grounding_dino_r50_scratch_8xb2_1x_coco.py

_base_ = [
    '../_base_/datasets/coco_detection.py',
    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
]
lang_model_name = 'bert-base-uncased'

model = dict(
    type='GroundingDINO',
    num_queries=900,
    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_mask=False,
    ),
    language_model=dict(
        type='BertModel',
        name=lang_model_name,
        pad_to_max=False,
        use_sub_sentence_represent=True,
        special_tokens_list=['[CLS]', '[SEP]', '.', '?'],
        add_pooling_layer=False,
    ),
    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,
        bias=True,
        norm_cfg=dict(type='GN', num_groups=32),
        num_outs=4),
    encoder=dict(
        num_layers=6,
        num_cp=6,
        # visual layer config
        layer_cfg=dict(
            self_attn_cfg=dict(embed_dims=256, num_levels=4, dropout=0.0),
            ffn_cfg=dict(
                embed_dims=256, feedforward_channels=2048, ffn_drop=0.0)),
        # text layer config
        text_layer_cfg=dict(
            self_attn_cfg=dict(num_heads=4, embed_dims=256, dropout=0.0),
            ffn_cfg=dict(
                embed_dims=256, feedforward_channels=1024, ffn_drop=0.0)),
        # fusion layer config
        fusion_layer_cfg=dict(
            v_dim=256,
            l_dim=256,
            embed_dim=1024,
            num_heads=4,
            init_values=1e-4),
    ),
    decoder=dict(
        num_layers=6,
        return_intermediate=True,
        layer_cfg=dict(
            # query self attention layer
            self_attn_cfg=dict(embed_dims=256, num_heads=8, dropout=0.0),
            # cross attention layer query to text
            cross_attn_text_cfg=dict(embed_dims=256, num_heads=8, dropout=0.0),
            # cross attention layer query to image
            cross_attn_cfg=dict(embed_dims=256, num_heads=8, dropout=0.0),
            ffn_cfg=dict(
                embed_dims=256, feedforward_channels=2048, ffn_drop=0.0)),
        post_norm_cfg=None),
    positional_encoding=dict(
        num_feats=128, normalize=True, offset=0.0, temperature=20),
    bbox_head=dict(
        type='GroundingDINOHead',
        num_classes=80,
        sync_cls_avg_factor=True,
        contrastive_cfg=dict(max_text_len=256, log_scale='auto', bias=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='BinaryFocalLossCost', 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))

# dataset settings
train_pipeline = [
    dict(type='LoadImageFromFile', backend_args=_base_.backend_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',
        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
                   'scale_factor', 'flip', 'flip_direction', 'text',
                   'custom_entities'))
]

test_pipeline = [
    dict(type='LoadImageFromFile', backend_args=_base_.backend_args),
    dict(type='FixScaleResize', scale=(800, 1333), keep_ratio=True),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(
        type='PackDetInputs',
        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
                   'scale_factor', 'text', 'custom_entities'))
]

train_dataloader = dict(
    dataset=dict(
        filter_cfg=dict(filter_empty_gt=False),
        pipeline=train_pipeline,
        return_classes=True))
val_dataloader = dict(
    dataset=dict(pipeline=test_pipeline, return_classes=True))
test_dataloader = val_dataloader

# We did not adopt the official 24e optimizer strategy
# because the results indicate that the current strategy is superior.
optim_wrapper = dict(
    _delete_=True,
    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={
        'absolute_pos_embed': dict(decay_mult=0.),
        'backbone': dict(lr_mult=0.1)
    }))
# 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(base_batch_size=16)