+                else _DEFAULT_DATASET_FROM_LIST_SERIALIZE_METHOD
+            )
+            logger.info(f"Serializing the dataset using: {serialize_method}")
+            self._lst = serialize_method(self._lst)
+
+    def __len__(self):
+        return len(self._lst)
+
+    def __getitem__(self, idx):
+        if self._copy and not self._serialize:
+            return copy.deepcopy(self._lst[idx])
+        else:
+            return self._lst[idx]
+
+
+class ToIterableDataset(data.IterableDataset):
+    """
+    Convert an old indices-based (also called map-style) dataset
+    to an iterable-style dataset.
+    """
+
+    def __init__(
+        self,
+        dataset: data.Dataset,
+        sampler: Sampler,
+        shard_sampler: bool = True,
+        shard_chunk_size: int = 1,
+    ):
+        """
+        Args:
+            dataset: an old-style dataset with ``__getitem__``
+            sampler: a cheap iterable that produces indices to be applied on ``dataset``.
+            shard_sampler: whether to shard the sampler based on the current pytorch data loader
+                worker id. When an IterableDataset is forked by pytorch's DataLoader into multiple
+                workers, it is responsible for sharding its data based on worker id so that workers
+                don't produce identical data.
+
+                Most samplers (like our TrainingSampler) do not shard based on dataloader worker id
+                and this argument should be set to True. But certain samplers may be already
+                sharded, in that case this argument should be set to False.
+            shard_chunk_size: when sharding the sampler, each worker will
+        """
+        assert not isinstance(dataset, data.IterableDataset), dataset
+        assert isinstance(sampler, Sampler), sampler
+        self.dataset = dataset
+        self.sampler = sampler
+        self.shard_sampler = shard_sampler
+        self.shard_chunk_size = shard_chunk_size
+
+    def __iter__(self):
+        if not self.shard_sampler:
+            sampler = self.sampler
+        else:
+            # With map-style dataset, `DataLoader(dataset, sampler)` runs the
+            # sampler in main process only. But `DataLoader(ToIterableDataset(dataset, sampler))`
+            # will run sampler in every of the N worker. So we should only keep 1/N of the ids on
+            # each worker. The assumption is that sampler is cheap to iterate so it's fine to
+            # discard ids in workers.
+            sampler = _shard_iterator_dataloader_worker(self.sampler, self.shard_chunk_size)
+        for idx in sampler:
+            yield self.dataset[idx]
+
+    def __len__(self):
+        return len(self.sampler)
+
+
+class AspectRatioGroupedDataset(data.IterableDataset):
+    """
+    Batch data that have similar aspect ratio together.
+    In this implementation, images whose aspect ratio < (or >) 1 will
+    be batched together.
+    This improves training speed because the images then need less padding
+    to form a batch.
+
+    It assumes the underlying dataset produces dicts with "width" and "height" keys.
+    It will then produce a list of original dicts with length = batch_size,
+    all with similar aspect ratios.
+    """
+
+    def __init__(self, dataset, batch_size):
+        """
+        Args:
+            dataset: an iterable. Each element must be a dict with keys
+                "width" and "height", which will be used to batch data.
+            batch_size (int):
+        """
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self._buckets = [[] for _ in range(2)]
+        # Hard-coded two aspect ratio groups: w > h and w < h.
+        # Can add support for more aspect ratio groups, but doesn't seem useful
+
+    def __iter__(self):
+        for d in self.dataset:
+            w, h = d["width"], d["height"]
+            bucket_id = 0 if w > h else 1
+            bucket = self._buckets[bucket_id]
+            bucket.append(d)
+            if len(bucket) == self.batch_size:
+                data = bucket[:]
+                # Clear bucket first, because code after yield is not
+                # guaranteed to execute
+                del bucket[:]
+                yield data

detectron2/data/dataset_mapper.py

@@ -0,0 +1,191 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import copy
+import logging
+import numpy as np
+from typing import List, Optional, Union
+import torch
+
+from detectron2.config import configurable
+
+from . import detection_utils as utils
+from . import transforms as T
+
+"""
+This file contains the default mapping that's applied to "dataset dicts".
+"""
+
+__all__ = ["DatasetMapper"]
+
+
+class DatasetMapper:
+    """
+    A callable which takes a dataset dict in Detectron2 Dataset format,
+    and map it into a format used by the model.
+
+    This is the default callable to be used to map your dataset dict into training data.
+    You may need to follow it to implement your own one for customized logic,
+    such as a different way to read or transform images.
+    See :doc:`/tutorials/data_loading` for details.
+
+    The callable currently does the following:
+
+    1. Read the image from "file_name"
+    2. Applies cropping/geometric transforms to the image and annotations
+    3. Prepare data and annotations to Tensor and :class:`Instances`
+    """
+
+    @configurable
+    def __init__(
+        self,
+        is_train: bool,
+        *,
+        augmentations: List[Union[T.Augmentation, T.Transform]],
+        image_format: str,
+        use_instance_mask: bool = False,
+        use_keypoint: bool = False,
+        instance_mask_format: str = "polygon",
+        keypoint_hflip_indices: Optional[np.ndarray] = None,
+        precomputed_proposal_topk: Optional[int] = None,
+        recompute_boxes: bool = False,
+    ):
+        """
+        NOTE: this interface is experimental.
+
+        Args:
+            is_train: whether it's used in training or inference
+            augmentations: a list of augmentations or deterministic transforms to apply
+            image_format: an image format supported by :func:`detection_utils.read_image`.
+            use_instance_mask: whether to process instance segmentation annotations, if available
+            use_keypoint: whether to process keypoint annotations if available
+            instance_mask_format: one of "polygon" or "bitmask". Process instance segmentation
+                masks into this format.
+            keypoint_hflip_indices: see :func:`detection_utils.create_keypoint_hflip_indices`
+            precomputed_proposal_topk: if given, will load pre-computed
+                proposals from dataset_dict and keep the top k proposals for each image.
+            recompute_boxes: whether to overwrite bounding box annotations
+                by computing tight bounding boxes from instance mask annotations.
+        """
+        if recompute_boxes:
+            assert use_instance_mask, "recompute_boxes requires instance masks"
+        # fmt: off
+        self.is_train               = is_train
+        self.augmentations          = T.AugmentationList(augmentations)
+        self.image_format           = image_format
+        self.use_instance_mask      = use_instance_mask
+        self.instance_mask_format   = instance_mask_format
+        self.use_keypoint           = use_keypoint
+        self.keypoint_hflip_indices = keypoint_hflip_indices
+        self.proposal_topk          = precomputed_proposal_topk
+        self.recompute_boxes        = recompute_boxes
+        # fmt: on
+        logger = logging.getLogger(__name__)
+        mode = "training" if is_train else "inference"
+        logger.info(f"[DatasetMapper] Augmentations used in {mode}: {augmentations}")
+
+    @classmethod
+    def from_config(cls, cfg, is_train: bool = True):
+        augs = utils.build_augmentation(cfg, is_train)
+        if cfg.INPUT.CROP.ENABLED and is_train:
+            augs.insert(0, T.RandomCrop(cfg.INPUT.CROP.TYPE, cfg.INPUT.CROP.SIZE))
+            recompute_boxes = cfg.MODEL.MASK_ON
+        else:
+            recompute_boxes = False
+
+        ret = {
+            "is_train": is_train,
+            "augmentations": augs,
+            "image_format": cfg.INPUT.FORMAT,
+            "use_instance_mask": cfg.MODEL.MASK_ON,
+            "instance_mask_format": cfg.INPUT.MASK_FORMAT,
+            "use_keypoint": cfg.MODEL.KEYPOINT_ON,
+            "recompute_boxes": recompute_boxes,
+        }
+
+        if cfg.MODEL.KEYPOINT_ON:
+            ret["keypoint_hflip_indices"] = utils.create_keypoint_hflip_indices(cfg.DATASETS.TRAIN)
+
+        if cfg.MODEL.LOAD_PROPOSALS:
+            ret["precomputed_proposal_topk"] = (
+                cfg.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TRAIN
+                if is_train
+                else cfg.DATASETS.PRECOMPUTED_PROPOSAL_TOPK_TEST
+            )
+        return ret
+
+    def _transform_annotations(self, dataset_dict, transforms, image_shape):
+        # USER: Modify this if you want to keep them for some reason.
+        for anno in dataset_dict["annotations"]:
+            if not self.use_instance_mask:
+                anno.pop("segmentation", None)
+            if not self.use_keypoint:
+                anno.pop("keypoints", None)
+
+        # USER: Implement additional transformations if you have other types of data
+        annos = [
+            utils.transform_instance_annotations(
+                obj, transforms, image_shape, keypoint_hflip_indices=self.keypoint_hflip_indices
+            )
+            for obj in dataset_dict.pop("annotations")
+            if obj.get("iscrowd", 0) == 0
+        ]
+        instances = utils.annotations_to_instances(
+            annos, image_shape, mask_format=self.instance_mask_format
+        )
+
+        # After transforms such as cropping are applied, the bounding box may no longer
+        # tightly bound the object. As an example, imagine a triangle object
+        # [(0,0), (2,0), (0,2)] cropped by a box [(1,0),(2,2)] (XYXY format). The tight
+        # bounding box of the cropped triangle should be [(1,0),(2,1)], which is not equal to
+        # the intersection of original bounding box and the cropping box.
+        if self.recompute_boxes:
+            instances.gt_boxes = instances.gt_masks.get_bounding_boxes()
+        dataset_dict["instances"] = utils.filter_empty_instances(instances)
+
+    def __call__(self, dataset_dict):
+        """
+        Args:
+            dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format.
+
+        Returns:
+            dict: a format that builtin models in detectron2 accept
+        """
+        dataset_dict = copy.deepcopy(dataset_dict)  # it will be modified by code below
+        # USER: Write your own image loading if it's not from a file
+        image = utils.read_image(dataset_dict["file_name"], format=self.image_format)
+        utils.check_image_size(dataset_dict, image)
+
+        # USER: Remove if you don't do semantic/panoptic segmentation.
+        if "sem_seg_file_name" in dataset_dict:
+            sem_seg_gt = utils.read_image(dataset_dict.pop("sem_seg_file_name"), "L").squeeze(2)
+        else:
+            sem_seg_gt = None
+
+        aug_input = T.AugInput(image, sem_seg=sem_seg_gt)
+        transforms = self.augmentations(aug_input)
+        image, sem_seg_gt = aug_input.image, aug_input.sem_seg
+
+        image_shape = image.shape[:2]  # h, w
+        # Pytorch's dataloader is efficient on torch.Tensor due to shared-memory,
+        # but not efficient on large generic data structures due to the use of pickle & mp.Queue.
+        # Therefore it's important to use torch.Tensor.
+        dataset_dict["image"] = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1)))
+        if sem_seg_gt is not None:
+            dataset_dict["sem_seg"] = torch.as_tensor(sem_seg_gt.astype("long"))
+
+        # USER: Remove if you don't use pre-computed proposals.
+        # Most users would not need this feature.
+        if self.proposal_topk is not None:
+            utils.transform_proposals(
+                dataset_dict, image_shape, transforms, proposal_topk=self.proposal_topk
+            )
+
+        if not self.is_train:
+            # USER: Modify this if you want to keep them for some reason.
+            dataset_dict.pop("annotations", None)
+            dataset_dict.pop("sem_seg_file_name", None)
+            return dataset_dict
+
+        if "annotations" in dataset_dict:
+            self._transform_annotations(dataset_dict, transforms, image_shape)
+
+        return dataset_dict

detectron2/data/datasets/README.md

@@ -0,0 +1,9 @@
+
+
+### Common Datasets
+
+The dataset implemented here do not need to load the data into the final format.
+It should provide the minimal data structure needed to use the dataset, so it can be very efficient.
+
+For example, for an image dataset, just provide the file names and labels, but don't read the images.
+Let the downstream decide how to read.

detectron2/data/datasets/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+from .coco import load_coco_json, load_sem_seg, register_coco_instances, convert_to_coco_json
+from .coco_panoptic import register_coco_panoptic, register_coco_panoptic_separated
+from .lvis import load_lvis_json, register_lvis_instances, get_lvis_instances_meta
+from .pascal_voc import load_voc_instances, register_pascal_voc
+from . import builtin as _builtin  # ensure the builtin datasets are registered
+
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

detectron2/data/datasets/__pycache__/lvis_v0_5_categories.cpython-311.pyc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f78d618393dda851251dec67a12ffcd7a3e3092d4fd6fdf912213c137ec2aef7
+size 269004

detectron2/data/datasets/__pycache__/lvis_v1_categories.cpython-311.pyc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f02029d25d9cd8351c594670b7b1f7cc0c61b6cfb2bf0f896b363abc9f33832
+size 263476

detectron2/data/datasets/builtin.py

@@ -0,0 +1,259 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+
+"""
+This file registers pre-defined datasets at hard-coded paths, and their metadata.
+
+We hard-code metadata for common datasets. This will enable:
+1. Consistency check when loading the datasets
+2. Use models on these standard datasets directly and run demos,
+   without having to download the dataset annotations
+
+We hard-code some paths to the dataset that's assumed to
+exist in "./datasets/".
+
+Users SHOULD NOT use this file to create new dataset / metadata for new dataset.
+To add new dataset, refer to the tutorial "docs/DATASETS.md".
+"""
+
+import os
+
+from detectron2.data import DatasetCatalog, MetadataCatalog
+
+from .builtin_meta import ADE20K_SEM_SEG_CATEGORIES, _get_builtin_metadata
+from .cityscapes import load_cityscapes_instances, load_cityscapes_semantic
+from .cityscapes_panoptic import register_all_cityscapes_panoptic
+from .coco import load_sem_seg, register_coco_instances
+from .coco_panoptic import register_coco_panoptic, register_coco_panoptic_separated
+from .lvis import get_lvis_instances_meta, register_lvis_instances
+from .pascal_voc import register_pascal_voc
+
+# ==== Predefined datasets and splits for COCO ==========
+
+_PREDEFINED_SPLITS_COCO = {}
+_PREDEFINED_SPLITS_COCO["coco"] = {
+    "coco_2014_train": ("coco/train2014", "coco/annotations/instances_train2014.json"),
+    "coco_2014_val": ("coco/val2014", "coco/annotations/instances_val2014.json"),
+    "coco_2014_minival": ("coco/val2014", "coco/annotations/instances_minival2014.json"),
+    "coco_2014_valminusminival": (
+        "coco/val2014",
+        "coco/annotations/instances_valminusminival2014.json",
+    ),
+    "coco_2017_train": ("coco/train2017", "coco/annotations/instances_train2017.json"),
+    "coco_2017_val": ("coco/val2017", "coco/annotations/instances_val2017.json"),
+    "coco_2017_test": ("coco/test2017", "coco/annotations/image_info_test2017.json"),
+    "coco_2017_test-dev": ("coco/test2017", "coco/annotations/image_info_test-dev2017.json"),
+    "coco_2017_val_100": ("coco/val2017", "coco/annotations/instances_val2017_100.json"),
+}
+
+_PREDEFINED_SPLITS_COCO["coco_person"] = {
+    "keypoints_coco_2014_train": (
+        "coco/train2014",
+        "coco/annotations/person_keypoints_train2014.json",
+    ),
+    "keypoints_coco_2014_val": ("coco/val2014", "coco/annotations/person_keypoints_val2014.json"),
+    "keypoints_coco_2014_minival": (
+        "coco/val2014",
+        "coco/annotations/person_keypoints_minival2014.json",
+    ),
+    "keypoints_coco_2014_valminusminival": (
+        "coco/val2014",
+        "coco/annotations/person_keypoints_valminusminival2014.json",
+    ),
+    "keypoints_coco_2017_train": (
+        "coco/train2017",
+        "coco/annotations/person_keypoints_train2017.json",
+    ),
+    "keypoints_coco_2017_val": ("coco/val2017", "coco/annotations/person_keypoints_val2017.json"),
+    "keypoints_coco_2017_val_100": (
+        "coco/val2017",
+        "coco/annotations/person_keypoints_val2017_100.json",
+    ),
+}
+
+
+_PREDEFINED_SPLITS_COCO_PANOPTIC = {
+    "coco_2017_train_panoptic": (
+        # This is the original panoptic annotation directory
+        "coco/panoptic_train2017",
+        "coco/annotations/panoptic_train2017.json",
+        # This directory contains semantic annotations that are
+        # converted from panoptic annotations.
+        # It is used by PanopticFPN.
+        # You can use the script at detectron2/datasets/prepare_panoptic_fpn.py
+        # to create these directories.
+        "coco/panoptic_stuff_train2017",
+    ),
+    "coco_2017_val_panoptic": (
+        "coco/panoptic_val2017",
+        "coco/annotations/panoptic_val2017.json",
+        "coco/panoptic_stuff_val2017",
+    ),
+    "coco_2017_val_100_panoptic": (
+        "coco/panoptic_val2017_100",
+        "coco/annotations/panoptic_val2017_100.json",
+        "coco/panoptic_stuff_val2017_100",
+    ),
+}
+
+
+def register_all_coco(root):
+    for dataset_name, splits_per_dataset in _PREDEFINED_SPLITS_COCO.items():
+        for key, (image_root, json_file) in splits_per_dataset.items():
+            # Assume pre-defined datasets live in `./datasets`.
+            register_coco_instances(
+                key,
+                _get_builtin_metadata(dataset_name),
+                os.path.join(root, json_file) if "://" not in json_file else json_file,
+                os.path.join(root, image_root),
+            )
+
+    for (
+        prefix,
+        (panoptic_root, panoptic_json, semantic_root),
+    ) in _PREDEFINED_SPLITS_COCO_PANOPTIC.items():
+        prefix_instances = prefix[: -len("_panoptic")]
+        instances_meta = MetadataCatalog.get(prefix_instances)
+        image_root, instances_json = instances_meta.image_root, instances_meta.json_file
+        # The "separated" version of COCO panoptic segmentation dataset,
+        # e.g. used by Panoptic FPN
+        register_coco_panoptic_separated(
+            prefix,
+            _get_builtin_metadata("coco_panoptic_separated"),
+            image_root,
+            os.path.join(root, panoptic_root),
+            os.path.join(root, panoptic_json),
+            os.path.join(root, semantic_root),
+            instances_json,
+        )
+        # The "standard" version of COCO panoptic segmentation dataset,
+        # e.g. used by Panoptic-DeepLab
+        register_coco_panoptic(
+            prefix,
+            _get_builtin_metadata("coco_panoptic_standard"),
+            image_root,
+            os.path.join(root, panoptic_root),
+            os.path.join(root, panoptic_json),
+            instances_json,
+        )
+
+
+# ==== Predefined datasets and splits for LVIS ==========
+
+
+_PREDEFINED_SPLITS_LVIS = {
+    "lvis_v1": {
+        "lvis_v1_train": ("coco/", "lvis/lvis_v1_train.json"),
+        "lvis_v1_val": ("coco/", "lvis/lvis_v1_val.json"),
+        "lvis_v1_test_dev": ("coco/", "lvis/lvis_v1_image_info_test_dev.json"),
+        "lvis_v1_test_challenge": ("coco/", "lvis/lvis_v1_image_info_test_challenge.json"),
+    },
+    "lvis_v0.5": {
+        "lvis_v0.5_train": ("coco/", "lvis/lvis_v0.5_train.json"),
+        "lvis_v0.5_val": ("coco/", "lvis/lvis_v0.5_val.json"),
+        "lvis_v0.5_val_rand_100": ("coco/", "lvis/lvis_v0.5_val_rand_100.json"),
+        "lvis_v0.5_test": ("coco/", "lvis/lvis_v0.5_image_info_test.json"),
+    },
+    "lvis_v0.5_cocofied": {
+        "lvis_v0.5_train_cocofied": ("coco/", "lvis/lvis_v0.5_train_cocofied.json"),
+        "lvis_v0.5_val_cocofied": ("coco/", "lvis/lvis_v0.5_val_cocofied.json"),
+    },
+}
+
+
+def register_all_lvis(root):
+    for dataset_name, splits_per_dataset in _PREDEFINED_SPLITS_LVIS.items():
+        for key, (image_root, json_file) in splits_per_dataset.items():
+            register_lvis_instances(
+                key,
+                get_lvis_instances_meta(dataset_name),
+                os.path.join(root, json_file) if "://" not in json_file else json_file,
+                os.path.join(root, image_root),
+            )
+
+
+# ==== Predefined splits for raw cityscapes images ===========
+_RAW_CITYSCAPES_SPLITS = {
+    "cityscapes_fine_{task}_train": ("cityscapes/leftImg8bit/train/", "cityscapes/gtFine/train/"),
+    "cityscapes_fine_{task}_val": ("cityscapes/leftImg8bit/val/", "cityscapes/gtFine/val/"),
+    "cityscapes_fine_{task}_test": ("cityscapes/leftImg8bit/test/", "cityscapes/gtFine/test/"),
+}
+
+
+def register_all_cityscapes(root):
+    for key, (image_dir, gt_dir) in _RAW_CITYSCAPES_SPLITS.items():
+        meta = _get_builtin_metadata("cityscapes")
+        image_dir = os.path.join(root, image_dir)
+        gt_dir = os.path.join(root, gt_dir)
+
+        inst_key = key.format(task="instance_seg")
+        DatasetCatalog.register(
+            inst_key,
+            lambda x=image_dir, y=gt_dir: load_cityscapes_instances(
+                x, y, from_json=True, to_polygons=True
+            ),
+        )
+        MetadataCatalog.get(inst_key).set(
+            image_dir=image_dir, gt_dir=gt_dir, evaluator_type="cityscapes_instance", **meta
+        )
+
+        sem_key = key.format(task="sem_seg")
+        DatasetCatalog.register(
+            sem_key, lambda x=image_dir, y=gt_dir: load_cityscapes_semantic(x, y)
+        )
+        MetadataCatalog.get(sem_key).set(
+            image_dir=image_dir,
+            gt_dir=gt_dir,
+            evaluator_type="cityscapes_sem_seg",
+            ignore_label=255,
+            **meta,
+        )
+
+
+# ==== Predefined splits for PASCAL VOC ===========
+def register_all_pascal_voc(root):
+    SPLITS = [
+        ("voc_2007_trainval", "VOC2007", "trainval"),
+        ("voc_2007_train", "VOC2007", "train"),
+        ("voc_2007_val", "VOC2007", "val"),
+        ("voc_2007_test", "VOC2007", "test"),
+        ("voc_2012_trainval", "VOC2012", "trainval"),
+        ("voc_2012_train", "VOC2012", "train"),
+        ("voc_2012_val", "VOC2012", "val"),
+    ]
+    for name, dirname, split in SPLITS:
+        year = 2007 if "2007" in name else 2012
+        register_pascal_voc(name, os.path.join(root, dirname), split, year)
+        MetadataCatalog.get(name).evaluator_type = "pascal_voc"
+
+
+def register_all_ade20k(root):
+    root = os.path.join(root, "ADEChallengeData2016")
+    for name, dirname in [("train", "training"), ("val", "validation")]:
+        image_dir = os.path.join(root, "images", dirname)
+        gt_dir = os.path.join(root, "annotations_detectron2", dirname)
+        name = f"ade20k_sem_seg_{name}"
+        DatasetCatalog.register(
+            name, lambda x=image_dir, y=gt_dir: load_sem_seg(y, x, gt_ext="png", image_ext="jpg")
+        )
+        MetadataCatalog.get(name).set(
+            stuff_classes=ADE20K_SEM_SEG_CATEGORIES[:],
+            image_root=image_dir,
+            sem_seg_root=gt_dir,
+            evaluator_type="sem_seg",
+            ignore_label=255,
+        )
+
+
+# True for open source;
+# Internally at fb, we register them elsewhere
+if __name__.endswith(".builtin"):
+    # Assume pre-defined datasets live in `./datasets`.
+    _root = os.path.expanduser(os.getenv("DETECTRON2_DATASETS", "datasets"))
+    register_all_coco(_root)
+    register_all_lvis(_root)
+    register_all_cityscapes(_root)
+    register_all_cityscapes_panoptic(_root)
+    register_all_pascal_voc(_root)
+    register_all_ade20k(_root)

detectron2/data/datasets/builtin_meta.py

@@ -0,0 +1,350 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+"""
+Note:
+For your custom dataset, there is no need to hard-code metadata anywhere in the code.
+For example, for COCO-format dataset, metadata will be obtained automatically
+when calling `load_coco_json`. For other dataset, metadata may also be obtained in other ways
+during loading.
+
+However, we hard-coded metadata for a few common dataset here.
+The only goal is to allow users who don't have these dataset to use pre-trained models.
+Users don't have to download a COCO json (which contains metadata), in order to visualize a
+COCO model (with correct class names and colors).
+"""
+
+
+# All coco categories, together with their nice-looking visualization colors
+# It's from https://github.com/cocodataset/panopticapi/blob/master/panoptic_coco_categories.json
+COCO_CATEGORIES = [
+    {"color": [220, 20, 60], "isthing": 1, "id": 1, "name": "person"},
+    {"color": [119, 11, 32], "isthing": 1, "id": 2, "name": "bicycle"},
+    {"color": [0, 0, 142], "isthing": 1, "id": 3, "name": "car"},
+    {"color": [0, 0, 230], "isthing": 1, "id": 4, "name": "motorcycle"},
+    {"color": [106, 0, 228], "isthing": 1, "id": 5, "name": "airplane"},
+    {"color": [0, 60, 100], "isthing": 1, "id": 6, "name": "bus"},
+    {"color": [0, 80, 100], "isthing": 1, "id": 7, "name": "train"},
+    {"color": [0, 0, 70], "isthing": 1, "id": 8, "name": "truck"},
+    {"color": [0, 0, 192], "isthing": 1, "id": 9, "name": "boat"},
+    {"color": [250, 170, 30], "isthing": 1, "id": 10, "name": "traffic light"},
+    {"color": [100, 170, 30], "isthing": 1, "id": 11, "name": "fire hydrant"},
+    {"color": [220, 220, 0], "isthing": 1, "id": 13, "name": "stop sign"},
+    {"color": [175, 116, 175], "isthing": 1, "id": 14, "name": "parking meter"},
+    {"color": [250, 0, 30], "isthing": 1, "id": 15, "name": "bench"},
+    {"color": [165, 42, 42], "isthing": 1, "id": 16, "name": "bird"},
+    {"color": [255, 77, 255], "isthing": 1, "id": 17, "name": "cat"},
+    {"color": [0, 226, 252], "isthing": 1, "id": 18, "name": "dog"},
+    {"color": [182, 182, 255], "isthing": 1, "id": 19, "name": "horse"},
+    {"color": [0, 82, 0], "isthing": 1, "id": 20, "name": "sheep"},
+    {"color": [120, 166, 157], "isthing": 1, "id": 21, "name": "cow"},
+    {"color": [110, 76, 0], "isthing": 1, "id": 22, "name": "elephant"},
+    {"color": [174, 57, 255], "isthing": 1, "id": 23, "name": "bear"},
+    {"color": [199, 100, 0], "isthing": 1, "id": 24, "name": "zebra"},
+    {"color": [72, 0, 118], "isthing": 1, "id": 25, "name": "giraffe"},
+    {"color": [255, 179, 240], "isthing": 1, "id": 27, "name": "backpack"},
+    {"color": [0, 125, 92], "isthing": 1, "id": 28, "name": "umbrella"},
+    {"color": [209, 0, 151], "isthing": 1, "id": 31, "name": "handbag"},
+    {"color": [188, 208, 182], "isthing": 1, "id": 32, "name": "tie"},
+    {"color": [0, 220, 176], "isthing": 1, "id": 33, "name": "suitcase"},
+    {"color": [255, 99, 164], "isthing": 1, "id": 34, "name": "frisbee"},
+    {"color": [92, 0, 73], "isthing": 1, "id": 35, "name": "skis"},
+    {"color": [133, 129, 255], "isthing": 1, "id": 36, "name": "snowboard"},
+    {"color": [78, 180, 255], "isthing": 1, "id": 37, "name": "sports ball"},
+    {"color": [0, 228, 0], "isthing": 1, "id": 38, "name": "kite"},
+    {"color": [174, 255, 243], "isthing": 1, "id": 39, "name": "baseball bat"},
+    {"color": [45, 89, 255], "isthing": 1, "id": 40, "name": "baseball glove"},
+    {"color": [134, 134, 103], "isthing": 1, "id": 41, "name": "skateboard"},
+    {"color": [145, 148, 174], "isthing": 1, "id": 42, "name": "surfboard"},
+    {"color": [255, 208, 186], "isthing": 1, "id": 43, "name": "tennis racket"},
+    {"color": [197, 226, 255], "isthing": 1, "id": 44, "name": "bottle"},
+    {"color": [171, 134, 1], "isthing": 1, "id": 46, "name": "wine glass"},
+    {"color": [109, 63, 54], "isthing": 1, "id": 47, "name": "cup"},
+    {"color": [207, 138, 255], "isthing": 1, "id": 48, "name": "fork"},
+    {"color": [151, 0, 95], "isthing": 1, "id": 49, "name": "knife"},
+    {"color": [9, 80, 61], "isthing": 1, "id": 50, "name": "spoon"},
+    {"color": [84, 105, 51], "isthing": 1, "id": 51, "name": "bowl"},
+    {"color": [74, 65, 105], "isthing": 1, "id": 52, "name": "banana"},
+    {"color": [166, 196, 102], "isthing": 1, "id": 53, "name": "apple"},
+    {"color": [208, 195, 210], "isthing": 1, "id": 54, "name": "sandwich"},
+    {"color": [255, 109, 65], "isthing": 1, "id": 55, "name": "orange"},
+    {"color": [0, 143, 149], "isthing": 1, "id": 56, "name": "broccoli"},
+    {"color": [179, 0, 194], "isthing": 1, "id": 57, "name": "carrot"},
+    {"color": [209, 99, 106], "isthing": 1, "id": 58, "name": "hot dog"},
+    {"color": [5, 121, 0], "isthing": 1, "id": 59, "name": "pizza"},
+    {"color": [227, 255, 205], "isthing": 1, "id": 60, "name": "donut"},
+    {"color": [147, 186, 208], "isthing": 1, "id": 61, "name": "cake"},
+    {"color": [153, 69, 1], "isthing": 1, "id": 62, "name": "chair"},
+    {"color": [3, 95, 161], "isthing": 1, "id": 63, "name": "couch"},
+    {"color": [163, 255, 0], "isthing": 1, "id": 64, "name": "potted plant"},
+    {"color": [119, 0, 170], "isthing": 1, "id": 65, "name": "bed"},
+    {"color": [0, 182, 199], "isthing": 1, "id": 67, "name": "dining table"},
+    {"color": [0, 165, 120], "isthing": 1, "id": 70, "name": "toilet"},
+    {"color": [183, 130, 88], "isthing": 1, "id": 72, "name": "tv"},
+    {"color": [95, 32, 0], "isthing": 1, "id": 73, "name": "laptop"},
+    {"color": [130, 114, 135], "isthing": 1, "id": 74, "name": "mouse"},
+    {"color": [110, 129, 133], "isthing": 1, "id": 75, "name": "remote"},
+    {"color": [166, 74, 118], "isthing": 1, "id": 76, "name": "keyboard"},
+    {"color": [219, 142, 185], "isthing": 1, "id": 77, "name": "cell phone"},
+    {"color": [79, 210, 114], "isthing": 1, "id": 78, "name": "microwave"},
+    {"color": [178, 90, 62], "isthing": 1, "id": 79, "name": "oven"},
+    {"color": [65, 70, 15], "isthing": 1, "id": 80, "name": "toaster"},
+    {"color": [127, 167, 115], "isthing": 1, "id": 81, "name": "sink"},
+    {"color": [59, 105, 106], "isthing": 1, "id": 82, "name": "refrigerator"},
+    {"color": [142, 108, 45], "isthing": 1, "id": 84, "name": "book"},
+    {"color": [196, 172, 0], "isthing": 1, "id": 85, "name": "clock"},
+    {"color": [95, 54, 80], "isthing": 1, "id": 86, "name": "vase"},
+    {"color": [128, 76, 255], "isthing": 1, "id": 87, "name": "scissors"},
+    {"color": [201, 57, 1], "isthing": 1, "id": 88, "name": "teddy bear"},
+    {"color": [246, 0, 122], "isthing": 1, "id": 89, "name": "hair drier"},
+    {"color": [191, 162, 208], "isthing": 1, "id": 90, "name": "toothbrush"},
+    {"color": [255, 255, 128], "isthing": 0, "id": 92, "name": "banner"},
+    {"color": [147, 211, 203], "isthing": 0, "id": 93, "name": "blanket"},
+    {"color": [150, 100, 100], "isthing": 0, "id": 95, "name": "bridge"},
+    {"color": [168, 171, 172], "isthing": 0, "id": 100, "name": "cardboard"},
+    {"color": [146, 112, 198], "isthing": 0, "id": 107, "name": "counter"},
+    {"color": [210, 170, 100], "isthing": 0, "id": 109, "name": "curtain"},
+    {"color": [92, 136, 89], "isthing": 0, "id": 112, "name": "door-stuff"},
+    {"color": [218, 88, 184], "isthing": 0, "id": 118, "name": "floor-wood"},
+    {"color": [241, 129, 0], "isthing": 0, "id": 119, "name": "flower"},
+    {"color": [217, 17, 255], "isthing": 0, "id": 122, "name": "fruit"},
+    {"color": [124, 74, 181], "isthing": 0, "id": 125, "name": "gravel"},
+    {"color": [70, 70, 70], "isthing": 0, "id": 128, "name": "house"},
+    {"color": [255, 228, 255], "isthing": 0, "id": 130, "name": "light"},
+    {"color": [154, 208, 0], "isthing": 0, "id": 133, "name": "mirror-stuff"},
+    {"color": [193, 0, 92], "isthing": 0, "id": 138, "name": "net"},
+    {"color": [76, 91, 113], "isthing": 0, "id": 141, "name": "pillow"},
+    {"color": [255, 180, 195], "isthing": 0, "id": 144, "name": "platform"},
+    {"color": [106, 154, 176], "isthing": 0, "id": 145, "name": "playingfield"},
+    {"color": [230, 150, 140], "isthing": 0, "id": 147, "name": "railroad"},
+    {"color": [60, 143, 255], "isthing": 0, "id": 148, "name": "river"},
+    {"color": [128, 64, 128], "isthing": 0, "id": 149, "name": "road"},
+    {"color": [92, 82, 55], "isthing": 0, "id": 151, "name": "roof"},
+    {"color": [254, 212, 124], "isthing": 0, "id": 154, "name": "sand"},
+    {"color": [73, 77, 174], "isthing": 0, "id": 155, "name": "sea"},
+    {"color": [255, 160, 98], "isthing": 0, "id": 156, "name": "shelf"},
+    {"color": [255, 255, 255], "isthing": 0, "id": 159, "name": "snow"},
+    {"color": [104, 84, 109], "isthing": 0, "id": 161, "name": "stairs"},
+    {"color": [169, 164, 131], "isthing": 0, "id": 166, "name": "tent"},
+    {"color": [225, 199, 255], "isthing": 0, "id": 168, "name": "towel"},
+    {"color": [137, 54, 74], "isthing": 0, "id": 171, "name": "wall-brick"},
+    {"color": [135, 158, 223], "isthing": 0, "id": 175, "name": "wall-stone"},
+    {"color": [7, 246, 231], "isthing": 0, "id": 176, "name": "wall-tile"},
+    {"color": [107, 255, 200], "isthing": 0, "id": 177, "name": "wall-wood"},
+    {"color": [58, 41, 149], "isthing": 0, "id": 178, "name": "water-other"},
+    {"color": [183, 121, 142], "isthing": 0, "id": 180, "name": "window-blind"},
+    {"color": [255, 73, 97], "isthing": 0, "id": 181, "name": "window-other"},
+    {"color": [107, 142, 35], "isthing": 0, "id": 184, "name": "tree-merged"},
+    {"color": [190, 153, 153], "isthing": 0, "id": 185, "name": "fence-merged"},
+    {"color": [146, 139, 141], "isthing": 0, "id": 186, "name": "ceiling-merged"},
+    {"color": [70, 130, 180], "isthing": 0, "id": 187, "name": "sky-other-merged"},
+    {"color": [134, 199, 156], "isthing": 0, "id": 188, "name": "cabinet-merged"},
+    {"color": [209, 226, 140], "isthing": 0, "id": 189, "name": "table-merged"},
+    {"color": [96, 36, 108], "isthing": 0, "id": 190, "name": "floor-other-merged"},
+    {"color": [96, 96, 96], "isthing": 0, "id": 191, "name": "pavement-merged"},
+    {"color": [64, 170, 64], "isthing": 0, "id": 192, "name": "mountain-merged"},
+    {"color": [152, 251, 152], "isthing": 0, "id": 193, "name": "grass-merged"},
+    {"color": [208, 229, 228], "isthing": 0, "id": 194, "name": "dirt-merged"},
+    {"color": [206, 186, 171], "isthing": 0, "id": 195, "name": "paper-merged"},
+    {"color": [152, 161, 64], "isthing": 0, "id": 196, "name": "food-other-merged"},
+    {"color": [116, 112, 0], "isthing": 0, "id": 197, "name": "building-other-merged"},
+    {"color": [0, 114, 143], "isthing": 0, "id": 198, "name": "rock-merged"},
+    {"color": [102, 102, 156], "isthing": 0, "id": 199, "name": "wall-other-merged"},
+    {"color": [250, 141, 255], "isthing": 0, "id": 200, "name": "rug-merged"},
+]
+
+# fmt: off
+COCO_PERSON_KEYPOINT_NAMES = (
+    "nose",
+    "left_eye", "right_eye",
+    "left_ear", "right_ear",
+    "left_shoulder", "right_shoulder",
+    "left_elbow", "right_elbow",
+    "left_wrist", "right_wrist",
+    "left_hip", "right_hip",
+    "left_knee", "right_knee",
+    "left_ankle", "right_ankle",
+)
+# fmt: on
+
+# Pairs of keypoints that should be exchanged under horizontal flipping
+COCO_PERSON_KEYPOINT_FLIP_MAP = (
+    ("left_eye", "right_eye"),
+    ("left_ear", "right_ear"),
+    ("left_shoulder", "right_shoulder"),
+    ("left_elbow", "right_elbow"),
+    ("left_wrist", "right_wrist"),
+    ("left_hip", "right_hip"),
+    ("left_knee", "right_knee"),
+    ("left_ankle", "right_ankle"),
+)
+
+# rules for pairs of keypoints to draw a line between, and the line color to use.
+KEYPOINT_CONNECTION_RULES = [
+    # face
+    ("left_ear", "left_eye", (102, 204, 255)),
+    ("right_ear", "right_eye", (51, 153, 255)),
+    ("left_eye", "nose", (102, 0, 204)),
+    ("nose", "right_eye", (51, 102, 255)),
+    # upper-body
+    ("left_shoulder", "right_shoulder", (255, 128, 0)),
+    ("left_shoulder", "left_elbow", (153, 255, 204)),
+    ("right_shoulder", "right_elbow", (128, 229, 255)),
+    ("left_elbow", "left_wrist", (153, 255, 153)),
+    ("right_elbow", "right_wrist", (102, 255, 224)),
+    # lower-body
+    ("left_hip", "right_hip", (255, 102, 0)),
+    ("left_hip", "left_knee", (255, 255, 77)),
+    ("right_hip", "right_knee", (153, 255, 204)),
+    ("left_knee", "left_ankle", (191, 255, 128)),
+    ("right_knee", "right_ankle", (255, 195, 77)),
+]
+
+# All Cityscapes categories, together with their nice-looking visualization colors
+# It's from https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/helpers/labels.py  # noqa
+CITYSCAPES_CATEGORIES = [
+    {"color": (128, 64, 128), "isthing": 0, "id": 7, "trainId": 0, "name": "road"},
+    {"color": (244, 35, 232), "isthing": 0, "id": 8, "trainId": 1, "name": "sidewalk"},
+    {"color": (70, 70, 70), "isthing": 0, "id": 11, "trainId": 2, "name": "building"},
+    {"color": (102, 102, 156), "isthing": 0, "id": 12, "trainId": 3, "name": "wall"},
+    {"color": (190, 153, 153), "isthing": 0, "id": 13, "trainId": 4, "name": "fence"},
+    {"color": (153, 153, 153), "isthing": 0, "id": 17, "trainId": 5, "name": "pole"},
+    {"color": (250, 170, 30), "isthing": 0, "id": 19, "trainId": 6, "name": "traffic light"},
+    {"color": (220, 220, 0), "isthing": 0, "id": 20, "trainId": 7, "name": "traffic sign"},
+    {"color": (107, 142, 35), "isthing": 0, "id": 21, "trainId": 8, "name": "vegetation"},
+    {"color": (152, 251, 152), "isthing": 0, "id": 22, "trainId": 9, "name": "terrain"},
+    {"color": (70, 130, 180), "isthing": 0, "id": 23, "trainId": 10, "name": "sky"},
+    {"color": (220, 20, 60), "isthing": 1, "id": 24, "trainId": 11, "name": "person"},
+    {"color": (255, 0, 0), "isthing": 1, "id": 25, "trainId": 12, "name": "rider"},
+    {"color": (0, 0, 142), "isthing": 1, "id": 26, "trainId": 13, "name": "car"},
+    {"color": (0, 0, 70), "isthing": 1, "id": 27, "trainId": 14, "name": "truck"},
+    {"color": (0, 60, 100), "isthing": 1, "id": 28, "trainId": 15, "name": "bus"},
+    {"color": (0, 80, 100), "isthing": 1, "id": 31, "trainId": 16, "name": "train"},
+    {"color": (0, 0, 230), "isthing": 1, "id": 32, "trainId": 17, "name": "motorcycle"},
+    {"color": (119, 11, 32), "isthing": 1, "id": 33, "trainId": 18, "name": "bicycle"},
+]
+
+# fmt: off
+ADE20K_SEM_SEG_CATEGORIES = [
+    "wall", "building", "sky", "floor", "tree", "ceiling", "road, route", "bed", "window ", "grass", "cabinet", "sidewalk, pavement", "person", "earth, ground", "door", "table", "mountain, mount", "plant", "curtain", "chair", "car", "water", "painting, picture", "sofa", "shelf", "house", "sea", "mirror", "rug", "field", "armchair", "seat", "fence", "desk", "rock, stone", "wardrobe, closet, press", "lamp", "tub", "rail", "cushion", "base, pedestal, stand", "box", "column, pillar", "signboard, sign", "chest of drawers, chest, bureau, dresser", "counter", "sand", "sink", "skyscraper", "fireplace", "refrigerator, icebox", "grandstand, covered stand", "path", "stairs", "runway", "case, display case, showcase, vitrine", "pool table, billiard table, snooker table", "pillow", "screen door, screen", "stairway, staircase", "river", "bridge, span", "bookcase", "blind, screen", "coffee table", "toilet, can, commode, crapper, pot, potty, stool, throne", "flower", "book", "hill", "bench", "countertop", "stove", "palm, palm tree", "kitchen island", "computer", "swivel chair", "boat", "bar", "arcade machine", "hovel, hut, hutch, shack, shanty", "bus", "towel", "light", "truck", "tower", "chandelier", "awning, sunshade, sunblind", "street lamp", "booth", "tv", "plane", "dirt track", "clothes", "pole", "land, ground, soil", "bannister, banister, balustrade, balusters, handrail", "escalator, moving staircase, moving stairway", "ottoman, pouf, pouffe, puff, hassock", "bottle", "buffet, counter, sideboard", "poster, posting, placard, notice, bill, card", "stage", "van", "ship", "fountain", "conveyer belt, conveyor belt, conveyer, conveyor, transporter", "canopy", "washer, automatic washer, washing machine", "plaything, toy", "pool", "stool", "barrel, cask", "basket, handbasket", "falls", "tent", "bag", "minibike, motorbike", "cradle", "oven", "ball", "food, solid food", "step, stair", "tank, storage tank", "trade name", "microwave", "pot", "animal", "bicycle", "lake", "dishwasher", "screen", "blanket, cover", "sculpture", "hood, exhaust hood", "sconce", "vase", "traffic light", "tray", "trash can", "fan", "pier", "crt screen", "plate", "monitor", "bulletin board", "shower", "radiator", "glass, drinking glass", "clock", "flag", # noqa
+]
+# After processed by `prepare_ade20k_sem_seg.py`, id 255 means ignore
+# fmt: on
+
+
+def _get_coco_instances_meta():
+    thing_ids = [k["id"] for k in COCO_CATEGORIES if k["isthing"] == 1]
+    thing_colors = [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 1]
+    assert len(thing_ids) == 80, len(thing_ids)
+    # Mapping from the incontiguous COCO category id to an id in [0, 79]
+    thing_dataset_id_to_contiguous_id = {k: i for i, k in enumerate(thing_ids)}
+    thing_classes = [k["name"] for k in COCO_CATEGORIES if k["isthing"] == 1]
+    ret = {
+        "thing_dataset_id_to_contiguous_id": thing_dataset_id_to_contiguous_id,
+        "thing_classes": thing_classes,
+        "thing_colors": thing_colors,
+    }
+    return ret
+
+
+def _get_coco_panoptic_separated_meta():
+    """
+    Returns metadata for "separated" version of the panoptic segmentation dataset.
+    """
+    stuff_ids = [k["id"] for k in COCO_CATEGORIES if k["isthing"] == 0]
+    assert len(stuff_ids) == 53, len(stuff_ids)
+
+    # For semantic segmentation, this mapping maps from contiguous stuff id
+    # (in [0, 53], used in models) to ids in the dataset (used for processing results)
+    # The id 0 is mapped to an extra category "thing".
+    stuff_dataset_id_to_contiguous_id = {k: i + 1 for i, k in enumerate(stuff_ids)}
+    # When converting COCO panoptic annotations to semantic annotations
+    # We label the "thing" category to 0
+    stuff_dataset_id_to_contiguous_id[0] = 0
+
+    # 54 names for COCO stuff categories (including "things")
+    stuff_classes = ["things"] + [
+        k["name"].replace("-other", "").replace("-merged", "")
+        for k in COCO_CATEGORIES
+        if k["isthing"] == 0
+    ]
+
+    # NOTE: I randomly picked a color for things
+    stuff_colors = [[82, 18, 128]] + [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 0]
+    ret = {
+        "stuff_dataset_id_to_contiguous_id": stuff_dataset_id_to_contiguous_id,
+        "stuff_classes": stuff_classes,
+        "stuff_colors": stuff_colors,
+    }
+    ret.update(_get_coco_instances_meta())
+    return ret
+
+
+def _get_builtin_metadata(dataset_name):
+    if dataset_name == "coco":
+        return _get_coco_instances_meta()
+    if dataset_name == "coco_panoptic_separated":
+        return _get_coco_panoptic_separated_meta()
+    elif dataset_name == "coco_panoptic_standard":
+        meta = {}
+        # The following metadata maps contiguous id from [0, #thing categories +
+        # #stuff categories) to their names and colors. We have to replica of the
+        # same name and color under "thing_*" and "stuff_*" because the current
+        # visualization function in D2 handles thing and class classes differently
+        # due to some heuristic used in Panoptic FPN. We keep the same naming to
+        # enable reusing existing visualization functions.
+        thing_classes = [k["name"] for k in COCO_CATEGORIES]
+        thing_colors = [k["color"] for k in COCO_CATEGORIES]
+        stuff_classes = [k["name"] for k in COCO_CATEGORIES]
+        stuff_colors = [k["color"] for k in COCO_CATEGORIES]
+
+        meta["thing_classes"] = thing_classes
+        meta["thing_colors"] = thing_colors
+        meta["stuff_classes"] = stuff_classes
+        meta["stuff_colors"] = stuff_colors
+
+        # Convert category id for training:
+        #   category id: like semantic segmentation, it is the class id for each
+        #   pixel. Since there are some classes not used in evaluation, the category
+        #   id is not always contiguous and thus we have two set of category ids:
+        #       - original category id: category id in the original dataset, mainly
+        #           used for evaluation.
+        #       - contiguous category id: [0, #classes), in order to train the linear
+        #           softmax classifier.
+        thing_dataset_id_to_contiguous_id = {}
+        stuff_dataset_id_to_contiguous_id = {}
+
+        for i, cat in enumerate(COCO_CATEGORIES):
+            if cat["isthing"]:
+                thing_dataset_id_to_contiguous_id[cat["id"]] = i
+            else:
+                stuff_dataset_id_to_contiguous_id[cat["id"]] = i
+
+        meta["thing_dataset_id_to_contiguous_id"] = thing_dataset_id_to_contiguous_id
+        meta["stuff_dataset_id_to_contiguous_id"] = stuff_dataset_id_to_contiguous_id
+
+        return meta
+    elif dataset_name == "coco_person":
+        return {
+            "thing_classes": ["person"],
+            "keypoint_names": COCO_PERSON_KEYPOINT_NAMES,
+            "keypoint_flip_map": COCO_PERSON_KEYPOINT_FLIP_MAP,
+            "keypoint_connection_rules": KEYPOINT_CONNECTION_RULES,
+        }
+    elif dataset_name == "cityscapes":
+        # fmt: off
+        CITYSCAPES_THING_CLASSES = [
+            "person", "rider", "car", "truck",
+            "bus", "train", "motorcycle", "bicycle",
+        ]
+        CITYSCAPES_STUFF_CLASSES = [
+            "road", "sidewalk", "building", "wall", "fence", "pole", "traffic light",
+            "traffic sign", "vegetation", "terrain", "sky", "person", "rider", "car",
+            "truck", "bus", "train", "motorcycle", "bicycle",
+        ]
+        # fmt: on
+        return {
+            "thing_classes": CITYSCAPES_THING_CLASSES,
+            "stuff_classes": CITYSCAPES_STUFF_CLASSES,
+        }
+    raise KeyError("No built-in metadata for dataset {}".format(dataset_name))