freshn
commented
2 years ago The config file I used is:
1. data
dataset_type = 'WIDERFaceDataset' data_root = 'data/WIDERFace/' img_norm_cfg = dict( mean=[123.675, 116.28, 103.53], std=[1, 1, 1], to_rgb=True) size_divisor = 32
data = dict( samples_per_gpu=4, workers_per_gpu=2, train=dict( typename=dataset_type, ann_file=data_root + 'WIDER_train/train.txt', img_prefix=data_root + 'WIDER_train/', min_size=1, offset=0, pipeline=[ dict(typename='LoadImageFromFile', to_float32=True), dict(typename='LoadAnnotations', with_bbox=True), dict(typename='RandomSquareCrop', crop_choice=[0.3, 0.45, 0.6, 0.8, 1.0]), dict( typename='PhotoMetricDistortion', brightness_delta=32, contrast_range=(0.5, 1.5), saturation_range=(0.5, 1.5), hue_delta=18), dict(typename='RandomFlip', flip_ratio=0.5), dict(typename='Resize', img_scale=(640, 640), keep_ratio=False), dict(typename='Normalize', img_norm_cfg), dict(typename='DefaultFormatBundle'), dict(typename='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_bboxes_ignore']), ]), val=dict( typename=dataset_type, ann_file=data_root + 'WIDER_val/val.txt', img_prefix=data_root + 'WIDER_val/', min_size=1, offset=0, pipeline=[ dict(typename='LoadImageFromFile'), dict( typename='MultiScaleFlipAug', img_scale=(1100, 1650), flip=False, transforms=[ dict(typename='Resize', keep_ratio=True), dict(typename='RandomFlip', flip_ratio=0.0), dict(typename='Normalize', img_norm_cfg), dict(typename='Pad', size_divisor=32, pad_val=0), dict(typename='ImageToTensor', keys=['img']), dict(typename='Collect', keys=['img']) ]) ]), )
2. model
num_classes = 1 strides = [4, 8, 16, 32, 64, 128] use_sigmoid = True scales_per_octave = 3 ratios = [1.3] num_anchors = scales_per_octave * len(ratios)
model = dict( typename='SingleStageDetector', backbone=dict( typename='ResNet', depth=50, num_stages=4, out_indices=(0, 1, 2, 3), norm_cfg=dict(typename='BN'), norm_eval=False, dcn=None, style='pytorch'), neck=[ dict( typename='FPN', in_channels=[256, 512, 1024, 2048], out_channels=256, start_level=0, add_extra_convs='on_input', num_outs=6, norm_cfg=dict(typename='BN'), upsample_cfg=dict(mode='bilinear')), dict( typename='Inception', in_channel=256, num_levels=6, norm_cfg=dict(typename='BN'), share=True) ], head=dict( typename='IoUAwareRetinaHead', num_classes=num_classes, num_anchors=num_anchors, in_channels=256, stacked_convs=4, feat_channels=256, norm_cfg=dict(typename='BN'), use_sigmoid=use_sigmoid))
3. engines
meshgrid = dict( typename='BBoxAnchorMeshGrid', strides=strides, base_anchor=dict( typename='BBoxBaseAnchor', octave_base_scale=2**(4 / 3), scales_per_octave=scales_per_octave, ratios=ratios, base_sizes=strides))
bbox_coder = dict( typename='DeltaXYWHBBoxCoder', target_means=[.0, .0, .0, .0], target_stds=[0.1, 0.1, 0.2, 0.2])
train_engine = dict( typename='TrainEngine', model=model, criterion=dict( typename='IoUBBoxAnchorCriterion', num_classes=num_classes, meshgrid=meshgrid, bbox_coder=bbox_coder, loss_cls=dict( typename='FocalLoss', use_sigmoid=use_sigmoid, gamma=2.0, alpha=0.25, loss_weight=1.0), reg_decoded_bbox=True, loss_bbox=dict(typename='DIoULoss', loss_weight=2.0), loss_iou=dict( typename='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0), train_cfg=dict( assigner=dict( typename='MaxIoUAssigner', pos_iou_thr=0.35, neg_iou_thr=0.35, min_pos_iou=0.35, ignore_iof_thr=-1, gpu_assign_thr=100), allowed_border=-1, pos_weight=-1, debug=False)), optimizer=dict(typename='SGD', lr=3.75e-3, momentum=0.9, weight_decay=5e-4)) # 3 GPUS
3.2 val engine
val_engine = dict( typename='ValEngine', model=model, meshgrid=meshgrid, converter=dict( typename='IoUBBoxAnchorConverter', num_classes=num_classes, bbox_coder=bbox_coder, nms_pre=-1, use_sigmoid=use_sigmoid), num_classes=num_classes, test_cfg=dict( min_bbox_size=0, score_thr=0.01, nms=dict(typename='lb_nms', iou_thr=0.45), max_per_img=-1), use_sigmoid=use_sigmoid, eval_metric=None)
hooks = [ dict(typename='OptimizerHook'), dict( typename='CosineRestartLrSchedulerHook', periods=[30] 21, restart_weights=[1] 21, warmup='linear', warmup_iters=500, warmup_ratio=1e-1, min_lr_ratio=1e-2), dict(typename='EvalHook'), dict(typename='SnapshotHook', interval=1, save_optim_flag=False, max_keep_ckpts=3), dict(typename='LoggerHook', interval=100) ]
5. work modes
modes = ['train']#, 'val'] max_epochs = 630
6. checkpoint
weights = dict( filepath='torchvision://resnet50', prefix='backbone')
7. misc
seed = 1234 dist_params = dict(backend='nccl') log_level = 'INFO'
dosemeion
@Media-Smart @mileistone @hxcai @mike112223 @DarthThomas Thank you for providing this repository!
I train tinaface with the command
tools/dist_trainval.sh configs/trainval/tinaface/tinaface_r50_fpn_bn.py '0,1,2'and then evaluate it by
python configs/trainval/tinaface/test_widerface.py configs/trainval/tinaface/tinaface_r50_fpn_bn.py workdir/tinaface_r50_fpn_bn/epoch_630_weights.pthThe result is +-------+-------+-----------+--------+-------+ | class | gts | dets | recall | ap | +-------+-------+-----------+--------+-------+ | face | 31957 | 241873177 | 0.910 | 0.000 | +-------+-------+-----------+--------+-------+ | mAP | | | | 0.000 | +-------+-------+-----------+--------+-------+I am sure the dataset is correctly prepared because when I load the checkpoint you provide, I can get the same evaluation results as yours. I am quite confused why the mAP comes to 0 when I train it myself.