File size: 6,886 Bytes
201693a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
_base_ = ('yolov8_l_syncbn_fast_8xb16-500e_coco.py')
custom_imports = dict(imports=['yolo_world'],
                      allow_failed_imports=False)

# hyper-parameters
num_classes = 1203
num_training_classes = 80
max_epochs = 100  # Maximum training epochs
close_mosaic_epochs = 2
save_epoch_intervals = 2
text_channels = 512
neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
base_lr = 2e-3
weight_decay = 0.05 / 2
train_batch_size_per_gpu = 16

# model settings
model = dict(
    type='YOLOWorldDetector',
    mm_neck=True,
    num_train_classes=num_training_classes,
    num_test_classes=num_classes,
    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
    backbone=dict(
        _delete_=True,
        type='MultiModalYOLOBackbone',
        image_model={{_base_.model.backbone}},
        text_model=dict(
            type='HuggingCLIPLanguageBackbone',
            model_name='openai/clip-vit-base-patch32',
            frozen_modules=['all'])),
    neck=dict(type='YOLOWolrdDualPAFPN',
              guide_channels=text_channels,
              embed_channels=neck_embed_channels,
              num_heads=neck_num_heads,
              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
              text_enhancder=dict(type='ImagePoolingAttentionModule',
                                  embed_channels=256,
                                  num_heads=8)),
    bbox_head=dict(type='YOLOWorldHead',
                   head_module=dict(type='YOLOWorldHeadModule',
                                    embed_dims=text_channels,
                                    num_classes=num_training_classes)),
    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))

# dataset settings
text_transform = [
    dict(type='RandomLoadText',
         num_neg_samples=(num_classes, num_classes),
         max_num_samples=num_training_classes,
         padding_to_max=True,
         padding_value=''),
    dict(type='mmdet.PackDetInputs',
         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
                    'flip_direction', 'texts'))
]
train_pipeline = [
    *_base_.pre_transform,
    dict(type='MultiModalMosaic',
         img_scale=_base_.img_scale,
         pad_val=114.0,
         pre_transform=_base_.pre_transform),
    dict(
        type='YOLOv5RandomAffine',
        max_rotate_degree=0.0,
        max_shear_degree=0.0,
        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
        max_aspect_ratio=_base_.max_aspect_ratio,
        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
        border_val=(114, 114, 114)),
    *_base_.last_transform[:-1],
    *text_transform,
]
train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
obj365v1_train_dataset = dict(
    type='MultiModalDataset',
    dataset=dict(
        type='YOLOv5Objects365V1Dataset',
        data_root='data/objects365v1/',
        ann_file='annotations/objects365_train.json',
        data_prefix=dict(img='train/'),
        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
    class_text_path='data/texts/obj365v1_class_texts.json',
    pipeline=train_pipeline)

mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
                        data_root='data/mixed_grounding/',
                        ann_file='annotations/final_mixed_train_no_coco.json',
                        data_prefix=dict(img='gqa/images/'),
                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
                        pipeline=train_pipeline)

flickr_train_dataset = dict(
    type='YOLOv5MixedGroundingDataset',
    data_root='data/flickr/',
    ann_file='annotations/final_flickr_separateGT_train.json',
    data_prefix=dict(img='full_images/'),
    filter_cfg=dict(filter_empty_gt=True, min_size=32),
    pipeline=train_pipeline)

train_dataloader = dict(batch_size=train_batch_size_per_gpu,
                        collate_fn=dict(type='yolow_collate'),
                        dataset=dict(_delete_=True,
                                     type='ConcatDataset',
                                     datasets=[
                                         obj365v1_train_dataset,
                                         flickr_train_dataset, mg_train_dataset
                                     ],
                                     ignore_keys=['classes', 'palette']))

test_pipeline = [
    *_base_.test_pipeline[:-1],
    dict(type='LoadText'),
    dict(type='mmdet.PackDetInputs',
         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
                    'scale_factor', 'pad_param', 'texts'))
]
coco_val_dataset = dict(
    _delete_=True,
    type='MultiModalDataset',
    dataset=dict(type='YOLOv5LVISV1Dataset',
                 data_root='data/coco/',
                 test_mode=True,
                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
                 data_prefix=dict(img=''),
                 batch_shapes_cfg=None),
    class_text_path='data/texts/lvis_v1_class_texts.json',
    pipeline=test_pipeline)
val_dataloader = dict(dataset=coco_val_dataset)
test_dataloader = val_dataloader

val_evaluator = dict(type='mmdet.LVISMetric',
                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
                     metric='bbox')
test_evaluator = val_evaluator

# training settings
default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
                     checkpoint=dict(interval=save_epoch_intervals,
                                     rule='greater'))
custom_hooks = [
    dict(type='EMAHook',
         ema_type='ExpMomentumEMA',
         momentum=0.0001,
         update_buffers=True,
         strict_load=False,
         priority=49),
    dict(type='mmdet.PipelineSwitchHook',
         switch_epoch=max_epochs - close_mosaic_epochs,
         switch_pipeline=train_pipeline_stage2)
]
train_cfg = dict(max_epochs=max_epochs,
                 val_interval=10,
                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
                                     _base_.val_interval_stage2)])
optim_wrapper = dict(optimizer=dict(
    _delete_=True,
    type='AdamW',
    lr=base_lr,
    weight_decay=weight_decay,
    batch_size_per_gpu=train_batch_size_per_gpu),
                     paramwise_cfg=dict(bias_decay_mult=0.0,
                                        norm_decay_mult=0.0,
                                        custom_keys={
                                            'backbone.text_model':
                                            dict(lr_mult=0.01),
                                            'logit_scale':
                                            dict(weight_decay=0.0)
                                        }),
                     constructor='YOLOWv5OptimizerConstructor')