_base_ = [ '../_base_/datasets/coco_detection.py', '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py', './centernet_tta.py' ] dataset_type = 'CocoDataset' data_root = 'data/coco/' # model settings model = dict( type='CenterNet', data_preprocessor=dict( type='DetDataPreprocessor', mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], bgr_to_rgb=True), backbone=dict( type='ResNet', depth=18, norm_eval=False, norm_cfg=dict(type='BN'), init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet18')), neck=dict( type='CTResNetNeck', in_channels=512, num_deconv_filters=(256, 128, 64), num_deconv_kernels=(4, 4, 4), use_dcn=True), bbox_head=dict( type='CenterNetHead', num_classes=80, in_channels=64, feat_channels=64, loss_center_heatmap=dict(type='GaussianFocalLoss', loss_weight=1.0), loss_wh=dict(type='L1Loss', loss_weight=0.1), loss_offset=dict(type='L1Loss', loss_weight=1.0)), train_cfg=None, test_cfg=dict(topk=100, local_maximum_kernel=3, max_per_img=100)) train_pipeline = [ dict(type='LoadImageFromFile', backend_args={{_base_.backend_args}}), 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', # The cropped images are padded into squares during training, # but may be less than crop_size. crop_size=(512, 512), 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), # Make sure the output is always crop_size. dict(type='Resize', scale=(512, 512), keep_ratio=True), dict(type='RandomFlip', prob=0.5), dict(type='PackDetInputs') ] test_pipeline = [ dict( type='LoadImageFromFile', backend_args={{_base_.backend_args}}, to_float32=True), # don't need Resize 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='LoadAnnotations', with_bbox=True), dict( type='PackDetInputs', meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'border')) ] # Use RepeatDataset to speed up training train_dataloader = dict( batch_size=16, num_workers=4, persistent_workers=True, sampler=dict(type='DefaultSampler', shuffle=True), dataset=dict( _delete_=True, type='RepeatDataset', times=5, 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, backend_args={{_base_.backend_args}}, ))) val_dataloader = dict(dataset=dict(pipeline=test_pipeline)) test_dataloader = val_dataloader # optimizer # 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. optim_wrapper = dict(clip_grad=dict(max_norm=35, norm_type=2)) max_epochs = 28 # learning policy # Based on the default settings of modern detectors, we added warmup settings. param_scheduler = [ dict( type='LinearLR', start_factor=0.001, by_epoch=False, begin=0, end=1000), dict( type='MultiStepLR', begin=0, end=max_epochs, by_epoch=True, milestones=[18, 24], # the real step is [18*5, 24*5] gamma=0.1) ] train_cfg = dict(max_epochs=max_epochs) # the real epoch is 28*5=140 # 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(base_batch_size=128)