tuofeilunhifi
diff --git a/‎configs/yolox/README.md
+2-2 b/‎configs/yolox/README.md
+2-2
diff --git a/‎configs/yolox/metafile.yml
+6-5 b/‎configs/yolox/metafile.yml
+6-5
diff --git a/‎configs/yolox/yolox_s_8x8_300e_coco.py
+49-31 b/‎configs/yolox/yolox_s_8x8_300e_coco.py
+49-31
diff --git a/‎configs/yolox/yolox_tiny_8x8_300e_coco.py
+12-37 b/‎configs/yolox/yolox_tiny_8x8_300e_coco.py
+12-37
diff --git a/‎mmdet/core/bbox/assigners/sim_ota_assigner.py
+2-1 b/‎mmdet/core/bbox/assigners/sim_ota_assigner.py
+2-1
diff --git a/‎mmdet/core/evaluation/eval_hooks.py
+65 b/‎mmdet/core/evaluation/eval_hooks.py
+65
diff --git a/‎mmdet/core/hook/yolox_mode_switch_hook.py
+14 b/‎mmdet/core/hook/yolox_mode_switch_hook.py
+14
@@ -17,9 +17,9 @@
 
 | Backbone  | size   | Mem (GB) |   box AP | Config | Download |
 |:---------:|:-------:|:-------:|:-------:|:--------:|:------:|
-| YOLOX-Tiny | 416 |   3.6      |   31.6  | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/yolox/yolox_tiny_8x8_300e_coco.py)       |[model](https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_tiny_8x8_300e_coco/yolox_tiny_8x8_300e_coco_20210806_234250-4ff3b67e.pth) &#124; [log](https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_tiny_8x8_300e_coco/yolox_tiny_8x8_300e_coco_20210806_234250.log.json) |
+| YOLOX-s | 640 |   7.6      |   40.5  | [config](https://github.com/open-mmlab/mmdetection/tree/master/configs/yolox/yolox_s_8x8_300e_coco.py)       |[model](https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_s_8x8_300e_coco/yolox_s_8x8_300e_coco_20211121_095711-4592a793.pth) &#124; [log](https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_s_8x8_300e_coco/yolox_s_8x8_300e_coco_20211121_095711.log.json) |
+
 
 **Note**:
 
 1. The test score threshold is 0.001.
-2. We find that the performance is unstable and may fluctuate by about 0.7 mAP. We will continue to investigate and improve it.
@@ -18,16 +18,17 @@ Collections:
       URL: https://github.com/open-mmlab/mmdetection/blob/v2.15.1/mmdet/models/detectors/yolox.py#L6
       Version: v2.15.1
 
+
 Models:
-  - Name: yolox_tiny_8x8_300e_coco
+  - Name: yolox_s_8x8_300e_coco
     In Collection: YOLOX
-    Config: configs/yolox/yolox_tiny_8x8_300e_coco.py
+    Config: configs/yolox/yolox_s_8x8_300e_coco.py
     Metadata:
-      Training Memory (GB): 3.6
+      Training Memory (GB): 7.6
       Epochs: 300
     Results:
       - Task: Object Detection
         Dataset: COCO
         Metrics:
-          box AP: 31.6
-    Weights: https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_tiny_8x8_300e_coco/yolox_tiny_8x8_300e_coco_20210806_234250-4ff3b67e.pth
+          box AP: 40.5
+    Weights: https://download.openmmlab.com/mmdetection/v2.0/yolox/yolox_s_8x8_300e_coco/yolox_s_8x8_300e_coco_20211121_095711-4592a793.pth
@@ -1,8 +1,13 @@
 _base_ = ['../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py']
 
+img_scale = (640, 640)
+
 # model settings
 model = dict(
     type='YOLOX',
+    input_size=img_scale,
+    random_size_range=(15, 25),
+    random_size_interval=10,
     backbone=dict(type='CSPDarknet', deepen_factor=0.33, widen_factor=0.5),
     neck=dict(
         type='YOLOXPAFPN',
@@ -20,11 +25,6 @@
 data_root = 'data/coco/'
 dataset_type = 'CocoDataset'
 
-img_norm_cfg = dict(
-    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
-
-img_scale = (640, 640)
-
 train_pipeline = [
     dict(type='Mosaic', img_scale=img_scale, pad_val=114.0),
     dict(
@@ -36,16 +36,19 @@
         img_scale=img_scale,
         ratio_range=(0.8, 1.6),
         pad_val=114.0),
-    dict(
-        type='PhotoMetricDistortion',
-        brightness_delta=32,
-        contrast_range=(0.5, 1.5),
-        saturation_range=(0.5, 1.5),
-        hue_delta=18),
+    dict(type='YOLOXHSVRandomAug'),
     dict(type='RandomFlip', flip_ratio=0.5),
-    dict(type='Resize', keep_ratio=True),
-    dict(type='Pad', pad_to_square=True, pad_val=114.0),
-    dict(type='Normalize', **img_norm_cfg),
+    # According to the official implementation, multi-scale
+    # training is not considered here but in the
+    # 'mmdet/models/detectors/yolox.py'.
+    dict(type='Resize', img_scale=img_scale, keep_ratio=True),
+    dict(
+        type='Pad',
+        pad_to_square=True,
+        # If the image is three-channel, the pad value needs
+        # to be set separately for each channel.
+        pad_val=dict(img=(114.0, 114.0, 114.0))),
+    dict(type='FilterAnnotations', min_gt_bbox_wh=(1, 1), keep_empty=False),
     dict(type='DefaultFormatBundle'),
     dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
 ]
@@ -57,13 +60,12 @@
         ann_file=data_root + 'annotations/instances_train2017.json',
         img_prefix=data_root + 'train2017/',
         pipeline=[
-            dict(type='LoadImageFromFile', to_float32=True),
+            dict(type='LoadImageFromFile'),
             dict(type='LoadAnnotations', with_bbox=True)
         ],
         filter_empty_gt=False,
     ),
-    pipeline=train_pipeline,
-    dynamic_scale=img_scale)
+    pipeline=train_pipeline)
 
 test_pipeline = [
     dict(type='LoadImageFromFile'),
@@ -74,16 +76,19 @@
         transforms=[
             dict(type='Resize', keep_ratio=True),
             dict(type='RandomFlip'),
-            dict(type='Pad', size=img_scale, pad_val=114.0),
-            dict(type='Normalize', **img_norm_cfg),
+            dict(
+                type='Pad',
+                pad_to_square=True,
+                pad_val=dict(img=(114.0, 114.0, 114.0))),
             dict(type='DefaultFormatBundle'),
             dict(type='Collect', keys=['img'])
         ])
 ]
 
 data = dict(
     samples_per_gpu=8,
-    workers_per_gpu=2,
+    workers_per_gpu=4,
+    persistent_workers=True,
     train=train_dataset,
     val=dict(
         type=dataset_type,
@@ -107,6 +112,11 @@
     paramwise_cfg=dict(norm_decay_mult=0., bias_decay_mult=0.))
 optimizer_config = dict(grad_clip=None)
 
+max_epochs = 300
+num_last_epochs = 15
+resume_from = None
+interval = 10
+
 # learning policy
 lr_config = dict(
     _delete_=True,
@@ -116,27 +126,35 @@
     warmup_by_epoch=True,
     warmup_ratio=1,
     warmup_iters=5,  # 5 epoch
-    num_last_epochs=15,
+    num_last_epochs=num_last_epochs,
     min_lr_ratio=0.05)
-runner = dict(type='EpochBasedRunner', max_epochs=300)
 
-resume_from = None
-interval = 10
+runner = dict(type='EpochBasedRunner', max_epochs=max_epochs)
 
 custom_hooks = [
-    dict(type='YOLOXModeSwitchHook', num_last_epochs=15, priority=48),
     dict(
-        type='SyncRandomSizeHook',
-        ratio_range=(14, 26),
-        img_scale=img_scale,
+        type='YOLOXModeSwitchHook',
+        num_last_epochs=num_last_epochs,
         priority=48),
     dict(
         type='SyncNormHook',
-        num_last_epochs=15,
+        num_last_epochs=num_last_epochs,
         interval=interval,
         priority=48),
-    dict(type='ExpMomentumEMAHook', resume_from=resume_from, priority=49)
+    dict(
+        type='ExpMomentumEMAHook',
+        resume_from=resume_from,
+        momentum=0.0001,
+        priority=49)
 ]
 checkpoint_config = dict(interval=interval)
-evaluation = dict(interval=interval, metric='bbox')
+evaluation = dict(
+    save_best='auto',
+    # The evaluation interval is 'interval' when running epoch is
+    # less than ‘max_epochs - num_last_epochs’.
+    # The evaluation interval is 1 when running epoch is greater than
+    # or equal to ‘max_epochs - num_last_epochs’.
+    interval=interval,
+    dynamic_intervals=[(max_epochs - num_last_epochs, 1)],
+    metric='bbox')
 log_config = dict(interval=50)
@@ -2,14 +2,11 @@
 
 # model settings
 model = dict(
+    random_size_range=(10, 20),
     backbone=dict(deepen_factor=0.33, widen_factor=0.375),
     neck=dict(in_channels=[96, 192, 384], out_channels=96),
     bbox_head=dict(in_channels=96, feat_channels=96))
 
-# dataset settings
-img_norm_cfg = dict(
-    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
-
 img_scale = (640, 640)
 
 train_pipeline = [
@@ -18,16 +15,14 @@
         type='RandomAffine',
         scaling_ratio_range=(0.5, 1.5),
         border=(-img_scale[0] // 2, -img_scale[1] // 2)),
-    dict(
-        type='PhotoMetricDistortion',
-        brightness_delta=32,
-        contrast_range=(0.5, 1.5),
-        saturation_range=(0.5, 1.5),
-        hue_delta=18),
+    dict(type='YOLOXHSVRandomAug'),
     dict(type='RandomFlip', flip_ratio=0.5),
-    dict(type='Resize', keep_ratio=True),
-    dict(type='Pad', pad_to_square=True, pad_val=114.0),
-    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Resize', img_scale=img_scale, keep_ratio=True),
+    dict(
+        type='Pad',
+        pad_to_square=True,
+        pad_val=dict(img=(114.0, 114.0, 114.0))),
+    dict(type='FilterAnnotations', min_gt_bbox_wh=(1, 1), keep_empty=False),
     dict(type='DefaultFormatBundle'),
     dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
 ]
@@ -41,8 +36,10 @@
         transforms=[
             dict(type='Resize', keep_ratio=True),
             dict(type='RandomFlip'),
-            dict(type='Pad', size=(416, 416), pad_val=114.0),
-            dict(type='Normalize', **img_norm_cfg),
+            dict(
+                type='Pad',
+                pad_to_square=True,
+                pad_val=dict(img=(114.0, 114.0, 114.0))),
             dict(type='DefaultFormatBundle'),
             dict(type='Collect', keys=['img'])
         ])
@@ -54,25 +51,3 @@
     train=train_dataset,
     val=dict(pipeline=test_pipeline),
     test=dict(pipeline=test_pipeline))
-
-resume_from = None
-interval = 10
-
-# Execute in the order of insertion when the priority is the same.
-# The smaller the value, the higher the priority
-custom_hooks = [
-    dict(type='YOLOXModeSwitchHook', num_last_epochs=15, priority=48),
-    dict(
-        type='SyncRandomSizeHook',
-        ratio_range=(10, 20),
-        img_scale=img_scale,
-        priority=48),
-    dict(
-        type='SyncNormHook',
-        num_last_epochs=15,
-        interval=interval,
-        priority=48),
-    dict(type='ExpMomentumEMAHook', resume_from=resume_from, priority=49)
-]
-checkpoint_config = dict(interval=interval)
-evaluation = dict(interval=interval, metric='bbox')
@@ -228,7 +228,8 @@ def get_in_gt_and_in_center_info(self, priors, gt_bboxes):
     def dynamic_k_matching(self, cost, pairwise_ious, num_gt, valid_mask):
         matching_matrix = torch.zeros_like(cost)
         # select candidate topk ious for dynamic-k calculation
-        topk_ious, _ = torch.topk(pairwise_ious, self.candidate_topk, dim=0)
+        candidate_topk = min(self.candidate_topk, pairwise_ious.size(0))
+        topk_ious, _ = torch.topk(pairwise_ious, candidate_topk, dim=0)
         # calculate dynamic k for each gt
         dynamic_ks = torch.clamp(topk_ious.sum(0).int(), min=1)
         for gt_idx in range(num_gt):
 
@@ -1,14 +1,52 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+import bisect
 import os.path as osp
 
+import mmcv
 import torch.distributed as dist
 from mmcv.runner import DistEvalHook as BaseDistEvalHook
 from mmcv.runner import EvalHook as BaseEvalHook
 from torch.nn.modules.batchnorm import _BatchNorm
 
 
+def _calc_dynamic_intervals(start_interval, dynamic_interval_list):
+    assert mmcv.is_list_of(dynamic_interval_list, tuple)
+
+    dynamic_milestones = [0]
+    dynamic_milestones.extend(
+        [dynamic_interval[0] for dynamic_interval in dynamic_interval_list])
+    dynamic_intervals = [start_interval]
+    dynamic_intervals.extend(
+        [dynamic_interval[1] for dynamic_interval in dynamic_interval_list])
+    return dynamic_milestones, dynamic_intervals
+
+
 class EvalHook(BaseEvalHook):
 
+    def __init__(self, *args, dynamic_intervals=None, **kwargs):
+        super(EvalHook, self).__init__(*args, **kwargs)
+
+        self.use_dynamic_intervals = dynamic_intervals is not None
+        if self.use_dynamic_intervals:
+            self.dynamic_milestones, self.dynamic_intervals = \
+                _calc_dynamic_intervals(self.interval, dynamic_intervals)
+
+    def _decide_interval(self, runner):
+        if self.use_dynamic_intervals:
+            progress = runner.epoch if self.by_epoch else runner.iter
+            step = bisect.bisect(self.dynamic_milestones, (progress + 1))
+            # Dynamically modify the evaluation interval
+            self.interval = self.dynamic_intervals[step - 1]
+
+    def before_train_epoch(self, runner):
+        """Evaluate the model only at the start of training by epoch."""
+        self._decide_interval(runner)
+        super().before_train_epoch(runner)
+
+    def before_train_iter(self, runner):
+        self._decide_interval(runner)
+        super().before_train_iter(runner)
+
     def _do_evaluate(self, runner):
         """perform evaluation and save ckpt."""
         if not self._should_evaluate(runner):
@@ -22,8 +60,35 @@ def _do_evaluate(self, runner):
             self._save_ckpt(runner, key_score)
 
 
+# Note: Considering that MMCV's EvalHook updated its interface in V1.3.16,
+# in order to avoid strong version dependency, we did not directly
+# inherit EvalHook but BaseDistEvalHook.
 class DistEvalHook(BaseDistEvalHook):
 
+    def __init__(self, *args, dynamic_intervals=None, **kwargs):
+        super(DistEvalHook, self).__init__(*args, **kwargs)
+
+        self.use_dynamic_intervals = dynamic_intervals is not None
+        if self.use_dynamic_intervals:
+            self.dynamic_milestones, self.dynamic_intervals = \
+                _calc_dynamic_intervals(self.interval, dynamic_intervals)
+
+    def _decide_interval(self, runner):
+        if self.use_dynamic_intervals:
+            progress = runner.epoch if self.by_epoch else runner.iter
+            step = bisect.bisect(self.dynamic_milestones, (progress + 1))
+            # Dynamically modify the evaluation interval
+            self.interval = self.dynamic_intervals[step - 1]
+
+    def before_train_epoch(self, runner):
+        """Evaluate the model only at the start of training by epoch."""
+        self._decide_interval(runner)
+        super().before_train_epoch(runner)
+
+    def before_train_iter(self, runner):
+        self._decide_interval(runner)
+        super().before_train_iter(runner)
+
     def _do_evaluate(self, runner):
         """perform evaluation and save ckpt."""
         # Synchronization of BatchNorm's buffer (running_mean
 
@@ -23,6 +23,7 @@ def __init__(self,
                  skip_type_keys=('Mosaic', 'RandomAffine', 'MixUp')):
         self.num_last_epochs = num_last_epochs
         self.skip_type_keys = skip_type_keys
+        self._restart_dataloader = False
 
     def before_train_epoch(self, runner):
         """Close mosaic and mixup augmentation and switches to use L1 loss."""
@@ -33,6 +34,19 @@ def before_train_epoch(self, runner):
             model = model.module
         if (epoch + 1) == runner.max_epochs - self.num_last_epochs:
             runner.logger.info('No mosaic and mixup aug now!')
+            # The dataset pipeline cannot be updated when persistent_workers
+            # is True, so we need to force the dataloader's multi-process
+            # restart. This is a very hacky approach.
             train_loader.dataset.update_skip_type_keys(self.skip_type_keys)
+            if hasattr(train_loader, 'persistent_workers'
+                       ) and train_loader.persistent_workers is True:
+                train_loader._DataLoader__initialized = False
+                train_loader._iterator = None
+                self._restart_dataloader = True
             runner.logger.info('Add additional L1 loss now!')
             model.bbox_head.use_l1 = True
+        else:
+            # Once the restart is complete, we need to restore
+            # the initialization flag.
+            if self._restart_dataloader:
+                train_loader._DataLoader__initialized = True