Training a customised Deformable DETR

[saadyousuf45]

In deformable DETR code I am implementing: ''' base = [ '../base/datasets/coco_detection.py', '../base/default_runtime.py' ]

model = dict( type='DeformableDETR', backbone=dict( type='ResNet', depth=152, num_stages=10, out_indices=(1, 2, 3, 4, 5, 6, 7, 8, 9), frozen_stages=1, norm_cfg=dict(type='BN', requires_grad=False), norm_eval=True, style='pytorch', init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet152')), neck=dict( type='ChannelMapper', in_channels=[128, 256, 512, 1024, 2048, 4096, 8192, 16384, 32768], kernel_size=1, out_channels=256, act_cfg=None, norm_cfg=dict(type='GN', num_groups=32), num_outs=12 ), bbox_head=dict( type='DeformableDETRHead', num_query=300, num_classes=162, in_channels=2048, sync_cls_avg_factor=True, as_two_stage=False, transformer=dict( type='DeformableDetrTransformer', encoder=dict( type='DetrTransformerEncoder', num_layers=6, transformerlayers=dict( type='BaseTransformerLayer', attn_cfgs=dict( type='MultiScaleDeformableAttention', embed_dims=256), feedforward_channels=1024, ffn_dropout=0.1, operation_order=('self_attn', 'norm', 'ffn', 'norm') ) ) ), positional_encoding=dict( type='SinePositionalEncoding', num_feats=128, normalize=True, offset=-0.5 ), loss_cls_config={}, train_cfg=dict( assigner=dict( type='HungarianAssigner', cls_cost=dict(type='FocalLossCost', weight=2.0), reg_cost=dict(type='BBoxL1Cost', weight=5.0, box_format='xywh'), iou_cost=dict(type='IoUCost', iou_mode='giou', weight=2.0) ) ), test_cfg=dict(max_per_img=100) ) )

loss_cls_config dictionary population

for num_classes in range(0, 161): if num_classes <= 59: model['bbox_head']['loss_cls_config'][num_classes] = { 'type': 'FocalLoss', 'use_sigmoid': True, 'gamma': 2.0, 'alpha': 0.25, 'loss_weight': 2.0 } elif 60 <= num_classes <= 61: model['bbox_head']['loss_cls_config'][num_classes] = { 'type': 'BinaryCrossEntropyLoss', 'loss_weight': 1.0 } else: model['bbox_head']['loss_cls_config'][num_classes] = { 'type': 'L1Loss', 'loss_weight': 1.0 }

Construct the loss configuration

model['bbox_head']['loss_cfg'] = dict( loss_cls=model['bbox_head']['loss_cls_config'], loss_bbox=dict(type='L1Loss', loss_weight=5.0), loss_iou=dict(type='GIoULoss', loss_weight=2.0) )

img_norm_cfg = dict( mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) train_pipeline = [

...

] test_pipeline = [

...

] data = dict( samples_per_gpu=2, workers_per_gpu=2, train=dict(filter_empty_gt=False, pipeline=train_pipeline), val=dict(pipeline=test_pipeline), test=dict(pipeline=test_pipeline) ) optimizer = dict(

...

) optimizer_config = dict(grad_clip=dict(max_norm=0.1, norm_type=2)) lr_config = dict(policy='step', step=[40]) runner = dict(type='EpochBasedRunner', max_epochs=50) auto_scale_lr = dict(base_batch_size=32) '''

Error I am getting:

TypeError Traceback (most recent call last) File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:69, in build_from_cfg(cfg, registry, default_args) 68 try: ---> 69 return obj_cls(**args) 70 except Exception as e: 71 # Normal TypeError does not print class name.

TypeError: init() got an unexpected keyword argument 'bg_cls_weight'

During handling of the above exception, another exception occurred:

TypeError Traceback (most recent call last) File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:69, in build_from_cfg(cfg, registry, default_args) 68 try: ---> 69 return obj_cls(**args) 70 except Exception as e: 71 # Normal TypeError does not print class name.

File ~/Desktop/mmdetection/mmdet/models/dense_heads/deformable_detr_head.py:47, in DeformableDETRHead.init(self, with_box_refine, as_two_stage, transformer, *args, *kwargs) 45 transformer['as_two_stage'] = self.as_two_stage ---> 47 super(DeformableDETRHead, self).init( 48 args, transformer=transformer, **kwargs)

File ~/Desktop/mmdetection/mmdet/models/dense_heads/detr_head.py:130, in DETRHead.init(self, num_classes, in_channels, num_query, num_reg_fcs, transformer, sync_cls_avg_factor, positional_encoding, loss_cls, loss_bbox, loss_iou, train_cfg, test_cfg, init_cfg, **kwargs) 129 self.fp16_enabled = False --> 130 self.loss_cls = build_loss(loss_cls) 131 self.loss_bbox = build_loss(loss_bbox)

File ~/Desktop/mmdetection/mmdet/models/builder.py:45, in build_loss(cfg) 44 """Build loss.""" ---> 45 return LOSSES.build(cfg)

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:237, in Registry.build(self, *args, kwargs) 236 def build(self, *args, *kwargs): --> 237 return self.build_func(args, kwargs, registry=self)

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/cnn/builder.py:27, in build_model_from_cfg(cfg, registry, default_args) 26 else: ---> 27 return build_from_cfg(cfg, registry, default_args)

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:72, in build_from_cfg(cfg, registry, default_args) 70 except Exception as e: 71 # Normal TypeError does not print class name. ---> 72 raise type(e)(f'{obj_cls.name}: {e}')

TypeError: CrossEntropyLoss: init() got an unexpected keyword argument 'bg_cls_weight'

During handling of the above exception, another exception occurred:

TypeError Traceback (most recent call last) File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:69, in build_from_cfg(cfg, registry, default_args) 68 try: ---> 69 return obj_cls(**args) 70 except Exception as e: 71 # Normal TypeError does not print class name.

File ~/Desktop/mmdetection/mmdet/models/detectors/deformable_detr.py:10, in DeformableDETR.init(self, *args, kwargs) 9 def init(self, *args, *kwargs): ---> 10 super(DETR, self).init(args, kwargs)

File ~/Desktop/mmdetection/mmdet/models/detectors/single_stage.py:37, in SingleStageDetector.init(self, backbone, neck, bbox_head, train_cfg, test_cfg, pretrained, init_cfg) 36 bbox_head.update(test_cfg=test_cfg) ---> 37 self.bbox_head = build_head(bbox_head) 38 self.train_cfg = train_cfg

File ~/Desktop/mmdetection/mmdet/models/builder.py:40, in build_head(cfg) 39 """Build head.""" ---> 40 return HEADS.build(cfg)

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:237, in Registry.build(self, *args, kwargs) 236 def build(self, *args, *kwargs): --> 237 return self.build_func(args, kwargs, registry=self)

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/cnn/builder.py:27, in build_model_from_cfg(cfg, registry, default_args) 26 else: ---> 27 return build_from_cfg(cfg, registry, default_args)

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:72, in build_from_cfg(cfg, registry, default_args) 70 except Exception as e: 71 # Normal TypeError does not print class name. ---> 72 raise type(e)(f'{obj_cls.name}: {e}')

TypeError: DeformableDETRHead: CrossEntropyLoss: init() got an unexpected keyword argument 'bg_cls_weight'

During handling of the above exception, another exception occurred:

TypeError Traceback (most recent call last) Input In [8], in <cell line: 9>() 4 from mmdet.utils import get_device 7 datasets = [build_dataset(cfg.data.train)] ----> 9 model = build_detector( 10 cfg.model, train_cfg=cfg.get('train_cfg'), test_cfg=cfg.get('test_cfg')) 12 model.CLASSES = datasets[0].CLASSES 14 #model.CLASSES = datasets[0].CLASSES[-100:]

File ~/Desktop/mmdetection/mmdet/models/builder.py:58, in build_detector(cfg, train_cfg, test_cfg) 54 assert cfg.get('train_cfg') is None or train_cfg is None, \ 55 'train_cfg specified in both outer field and model field ' 56 assert cfg.get('test_cfg') is None or test_cfg is None, \ 57 'test_cfg specified in both outer field and model field ' ---> 58 return DETECTORS.build( 59 cfg, default_args=dict(train_cfg=train_cfg, test_cfg=test_cfg))

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:237, in Registry.build(self, *args, kwargs) 236 def build(self, *args, *kwargs): --> 237 return self.build_func(args, kwargs, registry=self)

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/cnn/builder.py:27, in build_model_from_cfg(cfg, registry, default_args) 25 return Sequential(*modules) 26 else: ---> 27 return build_from_cfg(cfg, registry, default_args)

File ~/anaconda3/envs/openmmlab/lib/python3.8/site-packages/mmcv/utils/registry.py:72, in build_from_cfg(cfg, registry, default_args) 69 return obj_cls(**args) 70 except Exception as e: 71 # Normal TypeError does not print class name. ---> 72 raise type(e)(f'{obj_cls.name}: {e}')

TypeError: DeformableDETR: DeformableDETRHead: CrossEntropyLoss: init() got an unexpected keyword argument 'bg_cls_weight'

Please advise what can I do to get rid of it.

[saadyousuf45]

Any progress?

[tie-gang]

I am reading DETR paper. I noticed that most papers using the mmdetection framework have PyTorch version 1.9 and cuda 10.2. However, when I tried to reproduce it using an NVIDIA 3060 graphics card, I found that the versions didn't match. It seems that PyTorch 1.9.0 is a bit too low. How can I solve this issue? Looking forward to your reply.