mmdetection/config_rtdetrwithyolo.py

auto_scale_lr = dict(base_batch_size=32)
backend_args = None
pretrained = 'https://github.com/flytocc/mmdetection/releases/download/model_zoo/resnet34vd_pretrained_f6a72dc5.pth'
data_root = '../../../media/tricolops/T7/Dataset/coco_format_bd/'
dataset_type = 'CocoDataset'
interpolations = ['nearest', 'bilinear', 'bicubic', 'area', 'lanczos']
default_hooks = dict(
    checkpoint=dict(interval=5, save_best='auto', type='CheckpointHook'),
    logger=dict(interval=50, type='LoggerHook'),
    param_scheduler=dict(type='ParamSchedulerHook'),
    sampler_seed=dict(type='DistSamplerSeedHook'),
    timer=dict(type='IterTimerHook'),
    visualization=dict(type='DetVisualizationHook'))
custom_hooks = [
    dict(
        type='EMAHook',
        ema_type='ExpMomentumEMA',
        momentum=0.0001,
        update_buffers=True,
        priority=49)
]
default_scope = 'mmdet'
env_cfg = dict(
    cudnn_benchmark=False,
    dist_cfg=dict(backend='nccl'),
    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0))
launcher = 'none'
load_from = None
log_level = 'INFO'
log_processor = dict(by_epoch=True, type='LogProcessor', window_size=50)
max_epochs = 500
metainfo = dict(
    classes=('barcode', ), palette=[
        (
            220,
            20,
            60,
        ),
    ])
model = dict(
    backbone=dict(
            last_stage_out_channels= 768,
            plugins = None,
            deepen_factor = 0.67,
            widen_factor = 0.75,
            input_channels = 1,
            out_indices = (2, 3, 4),
        type='YOLOv8CSPDarknet'),
    bbox_head=dict(
            num_classes=1,
            embed_dims= 256,
            #same as hidding dims in decoder and hybridencoder
            num_reg_fcs= 2,
            sync_cls_avg_factor = False,
            loss_cls = dict(
                type='VarifocalLoss',
                use_sigmoid = True,
                alpha= 0.75,
                gamma= 2.0,
                with_logits=True,
                iou_weighted= True,
                reduction= 'mean',
                loss_weight= 1),
            loss_bbox= dict(type='L1Loss', loss_weight=5.0),
            loss_iou= dict(type='GIoULoss', loss_weight=2.0),
            num_pred_layer = 6,
        type='RTDETRHead'),
    data_preprocessor=dict(
        pad_size_divisor=1,
        type='DetDataPreprocessor'),
    decoder=dict(
        num_classes=1,
        hidden_dim=256,
        num_queries=300,
        position_type='sine',
        feat_channels=[256, 256, 256],
        #after neck it should be 256
        feat_strides=[8,16,32],
        #eval size should be same as encoder feat stride
        num_levels=3,
        num_crossattn_points=4,
        number_head=8,
        number_decoder_layer=4,
        #4 for 34 and 6 for 50
        dim_feedforward_ratio=4,
        dropout=0.0,
        act_cfg= dict(type='ReLU', inplace=True),
        num_denoising=100,
        label_noise_ratio=0.5,
        box_noise_scale=1.0,
        learnt_init_query=True,
        eval_idx=-1,
        eval_spatial_size=[640, 640],
        eps=1e-2
    ),
    neck=dict(
        in_channels=[192, 384, 576],
        #same as feature from back bone using last four layer
        #in_channels=[128,256,512] for resnet34 and [512,1024,2048] for resnet50
        feat_strides=[8,16,32],
        hidden_dim=256,
        n_head=8,
        dim_feedforward_ratio=4,
        drop_out=0.0,
        enc_act=dict(type='GELU'),
        use_encoder_idx=[2],
        num_encoder_layers=1,
        pe_temperature=100*100,
        norm_cfg= dict(type='BN', requires_grad=True),
        widen_factor=0.5,
        deepen_factor=1,
        input_proj_cfg=dict(        
                    type='ChannelMapper',
                        in_channels=[192, 384, 576],
                        #same as input channel in encoder
                        kernel_size=1,
                        out_channels=256,
                        act_cfg=None,
                        norm_cfg=dict(type='BN', requires_grad=True),
                        num_outs=3
                        ),
        eval_spatial_size=[640, 640],
        act_cfg= dict(type='SiLU', inplace=True),
        type='HybridEncoder'),
    num_queries=300,
    test_cfg=dict(max_per_img=100),
    train_cfg=dict(
        assigner=dict(
            match_costs=[
                dict(type='FocalLossCost', weight=2.0),
                dict(box_format='xywh', type='BBoxL1Cost', weight=5.0),
                dict(iou_mode='giou', type='IoUCost', weight=2.0),
            ],
            type='HungarianAssigner')),
    type='RtDetr',)
optim_wrapper = dict(
    clip_grad=dict(max_norm=10, norm_type=2),
    optimizer=dict(
        type='SGD',
        lr=0.0025,
        momentum=0.937,
        weight_decay= 0.0005,
        nesterov=True,),
    type='OptimWrapper')
param_scheduler = [
    dict(
        type='LinearLR', start_factor=0.001, by_epoch=False, begin=0, end=max_epochs)
]
resume = False
test_cfg = dict(type='TestLoop')
test_pipeline = [
    dict(backend_args=None,color_type='grayscale', type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(
        type='Resize',
        scale=(640, 640),
        keep_ratio=False,
        interpolation='bicubic'),
    dict(
        type='PackDetInputs',
        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape','scale_factor')),
]
test_dataloader = dict(
    batch_size=1,
    dataset=dict(
        ann_file='Val/Val.json',
        backend_args=None,
        data_prefix=dict(img='Val/'),
        data_root='../../../media/tricolops/T7/Dataset/coco_format_bd/',
        metainfo=dict(classes=('barcode', ), palette=[
            (
                220,
                20,
                60,
            ),
        ]),
        pipeline=test_pipeline,
        test_mode=True,
        type='CocoDataset'),
    drop_last=False,
    num_workers=2,
    persistent_workers=True,
    sampler=dict(shuffle=False, type='DefaultSampler'))
test_evaluator = dict(
    ann_file='../../../media/tricolops/T7/Dataset/coco_format_bd/Val/Val.json',
    backend_args=None,
    format_only=False,
    metric='bbox',
    type='CocoMetric')
train_cfg = dict(max_epochs=max_epochs, type='EpochBasedTrainLoop', val_interval=1)
train_pipeline = [
    dict(type='LoadImageFromFile', color_type='grayscale',backend_args = None),
    dict(type='LoadAnnotations', with_bbox=True),
    # dict(
    #     type='RandomApply',
    #     transforms=dict(type='PhotoMetricDistortion'),
    #     prob=0.8),
    dict(
        type='RandomApply', transforms=dict(type='MinIoURandomCrop'),
        prob=0.8),
    dict(type='FilterAnnotations', min_gt_bbox_wh=(1, 1), keep_empty=False),
    dict(type='RandomFlip', prob=0.5),
    dict(
        type='RandomChoice',
        transforms=[[
            dict(
                type='Resize',
                scale=(640, 640),
                keep_ratio=False,
                interpolation=interpolation)
        ] for interpolation in interpolations]),
    dict(type='PackDetInputs')
]
train_dataloader = dict(
    batch_sampler=dict(type='AspectRatioBatchSampler'),
    batch_size=12,
    dataset=dict(
        ann_file='Train/Train.json',
        backend_args=None,
        data_prefix=dict(img='Train/'),
        data_root='../../../media/tricolops/T7/Dataset/coco_format_bd/',
        filter_cfg=dict(filter_empty_gt=False, min_size=32),
        metainfo=dict(classes=('barcode', ), palette=[
            (
                220,
                20,
                60,
            ),
        ]),
        pipeline=train_pipeline,
        type='CocoDataset'),
    num_workers=2,
    persistent_workers=True,
    sampler=dict(shuffle=True, type='DefaultSampler'))

val_cfg = dict(type='ValLoop')
val_dataloader = dict(
    batch_size=1,
    dataset=dict(
        ann_file='Val/Val.json',
        backend_args=None,
        data_prefix=dict(img='Val/'),
        data_root='../../../media/tricolops/T7/Dataset/coco_format_bd/',
        metainfo=dict(classes=('barcode', ), palette=[
            (
                220,
                20,
                60,
            ),
        ]),
        pipeline=[
            dict(backend_args=None, color_type="grayscale",type='LoadImageFromFile'),
            dict(keep_ratio=False, scale=(
                640,
                640,
            ), type='Resize'),
            dict(type='LoadAnnotations', with_bbox=True),
            dict(
                meta_keys=(
                    'img_id',
                    'img_path',
                    'ori_shape',
                    'img_shape',
                    'scale_factor',
                ),
                type='PackDetInputs'),
        ],
        test_mode=True,
        type='CocoDataset'),
    drop_last=False,
    num_workers=2,
    persistent_workers=True,
    sampler=dict(shuffle=False, type='DefaultSampler'))
val_evaluator = dict(
    ann_file='../../../media/tricolops/T7/Dataset/coco_format_bd/Val/Val.json',
    backend_args=None,
    format_only=False,
    metric='bbox',
    type='CocoMetric')
vis_backends = [
    dict(type='LocalVisBackend'),
]
visualizer = dict(
    name='visualizer',
    type='DetLocalVisualizer',
    vis_backends=[
        dict(type='LocalVisBackend'),
    ])
work_dir = './work_dirs/rtdetr_yolo'