Merge branch 'add_v0.6' into 'v0.6-release'

support v0.6

See merge request !3

configs/PascalVOC-Detection/faster_rcnn_R_50_FPN.yaml

+_BASE_: "../Base-RCNN-FPN.yaml"
+MODEL:
+  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
+  MASK_ON: False
+  RESNETS:
+    DEPTH: 50
+  ROI_HEADS:
+    NUM_CLASSES: 20
+INPUT:
+  MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
+  MIN_SIZE_TEST: 800
+DATASETS:
+  TRAIN: ('voc_2007_trainval', 'voc_2012_trainval')
+  TEST: ('voc_2007_test',)
+SOLVER:
+  STEPS: (12000, 16000)
+  MAX_ITER: 18000  # 17.4 epochs
+  WARMUP_ITERS: 100

configs/common/README.md

+This directory provides definitions for a few common models, dataloaders, scheduler,
+and optimizers that are often used in training.
+The definition of these objects are provided in the form of lazy instantiation:
+their arguments can be edited by users before constructing the objects.
+
+They can be imported, or loaded by `model_zoo.get_config` API in users' own configs.

configs/common/coco_schedule.py

+from fvcore.common.param_scheduler import MultiStepParamScheduler
+
+from detectron2.config import LazyCall as L
+from detectron2.solver import WarmupParamScheduler
+
+
+def default_X_scheduler(num_X):
+    """
+    Returns the config for a default multi-step LR scheduler such as "1x", "3x",
+    commonly referred to in papers, where every 1x has the total length of 1440k
+    training images (~12 COCO epochs). LR is decayed twice at the end of training
+    following the strategy defined in "Rethinking ImageNet Pretraining", Sec 4.
+
+    Args:
+        num_X: a positive real number
+
+    Returns:
+        DictConfig: configs that define the multiplier for LR during training
+    """
+    # total number of iterations assuming 16 batch size, using 1440000/16=90000
+    total_steps_16bs = num_X * 90000
+
+    if num_X <= 2:
+        scheduler = L(MultiStepParamScheduler)(
+            values=[1.0, 0.1, 0.01],
+            # note that scheduler is scale-invariant. This is equivalent to
+            # milestones=[6, 8, 9]
+            milestones=[60000, 80000, 90000],
+        )
+    else:
+        scheduler = L(MultiStepParamScheduler)(
+            values=[1.0, 0.1, 0.01],
+            milestones=[total_steps_16bs - 60000, total_steps_16bs - 20000, total_steps_16bs],
+        )
+    return L(WarmupParamScheduler)(
+        scheduler=scheduler,
+        warmup_length=1000 / total_steps_16bs,
+        warmup_method="linear",
+        warmup_factor=0.001,
+    )
+
+
+lr_multiplier_1x = default_X_scheduler(1)
+lr_multiplier_2x = default_X_scheduler(2)
+lr_multiplier_3x = default_X_scheduler(3)
+lr_multiplier_6x = default_X_scheduler(6)
+lr_multiplier_9x = default_X_scheduler(9)

configs/common/data/coco.py

+from omegaconf import OmegaConf
+
+import detectron2.data.transforms as T
+from detectron2.config import LazyCall as L
+from detectron2.data import (
+    DatasetMapper,
+    build_detection_test_loader,
+    build_detection_train_loader,
+    get_detection_dataset_dicts,
+)
+from detectron2.evaluation import COCOEvaluator
+
+dataloader = OmegaConf.create()
+
+dataloader.train = L(build_detection_train_loader)(
+    dataset=L(get_detection_dataset_dicts)(names="coco_2017_train"),
+    mapper=L(DatasetMapper)(
+        is_train=True,
+        augmentations=[
+            L(T.ResizeShortestEdge)(
+                short_edge_length=(640, 672, 704, 736, 768, 800),
+                sample_style="choice",
+                max_size=1333,
+            ),
+            L(T.RandomFlip)(horizontal=True),
+        ],
+        image_format="BGR",
+        use_instance_mask=True,
+    ),
+    total_batch_size=16,
+    num_workers=4,
+)
+
+dataloader.test = L(build_detection_test_loader)(
+    dataset=L(get_detection_dataset_dicts)(names="coco_2017_val", filter_empty=False),
+    mapper=L(DatasetMapper)(
+        is_train=False,
+        augmentations=[
+            L(T.ResizeShortestEdge)(short_edge_length=800, max_size=1333),
+        ],
+        image_format="${...train.mapper.image_format}",
+    ),
+    num_workers=4,
+)
+
+dataloader.evaluator = L(COCOEvaluator)(
+    dataset_name="${..test.dataset.names}",
+)

configs/common/data/coco_keypoint.py

+from detectron2.data.detection_utils import create_keypoint_hflip_indices
+
+from .coco import dataloader
+
+dataloader.train.dataset.min_keypoints = 1
+dataloader.train.dataset.names = "keypoints_coco_2017_train"
+dataloader.test.dataset.names = "keypoints_coco_2017_val"
+
+dataloader.train.mapper.update(
+    use_instance_mask=False,
+    use_keypoint=True,
+    keypoint_hflip_indices=create_keypoint_hflip_indices(dataloader.train.dataset.names),
+)

configs/common/data/coco_panoptic_separated.py

+from detectron2.config import LazyCall as L
+from detectron2.evaluation import (
+    COCOEvaluator,
+    COCOPanopticEvaluator,
+    DatasetEvaluators,
+    SemSegEvaluator,
+)
+
+from .coco import dataloader
+
+dataloader.train.dataset.names = "coco_2017_train_panoptic_separated"
+dataloader.train.dataset.filter_empty = False
+dataloader.test.dataset.names = "coco_2017_val_panoptic_separated"
+
+
+dataloader.evaluator = [
+    L(COCOEvaluator)(
+        dataset_name="${...test.dataset.names}",
+    ),
+    L(SemSegEvaluator)(
+        dataset_name="${...test.dataset.names}",
+    ),
+    L(COCOPanopticEvaluator)(
+        dataset_name="${...test.dataset.names}",
+    ),
+]

configs/common/models/cascade_rcnn.py

+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.matcher import Matcher
+from detectron2.modeling.roi_heads import FastRCNNOutputLayers, FastRCNNConvFCHead, CascadeROIHeads
+
+from .mask_rcnn_fpn import model
+
+# arguments that don't exist for Cascade R-CNN
+[model.roi_heads.pop(k) for k in ["box_head", "box_predictor", "proposal_matcher"]]
+
+model.roi_heads.update(
+    _target_=CascadeROIHeads,
+    box_heads=[
+        L(FastRCNNConvFCHead)(
+            input_shape=ShapeSpec(channels=256, height=7, width=7),
+            conv_dims=[],
+            fc_dims=[1024, 1024],
+        )
+        for k in range(3)
+    ],
+    box_predictors=[
+        L(FastRCNNOutputLayers)(
+            input_shape=ShapeSpec(channels=1024),
+            test_score_thresh=0.05,
+            box2box_transform=L(Box2BoxTransform)(weights=(w1, w1, w2, w2)),
+            cls_agnostic_bbox_reg=True,
+            num_classes="${...num_classes}",
+        )
+        for (w1, w2) in [(10, 5), (20, 10), (30, 15)]
+    ],
+    proposal_matchers=[
+        L(Matcher)(thresholds=[th], labels=[0, 1], allow_low_quality_matches=False)
+        for th in [0.5, 0.6, 0.7]
+    ],
+)

configs/common/models/fcos.py

+from detectron2.modeling.meta_arch.fcos import FCOS, FCOSHead
+
+from .retinanet import model
+
+model._target_ = FCOS
+
+del model.anchor_generator
+del model.box2box_transform
+del model.anchor_matcher
+del model.input_format
+
+# Use P5 instead of C5 to compute P6/P7
+# (Sec 2.2 of https://arxiv.org/abs/2006.09214)
+model.backbone.top_block.in_feature = "p5"
+model.backbone.top_block.in_channels = 256
+
+# New score threshold determined based on sqrt(cls_score * centerness)
+model.test_score_thresh = 0.2
+model.test_nms_thresh = 0.6
+
+model.head._target_ = FCOSHead
+del model.head.num_anchors
+model.head.norm = "GN"

configs/common/models/keypoint_rcnn_fpn.py

+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling.poolers import ROIPooler
+from detectron2.modeling.roi_heads import KRCNNConvDeconvUpsampleHead
+
+from .mask_rcnn_fpn import model
+
+[model.roi_heads.pop(x) for x in ["mask_in_features", "mask_pooler", "mask_head"]]
+
+model.roi_heads.update(
+    num_classes=1,
+    keypoint_in_features=["p2", "p3", "p4", "p5"],
+    keypoint_pooler=L(ROIPooler)(
+        output_size=14,
+        scales=(1.0 / 4, 1.0 / 8, 1.0 / 16, 1.0 / 32),
+        sampling_ratio=0,
+        pooler_type="ROIAlignV2",
+    ),
+    keypoint_head=L(KRCNNConvDeconvUpsampleHead)(
+        input_shape=ShapeSpec(channels=256, width=14, height=14),
+        num_keypoints=17,
+        conv_dims=[512] * 8,
+        loss_normalizer="visible",
+    ),
+)
+
+# Detectron1 uses 2000 proposals per-batch, but this option is per-image in detectron2.
+# 1000 proposals per-image is found to hurt box AP.
+# Therefore we increase it to 1500 per-image.
+model.proposal_generator.post_nms_topk = (1500, 1000)
+
+# Keypoint AP degrades (though box AP improves) when using plain L1 loss
+model.roi_heads.box_predictor.smooth_l1_beta = 0.5

configs/common/models/mask_rcnn_c4.py

+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling.meta_arch import GeneralizedRCNN
+from detectron2.modeling.anchor_generator import DefaultAnchorGenerator
+from detectron2.modeling.backbone import BasicStem, BottleneckBlock, ResNet
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.matcher import Matcher
+from detectron2.modeling.poolers import ROIPooler
+from detectron2.modeling.proposal_generator import RPN, StandardRPNHead
+from detectron2.modeling.roi_heads import (
+    FastRCNNOutputLayers,
+    MaskRCNNConvUpsampleHead,
+    Res5ROIHeads,
+)
+
+model = L(GeneralizedRCNN)(
+    backbone=L(ResNet)(
+        stem=L(BasicStem)(in_channels=3, out_channels=64, norm="FrozenBN"),
+        stages=L(ResNet.make_default_stages)(
+            depth=50,
+            stride_in_1x1=True,
+            norm="FrozenBN",
+        ),
+        out_features=["res4"],
+    ),
+    proposal_generator=L(RPN)(
+        in_features=["res4"],
+        head=L(StandardRPNHead)(in_channels=1024, num_anchors=15),
+        anchor_generator=L(DefaultAnchorGenerator)(
+            sizes=[[32, 64, 128, 256, 512]],
+            aspect_ratios=[0.5, 1.0, 2.0],
+            strides=[16],
+            offset=0.0,
+        ),
+        anchor_matcher=L(Matcher)(
+            thresholds=[0.3, 0.7], labels=[0, -1, 1], allow_low_quality_matches=True
+        ),
+        box2box_transform=L(Box2BoxTransform)(weights=[1.0, 1.0, 1.0, 1.0]),
+        batch_size_per_image=256,
+        positive_fraction=0.5,
+        pre_nms_topk=(12000, 6000),
+        post_nms_topk=(2000, 1000),
+        nms_thresh=0.7,
+    ),
+    roi_heads=L(Res5ROIHeads)(
+        num_classes=80,
+        batch_size_per_image=512,
+        positive_fraction=0.25,
+        proposal_matcher=L(Matcher)(
+            thresholds=[0.5], labels=[0, 1], allow_low_quality_matches=False
+        ),
+        in_features=["res4"],
+        pooler=L(ROIPooler)(
+            output_size=14,
+            scales=(1.0 / 16,),
+            sampling_ratio=0,
+            pooler_type="ROIAlignV2",
+        ),
+        res5=L(ResNet.make_stage)(
+            block_class=BottleneckBlock,
+            num_blocks=3,
+            stride_per_block=[2, 1, 1],
+            in_channels=1024,
+            bottleneck_channels=512,
+            out_channels=2048,
+            norm="FrozenBN",
+            stride_in_1x1=True,
+        ),
+        box_predictor=L(FastRCNNOutputLayers)(
+            input_shape=L(ShapeSpec)(channels="${...res5.out_channels}", height=1, width=1),
+            test_score_thresh=0.05,
+            box2box_transform=L(Box2BoxTransform)(weights=(10, 10, 5, 5)),
+            num_classes="${..num_classes}",
+        ),
+        mask_head=L(MaskRCNNConvUpsampleHead)(
+            input_shape=L(ShapeSpec)(
+                channels="${...res5.out_channels}",
+                width="${...pooler.output_size}",
+                height="${...pooler.output_size}",
+            ),
+            num_classes="${..num_classes}",
+            conv_dims=[256],
+        ),
+    ),
+    pixel_mean=[103.530, 116.280, 123.675],
+    pixel_std=[1.0, 1.0, 1.0],
+    input_format="BGR",
+)

configs/common/models/mask_rcnn_fpn.py

+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling.meta_arch import GeneralizedRCNN
+from detectron2.modeling.anchor_generator import DefaultAnchorGenerator
+from detectron2.modeling.backbone.fpn import LastLevelMaxPool
+from detectron2.modeling.backbone import BasicStem, FPN, ResNet
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.matcher import Matcher
+from detectron2.modeling.poolers import ROIPooler
+from detectron2.modeling.proposal_generator import RPN, StandardRPNHead
+from detectron2.modeling.roi_heads import (
+    StandardROIHeads,
+    FastRCNNOutputLayers,
+    MaskRCNNConvUpsampleHead,
+    FastRCNNConvFCHead,
+)
+
+model = L(GeneralizedRCNN)(
+    backbone=L(FPN)(
+        bottom_up=L(ResNet)(
+            stem=L(BasicStem)(in_channels=3, out_channels=64, norm="FrozenBN"),
+            stages=L(ResNet.make_default_stages)(
+                depth=50,
+                stride_in_1x1=True,
+                norm="FrozenBN",
+            ),
+            out_features=["res2", "res3", "res4", "res5"],
+        ),
+        in_features="${.bottom_up.out_features}",
+        out_channels=256,
+        top_block=L(LastLevelMaxPool)(),
+    ),
+    proposal_generator=L(RPN)(
+        in_features=["p2", "p3", "p4", "p5", "p6"],
+        head=L(StandardRPNHead)(in_channels=256, num_anchors=3),
+        anchor_generator=L(DefaultAnchorGenerator)(
+            sizes=[[32], [64], [128], [256], [512]],
+            aspect_ratios=[0.5, 1.0, 2.0],
+            strides=[4, 8, 16, 32, 64],
+            offset=0.0,
+        ),
+        anchor_matcher=L(Matcher)(
+            thresholds=[0.3, 0.7], labels=[0, -1, 1], allow_low_quality_matches=True
+        ),
+        box2box_transform=L(Box2BoxTransform)(weights=[1.0, 1.0, 1.0, 1.0]),
+        batch_size_per_image=256,
+        positive_fraction=0.5,
+        pre_nms_topk=(2000, 1000),
+        post_nms_topk=(1000, 1000),
+        nms_thresh=0.7,
+    ),
+    roi_heads=L(StandardROIHeads)(
+        num_classes=80,
+        batch_size_per_image=512,
+        positive_fraction=0.25,
+        proposal_matcher=L(Matcher)(
+            thresholds=[0.5], labels=[0, 1], allow_low_quality_matches=False
+        ),
+        box_in_features=["p2", "p3", "p4", "p5"],
+        box_pooler=L(ROIPooler)(
+            output_size=7,
+            scales=(1.0 / 4, 1.0 / 8, 1.0 / 16, 1.0 / 32),
+            sampling_ratio=0,
+            pooler_type="ROIAlignV2",
+        ),
+        box_head=L(FastRCNNConvFCHead)(
+            input_shape=ShapeSpec(channels=256, height=7, width=7),
+            conv_dims=[],
+            fc_dims=[1024, 1024],
+        ),
+        box_predictor=L(FastRCNNOutputLayers)(
+            input_shape=ShapeSpec(channels=1024),
+            test_score_thresh=0.05,
+            box2box_transform=L(Box2BoxTransform)(weights=(10, 10, 5, 5)),
+            num_classes="${..num_classes}",
+        ),
+        mask_in_features=["p2", "p3", "p4", "p5"],
+        mask_pooler=L(ROIPooler)(
+            output_size=14,
+            scales=(1.0 / 4, 1.0 / 8, 1.0 / 16, 1.0 / 32),
+            sampling_ratio=0,
+            pooler_type="ROIAlignV2",
+        ),
+        mask_head=L(MaskRCNNConvUpsampleHead)(
+            input_shape=ShapeSpec(channels=256, width=14, height=14),
+            num_classes="${..num_classes}",
+            conv_dims=[256, 256, 256, 256, 256],
+        ),
+    ),
+    pixel_mean=[103.530, 116.280, 123.675],
+    pixel_std=[1.0, 1.0, 1.0],
+    input_format="BGR",
+)

configs/common/models/panoptic_fpn.py

+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling import PanopticFPN
+from detectron2.modeling.meta_arch.semantic_seg import SemSegFPNHead
+
+from .mask_rcnn_fpn import model
+
+model._target_ = PanopticFPN
+model.sem_seg_head = L(SemSegFPNHead)(
+    input_shape={
+        f: L(ShapeSpec)(stride=s, channels="${....backbone.out_channels}")
+        for f, s in zip(["p2", "p3", "p4", "p5"], [4, 8, 16, 32])
+    },
+    ignore_value=255,
+    num_classes=54,  # COCO stuff + 1
+    conv_dims=128,
+    common_stride=4,
+    loss_weight=0.5,
+    norm="GN",
+)

configs/common/models/retinanet.py

+# -*- coding: utf-8 -*-
+
+from detectron2.config import LazyCall as L
+from detectron2.layers import ShapeSpec
+from detectron2.modeling.meta_arch import RetinaNet
+from detectron2.modeling.anchor_generator import DefaultAnchorGenerator
+from detectron2.modeling.backbone.fpn import LastLevelP6P7
+from detectron2.modeling.backbone import BasicStem, FPN, ResNet
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.matcher import Matcher
+from detectron2.modeling.meta_arch.retinanet import RetinaNetHead
+
+model = L(RetinaNet)(
+    backbone=L(FPN)(
+        bottom_up=L(ResNet)(
+            stem=L(BasicStem)(in_channels=3, out_channels=64, norm="FrozenBN"),
+            stages=L(ResNet.make_default_stages)(
+                depth=50,
+                stride_in_1x1=True,
+                norm="FrozenBN",
+            ),
+            out_features=["res3", "res4", "res5"],
+        ),
+        in_features=["res3", "res4", "res5"],
+        out_channels=256,
+        top_block=L(LastLevelP6P7)(in_channels=2048, out_channels="${..out_channels}"),
+    ),
+    head=L(RetinaNetHead)(
+        # Shape for each input feature map
+        input_shape=[ShapeSpec(channels=256)] * 5,
+        num_classes="${..num_classes}",
+        conv_dims=[256, 256, 256, 256],
+        prior_prob=0.01,
+        num_anchors=9,
+    ),
+    anchor_generator=L(DefaultAnchorGenerator)(
+        sizes=[[x, x * 2 ** (1.0 / 3), x * 2 ** (2.0 / 3)] for x in [32, 64, 128, 256, 512]],
+        aspect_ratios=[0.5, 1.0, 2.0],
+        strides=[8, 16, 32, 64, 128],
+        offset=0.0,
+    ),
+    box2box_transform=L(Box2BoxTransform)(weights=[1.0, 1.0, 1.0, 1.0]),
+    anchor_matcher=L(Matcher)(
+        thresholds=[0.4, 0.5], labels=[0, -1, 1], allow_low_quality_matches=True
+    ),
+    num_classes=80,
+    head_in_features=["p3", "p4", "p5", "p6", "p7"],
+    focal_loss_alpha=0.25,
+    focal_loss_gamma=2.0,
+    pixel_mean=[103.530, 116.280, 123.675],
+    pixel_std=[1.0, 1.0, 1.0],
+    input_format="BGR",
+)

configs/common/optim.py

+import torch
+
+from detectron2.config import LazyCall as L
+from detectron2.solver.build import get_default_optimizer_params
+
+SGD = L(torch.optim.SGD)(
+    params=L(get_default_optimizer_params)(
+        # params.model is meant to be set to the model object, before instantiating
+        # the optimizer.
+        weight_decay_norm=0.0
+    ),
+    lr=0.02,
+    momentum=0.9,
+    weight_decay=1e-4,
+)

configs/common/train.py

+# Common training-related configs that are designed for "tools/lazyconfig_train_net.py"
+# You can use your own instead, together with your own train_net.py
+train = dict(
+    output_dir="./output",
+    init_checkpoint="",
+    max_iter=90000,
+    amp=dict(enabled=False),  # options for Automatic Mixed Precision
+    ddp=dict(  # options for DistributedDataParallel
+        broadcast_buffers=False,
+        find_unused_parameters=False,
+        fp16_compression=False,
+    ),
+    checkpointer=dict(period=5000, max_to_keep=100),  # options for PeriodicCheckpointer
+    eval_period=5000,
+    log_period=20,
+    device="cuda"
+    # ...
+)

configs/new_baselines/mask_rcnn_R_101_FPN_100ep_LSJ.py

+from .mask_rcnn_R_50_FPN_100ep_LSJ import (
+    dataloader,
+    lr_multiplier,
+    model,
+    optimizer,
+    train,
+)
+
+model.backbone.bottom_up.stages.depth = 101

configs/new_baselines/mask_rcnn_R_101_FPN_200ep_LSJ.py

+from .mask_rcnn_R_101_FPN_100ep_LSJ import (
+    dataloader,
+    lr_multiplier,
+    model,
+    optimizer,
+    train,
+)
+
+train.max_iter *= 2  # 100ep -> 200ep
+
+lr_multiplier.scheduler.milestones = [
+    milestone * 2 for milestone in lr_multiplier.scheduler.milestones
+]
+lr_multiplier.scheduler.num_updates = train.max_iter

configs/new_baselines/mask_rcnn_R_101_FPN_400ep_LSJ.py

+from .mask_rcnn_R_101_FPN_100ep_LSJ import (
+    dataloader,
+    lr_multiplier,
+    model,
+    optimizer,
+    train,
+)
+
+train.max_iter *= 4  # 100ep -> 400ep
+
+lr_multiplier.scheduler.milestones = [
+    milestone * 4 for milestone in lr_multiplier.scheduler.milestones
+]
+lr_multiplier.scheduler.num_updates = train.max_iter

configs/new_baselines/mask_rcnn_R_50_FPN_100ep_LSJ.py

+import detectron2.data.transforms as T
+from detectron2.config.lazy import LazyCall as L
+from detectron2.layers.batch_norm import NaiveSyncBatchNorm
+from detectron2.solver import WarmupParamScheduler
+from fvcore.common.param_scheduler import MultiStepParamScheduler
+
+from ..common.data.coco import dataloader
+from ..common.models.mask_rcnn_fpn import model
+from ..common.optim import SGD as optimizer
+from ..common.train import train
+
+# train from scratch
+train.init_checkpoint = ""
+train.amp.enabled = True
+train.ddp.fp16_compression = True
+model.backbone.bottom_up.freeze_at = 0
+
+# SyncBN
+# fmt: off
+model.backbone.bottom_up.stem.norm = \
+    model.backbone.bottom_up.stages.norm = \
+    model.backbone.norm = "SyncBN"
+
+# Using NaiveSyncBatchNorm becase heads may have empty input. That is not supported by
+# torch.nn.SyncBatchNorm. We can remove this after
+# https://github.com/pytorch/pytorch/issues/36530 is fixed.
+model.roi_heads.box_head.conv_norm = \
+    model.roi_heads.mask_head.conv_norm = lambda c: NaiveSyncBatchNorm(c,
+                                                                       stats_mode="N")
+# fmt: on
+
+# 2conv in RPN:
+# https://github.com/tensorflow/tpu/blob/b24729de804fdb751b06467d3dce0637fa652060/models/official/detection/modeling/architecture/heads.py#L95-L97  # noqa: E501, B950
+model.proposal_generator.head.conv_dims = [-1, -1]
+
+# 4conv1fc box head
+model.roi_heads.box_head.conv_dims = [256, 256, 256, 256]
+model.roi_heads.box_head.fc_dims = [1024]
+
+# resize_and_crop_image in:
+# https://github.com/tensorflow/tpu/blob/b24729de804fdb751b06467d3dce0637fa652060/models/official/detection/utils/input_utils.py#L127  # noqa: E501, B950
+image_size = 1024
+dataloader.train.mapper.augmentations = [
+    L(T.ResizeScale)(
+        min_scale=0.1, max_scale=2.0, target_height=image_size, target_width=image_size
+    ),
+    L(T.FixedSizeCrop)(crop_size=(image_size, image_size)),
+    L(T.RandomFlip)(horizontal=True),
+]
+
+# recompute boxes due to cropping
+dataloader.train.mapper.recompute_boxes = True
+
+# larger batch-size.
+dataloader.train.total_batch_size = 64
+
+# Equivalent to 100 epochs.
+# 100 ep = 184375 iters * 64 images/iter / 118000 images/ep
+train.max_iter = 184375
+
+lr_multiplier = L(WarmupParamScheduler)(
+    scheduler=L(MultiStepParamScheduler)(
+        values=[1.0, 0.1, 0.01],
+        milestones=[163889, 177546],
+        num_updates=train.max_iter,
+    ),
+    warmup_length=500 / train.max_iter,
+    warmup_factor=0.067,
+)
+
+optimizer.lr = 0.1
+optimizer.weight_decay = 4e-5

configs/new_baselines/mask_rcnn_R_50_FPN_200ep_LSJ.py

+from .mask_rcnn_R_50_FPN_100ep_LSJ import (
+    dataloader,
+    lr_multiplier,
+    model,
+    optimizer,
+    train,
+)
+
+train.max_iter *= 2  # 100ep -> 200ep
+
+lr_multiplier.scheduler.milestones = [
+    milestone * 2 for milestone in lr_multiplier.scheduler.milestones
+]
+lr_multiplier.scheduler.num_updates = train.max_iter