nwpu数据集map问题

[wzr0108]

作者您好，我用nwpu数据集试了samseg-maskrcnn和rsprompter_anchor两种方法，但是得到的结果和论文不一样，不知道是不是我哪里配置错了。 绿色是samseg-maskrcnn的结果(segm_mAP_copypaste: 0.670 0.927 0.708 0.192 0.629 0.799)，黄色是rsprompter_anchor的结果(segm_mAP_copypaste: 0.451 0.689 0.451 0.021 0.388 0.719)，samseg-maskrcnn更高，与论文的结论相反。 以下是我的环境：

python 3.8
torch==2.0.0+cu117
mmcv==2.1.0
mmengine==0.10.1

samseg-maskrcnn的配置

_base_ = ['_base_/samseg-maskrcnn.py']

work_dir = '/hy-tmp/work_dirs/rsprompter/samseg-maskrcnn-nwpu'

default_hooks = dict(
    timer=dict(type='IterTimerHook'),
    logger=dict(type='LoggerHook', interval=5),
    param_scheduler=dict(type='ParamSchedulerHook'),
    checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=5, save_best='coco/bbox_mAP', rule='greater', save_last=True),
    sampler_seed=dict(type='DistSamplerSeedHook'),
    # visualization=dict(type='DetVisualizationHook', draw=True, interval=1, test_out_dir='vis_data')
)

vis_backends = [dict(type='LocalVisBackend'),
                dict(type="TensorboardVisBackend"),
                ]
visualizer = dict(
    type='DetLocalVisualizer', vis_backends=vis_backends, name='visualizer')

num_classes = 10

# sam base model
hf_sam_pretrain_name = "checkpoints/sam-vit-base"
hf_sam_pretrain_ckpt_path = "checkpoints/sam-vit-base/pytorch_model.bin"
# # sam large model
# hf_sam_pretrain_name = "facebook/sam-vit-large"
# hf_sam_pretrain_ckpt_path = "~/.cache//huggingface/hub/models--facebook--sam-vit-large/snapshots/70009d56dac23ebb3265377257158b1d6ed4c802/pytorch_model.bin"
# # sam huge model
# hf_sam_pretrain_name = "facebook/sam-vit-huge"
# hf_sam_pretrain_ckpt_path = "~/.cache/huggingface/hub/models--facebook--sam-vit-huge/snapshots/89080d6dcd9a900ebd712b13ff83ecf6f072e798/pytorch_model.bin"

model = dict(
    type='SAMSegMaskRCNN',
    backbone=dict(
        hf_pretrain_name=hf_sam_pretrain_name,
        init_cfg=dict(type='Pretrained', checkpoint=hf_sam_pretrain_ckpt_path)),
    neck=dict(
        feature_aggregator=dict(
            in_channels="facebook/sam-vit-base",
            hidden_channels=32,
            select_layers=range(1, 13, 2),
            #### should be changed when using different pretrain model, base: range(1, 13, 2), large: range(1, 25, 2), huge: range(1, 33, 2)
        ),
    ),
    roi_head=dict(
        bbox_head=dict(
            num_classes=num_classes
        ),
        mask_head=dict(
            num_classes=num_classes
        ),
    ),
)

dataset_type = 'NWPUInsSegDataset'

#### should be changed align with your code root and data root
code_root = '/root/RSPrompter'
data_root = '/root/RSPrompter/data/NWPU'

batch_size_per_gpu = 2
num_workers = 2
persistent_workers = True
train_dataloader = dict(
    batch_size=batch_size_per_gpu,
    num_workers=num_workers,
    persistent_workers=persistent_workers,
    dataset=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file=code_root + '/data/NWPU/annotations/NWPU_instances_train.json',
        data_prefix=dict(img='imgs'),
    )
)

val_dataloader = dict(
    batch_size=batch_size_per_gpu,
    num_workers=num_workers,
    persistent_workers=persistent_workers,
    dataset=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file=code_root + '/data/NWPU/annotations/NWPU_instances_val.json',
        data_prefix=dict(img='imgs'),
    )
)

find_unused_parameters = True
test_dataloader = val_dataloader
resume = False
load_from = None

base_lr = 0.0002
max_epochs = 250
train_cfg = dict(max_epochs=max_epochs)

param_scheduler = [
    dict(
        type='LinearLR', start_factor=0.001, by_epoch=False, begin=0, end=50),
    dict(
        type='CosineAnnealingLR',
        eta_min=base_lr * 0.001,
        begin=1,
        end=max_epochs,
        T_max=max_epochs,
        by_epoch=True
    )
]

optim_wrapper = dict(
    type='AmpOptimWrapper',
    dtype='float16',
    optimizer=dict(
        type='AdamW',
        lr=base_lr,
        weight_decay=0.05
    )
)

rsprompter_anchor的配置

_base_ = ['_base_/rsprompter_anchor.py']

work_dir = '/hy-tmp/work_dirs/rsprompter/rsprompter_anchor-nwpu'

default_hooks = dict(
    timer=dict(type='IterTimerHook'),
    logger=dict(type='LoggerHook', interval=5),
    param_scheduler=dict(type='ParamSchedulerHook'),
    checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=4, save_best='coco/bbox_mAP', rule='greater', save_last=True),
    sampler_seed=dict(type='DistSamplerSeedHook'),
    # visualization=dict(type='DetVisualizationHook', draw=True, interval=1, test_out_dir='vis_data')
)

vis_backends = [dict(type='LocalVisBackend'),
                dict(type="TensorboardVisBackend"),
                ]
visualizer = dict(
    type='DetLocalVisualizer', vis_backends=vis_backends, name='visualizer')

num_classes = 10
prompt_shape = (70, 5)  # (per img pointset, per pointset point)

#### should be changed when using different pretrain model

# sam base model
hf_sam_pretrain_name = "checkpoints/sam-vit-base"
hf_sam_pretrain_ckpt_path = "checkpoints/sam-vit-base/pytorch_model.bin"
# # sam large model
# hf_sam_pretrain_name = "facebook/sam-vit-large"
# hf_sam_pretrain_ckpt_path = "~/.cache//huggingface/hub/models--facebook--sam-vit-large/snapshots/70009d56dac23ebb3265377257158b1d6ed4c802/pytorch_model.bin"
# # sam huge model
# hf_sam_pretrain_name = "facebook/sam-vit-huge"
# hf_sam_pretrain_ckpt_path = "~/.cache/huggingface/hub/models--facebook--sam-vit-huge/snapshots/89080d6dcd9a900ebd712b13ff83ecf6f072e798/pytorch_model.bin"

model = dict(
    decoder_freeze=False,
    shared_image_embedding=dict(
        hf_pretrain_name=hf_sam_pretrain_name,
        init_cfg=dict(type='Pretrained', checkpoint=hf_sam_pretrain_ckpt_path),
    ),
    backbone=dict(
        hf_pretrain_name=hf_sam_pretrain_name,
        init_cfg=dict(type='Pretrained', checkpoint=hf_sam_pretrain_ckpt_path)
    ),
    neck=dict(
        feature_aggregator=dict(
            in_channels="facebook/sam-vit-base",
            hidden_channels=32,
            select_layers=range(1, 13, 2),  #### should be changed when using different pretrain model, base: range(1, 13, 2), large: range(1, 25, 2), huge: range(1, 33, 2)
        ),
    ),
    roi_head=dict(
        bbox_head=dict(
            num_classes=num_classes,
        ),
        mask_head=dict(
            mask_decoder=dict(
                hf_pretrain_name=hf_sam_pretrain_name,
                init_cfg=dict(type='Pretrained', checkpoint=hf_sam_pretrain_ckpt_path)
            ),
            per_pointset_point=prompt_shape[1],
            with_sincos=True,
        ),
    ),
)

dataset_type = 'NWPUInsSegDataset'

#### should be changed align with your code root and data root
code_root = '/root/RSPrompter'
data_root = '/root/RSPrompter/data/NWPU'

batch_size_per_gpu = 1
num_workers = 1
persistent_workers = True
train_dataloader = dict(
    batch_size=batch_size_per_gpu,
    num_workers=num_workers,
    persistent_workers=persistent_workers,
    dataset=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file=code_root + '/data/NWPU/annotations/NWPU_instances_train.json',
        data_prefix=dict(img='imgs'),
    )
)

val_dataloader = dict(
    batch_size=batch_size_per_gpu,
    num_workers=num_workers,
    persistent_workers=persistent_workers,
    dataset=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file=code_root + '/data/NWPU/annotations/NWPU_instances_val.json',
        data_prefix=dict(img='imgs'),
    )
)

find_unused_parameters = True

test_dataloader = val_dataloader
resume = False
load_from = None

base_lr = 0.0002
max_epochs = 500

train_cfg = dict(max_epochs=max_epochs)

param_scheduler = [
    dict(
        type='LinearLR', start_factor=0.001, by_epoch=False, begin=0, end=50),
    dict(
        type='CosineAnnealingLR',
        eta_min=base_lr * 0.001,
        begin=1,
        end=max_epochs,
        T_max=max_epochs,
        by_epoch=True
    )
]

#### AMP training config
runner_type = 'Runner'
optim_wrapper = dict(
    type='AmpOptimWrapper',
    dtype='float16',
    optimizer=dict(
        type='AdamW',
        lr=base_lr,
        weight_decay=0.05)
)

#### DeepSpeed training config
# runner_type = 'FlexibleRunner'
# strategy = dict(
#     type='DeepSpeedStrategy',
#     fp16=dict(
#         enabled=True,
#         auto_cast=False,
#         fp16_master_weights_and_grads=False,
#         loss_scale=0,
#         loss_scale_window=500,
#         hysteresis=2,
#         min_loss_scale=1,
#         initial_scale_power=15,
#     ),
#     inputs_to_half=['inputs'],
#     zero_optimization=dict(
#         stage=2,
#         allgather_partitions=True,
#         allgather_bucket_size=2e8,
#         reduce_scatter=True,
#         reduce_bucket_size='auto',
#         overlap_comm=True,
#         contiguous_gradients=True,
#     ),
# )
# optim_wrapper = dict(
#     type='DeepSpeedOptimWrapper',
#     optimizer=dict(
#         type='AdamW',
#         lr=base_lr,
#         weight_decay=0.05
#     )
# )

[wzr0108]

samseg-maskrcnn的log 20231213_155950.log

[wzr0108]

rsprompter_anchor的log 20231213_200229.log

[KyanChen]

从图上可以看出，你这个黄色的还没有收敛，指标还在上升区间，需要继续训练