Double flip

[konyul]

Do we have to train model with double flip to enhance mAP? Because when I test model with and without doble flip, model with doble flip get worse mAP.

[konyul]

And also in config, there is double_flip args in train_pipeline.

[tianweiy]

Could you send a link to the config you talk about? You don't need to do double flip during training. It is only for test time aug

[konyul]

in nusc_centerpoint_voxelnet_0075voxel_dcn_flip.py, there is DOUBLE_FLIP = True and this DOUBLE_FLIP args goes to voxel_generator which is in train_pipeline.

voxel_generator = dict( range=[-54, -54, -5.0, 54, 54, 3.0], voxel_size=[0.075, 0.075, 0.2], max_points_in_voxel=10, max_voxel_num=[120000, 160000], double_flip=DOUBLE_FLIP )

train_pipeline = [ dict(type="LoadPointCloudFromFile", dataset=dataset_type), dict(type="LoadPointCloudAnnotations", with_bbox=True), dict(type="Preprocess", cfg=train_preprocessor), dict(type="Voxelization", cfg=voxel_generator), dict(type="AssignLabel", cfg=train_cfg["assigner"]), dict(type="Reformat"), # dict(type='PointCloudCollect', keys=['points', 'voxels', 'annotations', 'calib']), ]

[tianweiy]

flip config is only used for testing. you can use nusc_centerpoint_voxelnet_0075voxel_dcn.py for training.

[tianweiy]

the double flip modifies the data arguments so training using this flip config may have some undesired behavior