fcenet_resnet50-dcnv2_fpn_1500e_ctw1500.py

_base_ = [
    '_base_fcenet_resnet50-dcnv2_fpn.py',
    '../_base_/datasets/ctw1500.py',
    '../_base_/default_runtime.py',
    '../_base_/schedules/schedule_sgd_base.py',
]

optim_wrapper = dict(optimizer=dict(lr=1e-3, weight_decay=5e-4))
train_cfg = dict(max_epochs=1500)
# learning policy
param_scheduler = [
    dict(type='PolyLR', power=0.9, eta_min=1e-7, end=1500),
]

# dataset settings
ctw1500_textdet_train = _base_.ctw1500_textdet_train
ctw1500_textdet_test = _base_.ctw1500_textdet_test

# test pipeline for CTW1500
ctw_test_pipeline = [
    dict(type='LoadImageFromFile', color_type='color_ignore_orientation'),
    dict(type='Resize', scale=(1080, 736), keep_ratio=True),
    # add loading annotation after ``Resize`` because ground truth
    # does not need to do resize data transform
    dict(
        type='LoadOCRAnnotations',
        with_polygon=True,
        with_bbox=True,
        with_label=True),
    dict(
        type='PackTextDetInputs',
        meta_keys=('img_path', 'ori_shape', 'img_shape', 'scale_factor'))
]

ctw1500_textdet_train.pipeline = _base_.train_pipeline
ctw1500_textdet_test.pipeline = ctw_test_pipeline

train_dataloader = dict(
    batch_size=8,
    num_workers=4,
    persistent_workers=True,
    sampler=dict(type='DefaultSampler', shuffle=True),
    dataset=ctw1500_textdet_train)

val_dataloader = dict(
    batch_size=1,
    num_workers=1,
    persistent_workers=True,
    sampler=dict(type='DefaultSampler', shuffle=False),
    dataset=ctw1500_textdet_test)

test_dataloader = val_dataloader

auto_scale_lr = dict(base_batch_size=8)