centernet_resnet34_140e_mydet.py

_base_ = './centernet_resnet34_dcnv2_140e_coco.py'

# model settings
model = dict(
    neck=dict(use_dcn=False),
    bbox_head=dict(num_classes=3,))

# data settings
# dataset type
dataset_type = 'CocoDataset'

# set the classes according to mydet
classes = ('rust', 'scratch', 'spot')

# We fixed the incorrect img_norm_cfg problem in the source code.
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)

# train pipeline
train_pipeline = [
    dict(type='LoadImageFromFile', to_float32=True, color_type='color'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(
        type='PhotoMetricDistortion',
        brightness_delta=32,
        contrast_range=(0.5, 1.5),
        saturation_range=(0.5, 1.5),
        hue_delta=18),
    dict(
        type='RandomCenterCropPad',
        crop_size=(1000, 600),
        ratios=(0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3),
        mean=[0, 0, 0],
        std=[1, 1, 1],
        to_rgb=True,
        test_pad_mode=None),
    dict(type='Resize', img_scale=(1000, 600), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]

# test pipeline
test_pipeline = [
    dict(type='LoadImageFromFile', to_float32=True),
    dict(
        type='MultiScaleFlipAug',
        flip=False,
        scale_factor=1.0,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(
                type='RandomCenterCropPad',
                ratios=None,
                border=None,
                mean=[0, 0, 0],
                std=[1, 1, 1],
                to_rgb=True,
                test_mode=True,
                test_pad_mode=['logical_or', 31],
                test_pad_add_pix=1),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='DefaultFormatBundle'),
            dict(
                type='Collect',
                meta_keys=('filename', 'ori_filename', 'ori_shape',
                           'img_shape', 'pad_shape', 'scale_factor', 'flip',
                           'flip_direction', 'img_norm_cfg', 'border'),
                keys=['img'])
        ])
]

# Use RepeatDataset to speed up training
data = dict(
    samples_per_gpu=16,
    workers_per_gpu=4,
    train=dict(
        _delete_=True,
        type='RepeatDataset',
        times=5,
        dataset=dict(
            type=dataset_type,
            img_prefix='data/images/',
            classes=classes,
            ann_file='data/annotations/train2017.json',
            pipeline=train_pipeline)),
    val=dict(
        img_prefix='data/images/',
        classes=classes,
        ann_file='data/annotations/val2017.json',
        pipeline=test_pipeline),
    test=dict(
        img_prefix='data/images/',
        classes=classes,
        ann_file='data/annotations/test2017.json',
        pipeline=test_pipeline))


# optimizer LR = 0.02(default)
# Based on the default settings of modern detectors, the SGD effect is better
# than the Adam in the source code, so we use SGD default settings and
# if you use adam+lr5e-4, the map is 29.1.
# set LR according to the original paper's settings
# optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(_delete_=True, grad_clip=dict(max_norm=35, norm_type=2))

# NOTE: `auto_scale_lr` is for automatically scaling LR,
# USER SHOULD NOT CHANGE ITS VALUES.
# base_batch_size = (8 GPUs) x (16 samples per GPU)
auto_scale_lr = dict(enable=True, base_batch_size=128)

# learning policy
# Based on the default settings of modern detectors, we added warmup settings.
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=1000,
    warmup_ratio=1.0 / 1000,
    step=[24, 33])  # the real step is [24*5, 33*5]

runner = dict(max_epochs=36)  # the real epoch is 36*5=180

# set checkpoints save every: 5 epochs
checkpoint_config = dict(interval=5)