Merge pull request #26 from okotaku/feat/dino

[Feature] Support DINO

.pre-commit-config.yaml

@@ -3,8 +3,8 @@ repos:
     rev: 4.0.1
     hooks:
       - id: flake8
-  - repo: https://github.com/zhouzaida/isort
-    rev: 5.12.1
+  - repo: https://github.com/PyCQA/isort
+    rev: 5.11.5
     hooks:
       - id: isort
   - repo: https://github.com/pre-commit/mirrors-yapf

configs/_base_/datasets/coco/coco_detection_dino.py

@@ -0,0 +1,87 @@
+# dataset settings
+dataset_type = 'CocoDataset'
+data_root = 'data/coco/'
+
+file_client_args = dict(backend='disk')
+
+train_pipeline = [
+    dict(type='LoadImageFromFile', file_client_args=file_client_args),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='RandomFlip', prob=0.5),
+    dict(
+        type='RandomChoice',
+        transforms=[
+            [
+                dict(
+                    type='RandomChoiceResize',
+                    scales=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
+                            (608, 1333), (640, 1333), (672, 1333), (704, 1333),
+                            (736, 1333), (768, 1333), (800, 1333)],
+                    keep_ratio=True)
+            ],
+            [
+                dict(
+                    type='RandomChoiceResize',
+                    # The radio of all image in train dataset < 7
+                    # follow the original implement
+                    scales=[(400, 4200), (500, 4200), (600, 4200)],
+                    keep_ratio=True),
+                dict(
+                    type='RandomCrop',
+                    crop_type='absolute_range',
+                    crop_size=(384, 600),
+                    allow_negative_crop=True),
+                dict(
+                    type='RandomChoiceResize',
+                    scales=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
+                            (608, 1333), (640, 1333), (672, 1333), (704, 1333),
+                            (736, 1333), (768, 1333), (800, 1333)],
+                    keep_ratio=True)
+            ]
+        ]),
+    dict(type='PackDetInputs')
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile', file_client_args=file_client_args),
+    dict(type='Resize', scale=(1333, 800), keep_ratio=True),
+    # If you don't have a gt annotation, delete the pipeline
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(
+        type='PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor'))
+]
+train_dataloader = dict(
+    batch_size=2,
+    num_workers=2,
+    persistent_workers=True,
+    sampler=dict(type='DefaultSampler', shuffle=True),
+    batch_sampler=dict(type='AspectRatioBatchSampler'),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        ann_file='annotations/instances_train2017.json',
+        data_prefix=dict(img='train2017/'),
+        filter_cfg=dict(filter_empty_gt=True, min_size=32),
+        pipeline=train_pipeline))
+val_dataloader = dict(
+    batch_size=1,
+    num_workers=2,
+    persistent_workers=True,
+    drop_last=False,
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        test_mode=True,
+        pipeline=test_pipeline))
+test_dataloader = val_dataloader
+
+val_evaluator = dict(
+    type='CocoMetric',
+    ann_file=data_root + 'annotations/instances_val2017.json',
+    metric='bbox',
+    format_only=False)
+test_evaluator = val_evaluator

configs/_base_/datasets/lvis/lvis_detection_dino.py

@@ -0,0 +1,89 @@
+# dataset settings
+dataset_type = 'LVISV1Dataset'
+data_root = 'data/lvis_v1/'
+
+file_client_args = dict(backend='disk')
+
+train_pipeline = [
+    dict(type='LoadImageFromFile', file_client_args=file_client_args),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='RandomFlip', prob=0.5),
+    dict(
+        type='RandomChoice',
+        transforms=[
+            [
+                dict(
+                    type='RandomChoiceResize',
+                    scales=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
+                            (608, 1333), (640, 1333), (672, 1333), (704, 1333),
+                            (736, 1333), (768, 1333), (800, 1333)],
+                    keep_ratio=True)
+            ],
+            [
+                dict(
+                    type='RandomChoiceResize',
+                    # The radio of all image in train dataset < 7
+                    # follow the original implement
+                    scales=[(400, 4200), (500, 4200), (600, 4200)],
+                    keep_ratio=True),
+                dict(
+                    type='RandomCrop',
+                    crop_type='absolute_range',
+                    crop_size=(384, 600),
+                    allow_negative_crop=True),
+                dict(
+                    type='RandomChoiceResize',
+                    scales=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
+                            (608, 1333), (640, 1333), (672, 1333), (704, 1333),
+                            (736, 1333), (768, 1333), (800, 1333)],
+                    keep_ratio=True)
+            ]
+        ]),
+    dict(type='PackDetInputs')
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile', file_client_args=file_client_args),
+    dict(type='Resize', scale=(1333, 800), keep_ratio=True),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(
+        type='PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor'))
+]
+
+train_dataloader = dict(
+    batch_size=4,
+    num_workers=4,
+    persistent_workers=True,
+    sampler=dict(type='DefaultSampler', shuffle=True),
+    batch_sampler=dict(type='AspectRatioBatchSampler'),
+    dataset=dict(
+        type='ClassBalancedDataset',
+        oversample_thr=1e-3,
+        dataset=dict(
+            type=dataset_type,
+            data_root=data_root,
+            ann_file='annotations/lvis_v1_train.json',
+            data_prefix=dict(img=''),
+            filter_cfg=dict(filter_empty_gt=True, min_size=32),
+            pipeline=train_pipeline)))
+val_dataloader = dict(
+    batch_size=1,
+    num_workers=2,
+    persistent_workers=True,
+    drop_last=False,
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        ann_file='annotations/lvis_v1_val.json',
+        data_prefix=dict(img=''),
+        test_mode=True,
+        pipeline=test_pipeline))
+test_dataloader = val_dataloader
+
+val_evaluator = dict(
+    type='LVISMetric',
+    ann_file=data_root + 'annotations/lvis_v1_val.json',
+    metric=['bbox'])
+test_evaluator = val_evaluator

configs/_base_/datasets/paco/paco_lvis_detection_dino.py

@@ -0,0 +1,89 @@
+# dataset settings
+dataset_type = 'LVISV1Dataset'
+data_root = 'data/lvis_v1/'
+
+file_client_args = dict(backend='disk')
+
+train_pipeline = [
+    dict(type='LoadImageFromFile', file_client_args=file_client_args),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='RandomFlip', prob=0.5),
+    dict(
+        type='RandomChoice',
+        transforms=[
+            [
+                dict(
+                    type='RandomChoiceResize',
+                    scales=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
+                            (608, 1333), (640, 1333), (672, 1333), (704, 1333),
+                            (736, 1333), (768, 1333), (800, 1333)],
+                    keep_ratio=True)
+            ],
+            [
+                dict(
+                    type='RandomChoiceResize',
+                    # The radio of all image in train dataset < 7
+                    # follow the original implement
+                    scales=[(400, 4200), (500, 4200), (600, 4200)],
+                    keep_ratio=True),
+                dict(
+                    type='RandomCrop',
+                    crop_type='absolute_range',
+                    crop_size=(384, 600),
+                    allow_negative_crop=True),
+                dict(
+                    type='RandomChoiceResize',
+                    scales=[(480, 1333), (512, 1333), (544, 1333), (576, 1333),
+                            (608, 1333), (640, 1333), (672, 1333), (704, 1333),
+                            (736, 1333), (768, 1333), (800, 1333)],
+                    keep_ratio=True)
+            ]
+        ]),
+    dict(type='PackDetInputs')
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile', file_client_args=file_client_args),
+    dict(type='Resize', scale=(1333, 800), keep_ratio=True),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(
+        type='PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor'))
+]
+
+train_dataloader = dict(
+    batch_size=4,
+    num_workers=4,
+    persistent_workers=True,
+    sampler=dict(type='DefaultSampler', shuffle=True),
+    batch_sampler=dict(type='AspectRatioBatchSampler'),
+    dataset=dict(
+        type='ClassBalancedDataset',
+        oversample_thr=1e-3,
+        dataset=dict(
+            type=dataset_type,
+            data_root=data_root,
+            ann_file='annotations/paco_lvis_v1_train.json',
+            data_prefix=dict(img=''),
+            filter_cfg=dict(filter_empty_gt=True, min_size=32),
+            pipeline=train_pipeline)))
+val_dataloader = dict(
+    batch_size=1,
+    num_workers=2,
+    persistent_workers=True,
+    drop_last=False,
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        ann_file='annotations/paco_lvis_v1_val.json',
+        data_prefix=dict(img=''),
+        test_mode=True,
+        pipeline=test_pipeline))
+test_dataloader = val_dataloader
+
+val_evaluator = dict(
+    type='LVISMetric',
+    ann_file=data_root + 'annotations/paco_lvis_v1_val.json',
+    metric=['bbox'])
+test_evaluator = val_evaluator

configs/_base_/models/dino/dino-4scale_r50.py

@@ -0,0 +1,83 @@
+model = dict(
+    type='DINO',
+    num_queries=900,  # num_matching_queries
+    with_box_refine=True,
+    as_two_stage=True,
+    data_preprocessor=dict(
+        type='DetDataPreprocessor',
+        mean=[123.675, 116.28, 103.53],
+        std=[58.395, 57.12, 57.375],
+        bgr_to_rgb=True,
+        pad_size_divisor=1),
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        out_indices=(1, 2, 3),
+        frozen_stages=1,
+        norm_cfg=dict(type='BN', requires_grad=False),
+        norm_eval=True,
+        style='pytorch',
+        init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')),
+    neck=dict(
+        type='ChannelMapper',
+        in_channels=[512, 1024, 2048],
+        kernel_size=1,
+        out_channels=256,
+        act_cfg=None,
+        norm_cfg=dict(type='GN', num_groups=32),
+        num_outs=4),
+    encoder=dict(
+        num_layers=6,
+        layer_cfg=dict(
+            self_attn_cfg=dict(embed_dims=256, num_levels=4,
+                               dropout=0.0),  # 0.1 for DeformDETR
+            ffn_cfg=dict(
+                embed_dims=256,
+                feedforward_channels=2048,  # 1024 for DeformDETR
+                ffn_drop=0.0))),  # 0.1 for DeformDETR
+    decoder=dict(
+        num_layers=6,
+        return_intermediate=True,
+        layer_cfg=dict(
+            self_attn_cfg=dict(embed_dims=256, num_heads=8,
+                               dropout=0.0),  # 0.1 for DeformDETR
+            cross_attn_cfg=dict(embed_dims=256, num_levels=4,
+                                dropout=0.0),  # 0.1 for DeformDETR
+            ffn_cfg=dict(
+                embed_dims=256,
+                feedforward_channels=2048,  # 1024 for DeformDETR
+                ffn_drop=0.0)),  # 0.1 for DeformDETR
+        post_norm_cfg=None),
+    positional_encoding=dict(
+        num_feats=128,
+        normalize=True,
+        offset=0.0,  # -0.5 for DeformDETR
+        temperature=20),  # 10000 for DeformDETR
+    bbox_head=dict(
+        type='DINOHead',
+        num_classes=80,
+        sync_cls_avg_factor=True,
+        loss_cls=dict(
+            type='FocalLoss',
+            use_sigmoid=True,
+            gamma=2.0,
+            alpha=0.25,
+            loss_weight=1.0),  # 2.0 in DeformDETR
+        loss_bbox=dict(type='L1Loss', loss_weight=5.0),
+        loss_iou=dict(type='GIoULoss', loss_weight=2.0)),
+    dn_cfg=dict(  # TODO: Move to model.train_cfg ?
+        label_noise_scale=0.5,
+        box_noise_scale=1.0,  # 0.4 for DN-DETR
+        group_cfg=dict(dynamic=True, num_groups=None,
+                       num_dn_queries=100)),  # TODO: half num_dn_queries
+    # training and testing settings
+    train_cfg=dict(
+        assigner=dict(
+            type='HungarianAssigner',
+            match_costs=[
+                dict(type='FocalLossCost', weight=2.0),
+                dict(type='BBoxL1Cost', weight=5.0, box_format='xywh'),
+                dict(type='IoUCost', iou_mode='giou', weight=2.0)
+            ])),
+    test_cfg=dict(max_per_img=300))  # 100 for DeformDETR

configs/_base_/schedules/dino/dino_12e.py

@@ -0,0 +1,33 @@
+# optimizer
+optim_wrapper = dict(
+    type='OptimWrapper',
+    optimizer=dict(
+        type='AdamW',
+        lr=0.0001,  # 0.0002 for DeformDETR
+        weight_decay=0.0001),
+    clip_grad=dict(max_norm=0.1, norm_type=2),
+    paramwise_cfg=dict(custom_keys={'backbone': dict(lr_mult=0.1)})
+)  # custom_keys contains sampling_offsets and reference_points in DeformDETR  # noqa
+
+# learning policy
+max_epochs = 12
+train_cfg = dict(
+    type='EpochBasedTrainLoop', max_epochs=max_epochs, val_interval=1)
+
+val_cfg = dict(type='ValLoop')
+test_cfg = dict(type='TestLoop')
+
+param_scheduler = [
+    dict(
+        type='MultiStepLR',
+        begin=0,
+        end=max_epochs,
+        by_epoch=True,
+        milestones=[11],
+        gamma=0.1)
+]
+
+# NOTE: `auto_scale_lr` is for automatically scaling LR,
+# USER SHOULD NOT CHANGE ITS VALUES.
+# base_batch_size = (8 GPUs) x (2 samples per GPU)
+auto_scale_lr = dict(enable=True, base_batch_size=16)