Config

  1. model = dict(
  2.     type='FastRCNN',
  3.     backbone=dict(
  4.         type='ResNet',
  5.         depth=50,
  6.         num_stages=4,
  7.         out_indices=(0, 1, 2, 3),
  8.         frozen_stages=1,
  9.         norm_cfg=dict(type='BN', requires_grad=True),
  10.         norm_eval=True,
  11.         style='pytorch',
  12.         init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')),
  13.     neck=dict(
  14.         type='FPN',
  15.         in_channels=[256, 512, 1024, 2048],
  16.         out_channels=256,
  17.         num_outs=5),
  18.     roi_head=dict(
  19.         type='StandardRoIHead',
  20.         bbox_roi_extractor=dict(
  21.             type='SingleRoIExtractor',
  22.             roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
  23.             out_channels=256,
  24.             featmap_strides=[4, 8, 16, 32]),
  25.         bbox_head=dict(
  26.             type='Shared2FCBBoxHead',
  27.             in_channels=256,
  28.             fc_out_channels=1024,
  29.             roi_feat_size=7,
  30.             num_classes=1,
  31.             bbox_coder=dict(
  32.                 type='DeltaXYWHBBoxCoder',
  33.                 target_means=[0.0, 0.0, 0.0, 0.0],
  34.                 target_stds=[0.1, 0.1, 0.2, 0.2]),
  35.             reg_class_agnostic=False,
  36.             loss_cls=dict(
  37.                 type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
  38.             loss_bbox=dict(type='L1Loss', loss_weight=1.0))),
  39.     train_cfg=dict(
  40.         rcnn=dict(
  41.             assigner=dict(
  42.                 type='MaxIoUAssigner',
  43.                 pos_iou_thr=0.5,
  44.                 neg_iou_thr=0.5,
  45.                 min_pos_iou=0.5,
  46.                 match_low_quality=False,
  47.                 ignore_iof_thr=-1),
  48.             sampler=dict(
  49.                 type='RandomSampler',
  50.                 num=512,
  51.                 pos_fraction=0.25,
  52.                 neg_pos_ub=-1,
  53.                 add_gt_as_proposals=True),
  54.             pos_weight=-1,
  55.             debug=False)),
  56.     test_cfg=dict(
  57.         rcnn=dict(
  58.             score_thr=0.05,
  59.             nms=dict(type='nms', iou_threshold=0.5),
  60.             max_per_img=100)))
  61. dataset_type = 'CocoDataset'
  62. data_root = 'data/coco/'
  63. img_norm_cfg = dict(
  64.     mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
  65. train_pipeline = [
  66.     dict(type='LoadImageFromFile'),
  67.     dict(type='LoadProposals', num_max_proposals=2000),
  68.     dict(type='LoadAnnotations', with_bbox=True),
  69.     dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
  70.     dict(type='RandomFlip', flip_ratio=0.5),
  71.     dict(
  72.         type='Normalize',
  73.         mean=[123.675, 116.28, 103.53],
  74.         std=[58.395, 57.12, 57.375],
  75.         to_rgb=True),
  76.     dict(type='Pad', size_divisor=32),
  77.     dict(type='DefaultFormatBundle'),
  78.     dict(type='Collect', keys=['img', 'proposals', 'gt_bboxes', 'gt_labels'])
  79. ]
  80. test_pipeline = [
  81.     dict(type='LoadImageFromFile'),
  82.     dict(type='LoadProposals', num_max_proposals=None),
  83.     dict(
  84.         type='MultiScaleFlipAug',
  85.         img_scale=(1333, 800),
  86.         flip=False,
  87.         transforms=[
  88.             dict(type='Resize', keep_ratio=True),
  89.             dict(type='RandomFlip'),
  90.             dict(
  91.                 type='Normalize',
  92.                 mean=[123.675, 116.28, 103.53],
  93.                 std=[58.395, 57.12, 57.375],
  94.                 to_rgb=True),
  95.             dict(type='Pad', size_divisor=32),
  96.             dict(type='ImageToTensor', keys=['img']),
  97.             dict(type='ToTensor', keys=['proposals']),
  98.             dict(
  99.                 type='ToDataContainer',
  100.                 fields=[dict(key='proposals', stack=False)]),
  101.             dict(type='Collect', keys=['img', 'proposals'])
  102.         ])
  103. ]
  104. data = dict(
  105.     samples_per_gpu=2,
  106.     workers_per_gpu=2,
  107.     train=dict(
  108.         type='TableDataset',
  109.         ann_file='data/icdar2019/test.json',
  110.         img_prefix='data/icdar2019/test/TRACKA/',
  111.         pipeline=[
  112.             dict(type='LoadImageFromFile'),
  113.             dict(type='LoadProposals', num_max_proposals=2000),
  114.             dict(type='LoadAnnotations', with_bbox=True),
  115.             dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
  116.             dict(type='RandomFlip', flip_ratio=0.5),
  117.             dict(
  118.                 type='Normalize',
  119.                 mean=[123.675, 116.28, 103.53],
  120.                 std=[58.395, 57.12, 57.375],
  121.                 to_rgb=True),
  122.             dict(type='Pad', size_divisor=32),
  123.             dict(type='DefaultFormatBundle'),
  124.             dict(
  125.                 type='Collect',
  126.                 keys=['img', 'proposals', 'gt_bboxes', 'gt_labels'])
  127.         ],
  128.         proposal_file='work_dirs/rpn_r50_fpn_1x_icdar2019/epoch_12.pkl'),
  129.     val=dict(
  130.         type='TableDataset',
  131.         ann_file='data/icdar2019/test.json',
  132.         img_prefix='data/icdar2019/test/TRACKA/',
  133.         pipeline=[
  134.             dict(type='LoadImageFromFile'),
  135.             dict(type='LoadProposals', num_max_proposals=None),
  136.             dict(
  137.                 type='MultiScaleFlipAug',
  138.                 img_scale=(1333, 800),
  139.                 flip=False,
  140.                 transforms=[
  141.                     dict(type='Resize', keep_ratio=True),
  142.                     dict(type='RandomFlip'),
  143.                     dict(
  144.                         type='Normalize',
  145.                         mean=[123.675, 116.28, 103.53],
  146.                         std=[58.395, 57.12, 57.375],
  147.                         to_rgb=True),
  148.                     dict(type='Pad', size_divisor=32),
  149.                     dict(type='ImageToTensor', keys=['img']),
  150.                     dict(type='ToTensor', keys=['proposals']),
  151.                     dict(
  152.                         type='ToDataContainer',
  153.                         fields=[dict(key='proposals', stack=False)]),
  154.                     dict(type='Collect', keys=['img', 'proposals'])
  155.                 ])
  156.         ],
  157.         proposal_file='work_dirs/rpn_r50_fpn_1x_icdar2019/epoch_12.pkl'),
  158.     test=dict(
  159.         type='TableDataset',
  160.         ann_file='data/icdar2019/test.json',
  161.         img_prefix='data/icdar2019/test/TRACKA/',
  162.         pipeline=[
  163.             dict(type='LoadImageFromFile'),
  164.             dict(type='LoadProposals', num_max_proposals=None),
  165.             dict(
  166.                 type='MultiScaleFlipAug',
  167.                 img_scale=(1333, 800),
  168.                 flip=False,
  169.                 transforms=[
  170.                     dict(type='Resize', keep_ratio=True),
  171.                     dict(type='RandomFlip'),
  172.                     dict(
  173.                         type='Normalize',
  174.                         mean=[123.675, 116.28, 103.53],
  175.                         std=[58.395, 57.12, 57.375],
  176.                         to_rgb=True),
  177.                     dict(type='Pad', size_divisor=32),
  178.                     dict(type='ImageToTensor', keys=['img']),
  179.                     dict(type='ToTensor', keys=['proposals']),
  180.                     dict(
  181.                         type='ToDataContainer',
  182.                         fields=[dict(key='proposals', stack=False)]),
  183.                     dict(type='Collect', keys=['img', 'proposals'])
  184.                 ])
  185.         ],
  186.         proposal_file='data/coco/proposals/rpn_r50_fpn_1x_val2017.pkl'))
  187. evaluation = dict(interval=1, metric='bbox')
  188. optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
  189. optimizer_config = dict(grad_clip=None)
  190. lr_config = dict(
  191.     policy='step',
  192.     warmup='linear',
  193.     warmup_iters=500,
  194.     warmup_ratio=0.001,
  195.     step=[8, 11])
  196. runner = dict(type='EpochBasedRunner', max_epochs=12)
  197. checkpoint_config = dict(interval=1)
  198. log_config = dict(interval=50, hooks=[dict(type='TextLoggerHook')])
  199. custom_hooks = [dict(type='NumClassCheckHook'), dict(type='UploadHook')]
  200. dist_params = dict(backend='nccl')
  201. log_level = 'INFO'
  202. load_from = None
  203. resume_from = None
  204. workflow = [('train', 1)]
  205. work_dir = './work_dirs/fast_rcnn_r50_fpn_1x_table_icdar2019'
  206. gpu_ids = range(0, 1)

前置基础知识

Config (定义了Config文件的格式以及引入的格式)
Registry (注册器实现编写Config文件实例化类)
Runner（实例化是实例化了，使用）

计算参数量

  1. python tools/analysis_tools/get_flops.py configs/tabnet/fast_rcnn_r50_fpn_1x_table_icdar2019.py

image.png