sskim0126 commented 9 months ago

Hello, I would like to attempt semantic segmentation on the Potsdam dataset using your approach. I have followed all the instructions you provided, but when I conducted tests, I found that the performance was slightly lower than what is indicated.

The performance metrics I obtained are as follows: Metrics from GitHub

ResNet-50
- mIoU: 87.35
Swin-B
- mIoU: 88.36

Metrics from my own experiments

ResNet-50
- mIoU: 70.41
Swin-B
- mIoU: 69.62

I would like to know if I made any mistakes that could have led to the lower performance.

For reference, here is the environment in which I conducted the experiments:

Model weights: CMID - ResNet-50: Link to Download CMID - Swin-B: Link to Download
Conda environment: Name Version Build Channel _libgcc_mutex 0.1 conda_forge conda-forge _openmp_mutex 4.5 2_gnu conda-forge absl-py 2.0.0 pypi_0 pypi addict 2.4.0 pypi_0 pypi aliyun-python-sdk-core 2.14.0 pypi_0 pypi aliyun-python-sdk-kms 2.16.2 pypi_0 pypi appdirs 1.4.4 pypi_0 pypi attrs 23.1.0 pypi_0 pypi bzip2 1.0.8 h7f98852_4 conda-forge ca-certificates 2023.08.22 h06a4308_0 defaults cachetools 5.3.2 pypi_0 pypi certifi 2023.7.22 pypi_0 pypi cffi 1.16.0 pypi_0 pypi charset-normalizer 3.3.1 pypi_0 pypi cityscapesscripts 2.2.2 pypi_0 pypi click 8.1.7 pypi_0 pypi codecov 2.1.13 pypi_0 pypi colorama 0.4.6 pypi_0 pypi coloredlogs 15.0.1 pypi_0 pypi contourpy 1.1.1 pypi_0 pypi coverage 7.3.2 pypi_0 pypi crcmod 1.7 pypi_0 pypi cryptography 41.0.5 pypi_0 pypi cycler 0.12.1 pypi_0 pypi exceptiongroup 1.1.3 pypi_0 pypi flake8 6.1.0 pypi_0 pypi fonttools 4.43.1 pypi_0 pypi google-auth 2.23.3 pypi_0 pypi google-auth-oauthlib 1.0.0 pypi_0 pypi grpcio 1.59.0 pypi_0 pypi humanfriendly 10.0 pypi_0 pypi idna 3.4 pypi_0 pypi importlib-metadata 6.8.0 pypi_0 pypi importlib-resources 6.1.0 pypi_0 pypi iniconfig 2.0.0 pypi_0 pypi interrogate 1.5.0 pypi_0 pypi jmespath 0.10.0 pypi_0 pypi kiwisolver 1.4.5 pypi_0 pypi ld_impl_linux-64 2.40 h41732ed_0 conda-forge libffi 3.4.4 h6a678d5_0 defaults libgcc-ng 13.2.0 h807b86a_2 conda-forge libgomp 13.2.0 h807b86a_2 conda-forge libnsl 2.0.1 hd590300_0 conda-forge libsqlite 3.43.2 h2797004_0 conda-forge libstdcxx-ng 13.2.0 h7e041cc_2 conda-forge libuuid 2.38.1 h0b41bf4_0 conda-forge libzlib 1.2.13 hd590300_5 conda-forge markdown 3.5 pypi_0 pypi markdown-it-py 3.0.0 pypi_0 pypi markupsafe 2.1.3 pypi_0 pypi matplotlib 3.7.3 pypi_0 pypi mccabe 0.7.0 pypi_0 pypi mdurl 0.1.2 pypi_0 pypi mmcls 0.25.0 pypi_0 pypi mmcv-full 1.6.1 pypi_0 pypi mmengine 0.9.0 pypi_0 pypi mmsegmentation 0.27.0 dev_0 model-index 0.1.11 pypi_0 pypi ncurses 6.4 hcb278e6_0 conda-forge numpy 1.24.4 pypi_0 pypi oauthlib 3.2.2 pypi_0 pypi opencv-python 4.8.1.78 pypi_0 pypi opendatalab 0.0.10 pypi_0 pypi openmim 0.3.9 pypi_0 pypi openssl 3.1.4 hd590300_0 conda-forge openxlab 0.0.28 pypi_0 pypi ordered-set 4.1.0 pypi_0 pypi oss2 2.17.0 pypi_0 pypi packaging 23.2 pypi_0 pypi pandas 2.0.3 pypi_0 pypi pillow 10.1.0 pypi_0 pypi pip 23.3.1 pyhd8ed1ab_0 conda-forge platformdirs 3.11.0 pypi_0 pypi pluggy 1.3.0 pypi_0 pypi prettytable 3.9.0 pypi_0 pypi protobuf 4.24.4 pypi_0 pypi py 1.11.0 pypi_0 pypi pyasn1 0.5.0 pypi_0 pypi pyasn1-modules 0.3.0 pypi_0 pypi pycodestyle 2.11.1 pypi_0 pypi pycparser 2.21 pypi_0 pypi pycryptodome 3.19.0 pypi_0 pypi pyflakes 3.1.0 pypi_0 pypi pygments 2.16.1 pypi_0 pypi pyparsing 3.1.1 pypi_0 pypi pyquaternion 0.9.9 pypi_0 pypi pytest 7.4.3 pypi_0 pypi python 3.8.18 hd12c33a_0_cpython conda-forge python-dateutil 2.8.2 pypi_0 pypi pytz 2023.3.post1 pypi_0 pypi pyyaml 6.0.1 pypi_0 pypi readline 8.2 h8228510_1 conda-forge requests 2.28.2 pypi_0 pypi requests-oauthlib 1.3.1 pypi_0 pypi rich 13.4.2 pypi_0 pypi rsa 4.9 pypi_0 pypi setuptools 59.5.0 pypi_0 pypi six 1.16.0 pypi_0 pypi tabulate 0.9.0 pypi_0 pypi tensorboard 2.14.0 pypi_0 pypi tensorboard-data-server 0.7.2 pypi_0 pypi termcolor 2.3.0 pypi_0 pypi tk 8.6.13 h2797004_0 conda-forge toml 0.10.2 pypi_0 pypi tomli 2.0.1 pypi_0 pypi torch 1.10.1+cu111 pypi_0 pypi torchaudio 0.10.1+cu111 pypi_0 pypi torchvision 0.11.2+cu111 pypi_0 pypi tqdm 4.65.2 pypi_0 pypi typing 3.7.4.3 pypi_0 pypi typing-extensions 4.8.0 pypi_0 pypi tzdata 2023.3 pypi_0 pypi urllib3 1.26.18 pypi_0 pypi wcwidth 0.2.8 pypi_0 pypi werkzeug 3.0.1 pypi_0 pypi wheel 0.41.2 pyhd8ed1ab_0 conda-forge xdoctest 1.1.1 pypi_0 pypi xz 5.4.2 h5eee18b_0 defaults yapf 0.40.1 pypi_0 pypi zipp 3.17.0 pypi_0 pypi

pUmpKin-Co commented 9 months ago

Hi~.Thanks for you interest for our work. Could you provide your log and config files for better analysis the problem?

sskim0126 commented 9 months ago

Thank you for your response!

First of all, config files are identical to the files provided on your GitHub repository, except for the changed paths.

Here are the config files we used.

CMID-ResNet


log_config = dict(
interval=50,
hooks=[
    dict(type='TextLoggerHook', by_epoch=False),
    dict(type='TensorboardLoggerHook')
])
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
resume_from = None
workflow = [('train', 1)]
cudnn_benchmark = True

optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)

optimizer_config = dict(type='Fp16OptimizerHook', distributed=False)

lr_config = dict( policy='CosineAnnealing', warmup='linear', warmup_iters=100, min_lr=0.000001, by_epoch=False)

runner = dict(type='EpochBasedRunner', max_epochs=50) checkpoint_config = dict(by_epoch=True, interval=1) evaluation = dict(interval=5, metric=['mIoU', "mFscore"], pre_eval=True, by_epoch=True)

norm_cfg = dict(type='BN', requires_grad=True) checkpoint_path = "./pretrained/CMID-ResNet50-millionAID.pth" model = dict( type='EncoderDecoder', pretrained=checkpoint_path, backbone=dict( type='ResNet', depth=50, num_stages=4, out_indices=(0, 1, 2, 3), dilations=(1, 1, 1, 1), strides=(1, 2, 2, 2), norm_cfg=dict(type='BN', requires_grad=True), norm_eval=False, style='pytorch', contract_dilation=True), decode_head=dict( type='UPerHead', in_channels=[256, 512, 1024, 2048], in_index=[0, 1, 2, 3], pool_scales=(1, 2, 3, 6), channels=512, ignore_index=255, dropout_ratio=0.1, num_classes=6, norm_cfg=dict(type='BN', requires_grad=True), align_corners=False, loss_decode=dict( type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)), auxiliary_head=dict( type='FCNHead', in_channels=1024, in_index=2, channels=256, ignore_index=255, num_convs=1, concat_input=False, dropout_ratio=0.1, num_classes=6, norm_cfg=dict(type='BN', requires_grad=True), align_corners=False, loss_decode=dict( type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)), train_cfg=dict(), test_cfg=dict(mode='whole'))

dataset_type = 'PotsdamAllDataset'

dataset_type="PotsdamDataset" data_root = 'data/potsdam' img_norm_cfg = dict( mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) crop_size = (512, 512) train_pipeline = [ dict(type='LoadImageFromFile'),

dict(type='LoadAnnotations', reduce_zero_label=True),

dict(type='LoadAnnotationsReduceIgnoreIndex', reduce_zero_label=True, ignore_index=6),
dict(type='Resize', img_scale=(512, 512), ratio_range=(0.5, 2.0)),
dict(type='RandomCrop', crop_size=(512, 512), cat_max_ratio=0.75),
dict(type='RandomFlip', prob=0.5),
dict(type='PhotoMetricDistortion'),
dict(
    type='Normalize',
    mean=[123.675, 116.28, 103.53],
    std=[58.395, 57.12, 57.375],
    to_rgb=True),
dict(type='Pad', size=(512, 512), pad_val=0, seg_pad_val=255),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_semantic_seg'])

] test_pipeline = [ dict(type='LoadImageFromFile'), dict( type='MultiScaleFlipAug', img_scale=(512, 512), flip=False, transforms=[ dict(type='Resize', keep_ratio=True), dict(type='RandomFlip'), dict( type='Normalize', mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True), dict(type='ImageToTensor', keys=['img']), dict(type='Collect', keys=['img']) ]) ] data = dict( samples_per_gpu=8, workers_per_gpu=4, train=dict( type=dataset_type, data_root=data_root, img_dir='img_dir/train', ann_dir='ann_dir/train', pipeline=train_pipeline), val=dict( type=dataset_type, data_root=data_root, img_dir='img_dir/val', ann_dir='ann_dir/val', pipeline=test_pipeline), test=dict( type=dataset_type, data_root=data_root, img_dir='img_dir/val', ann_dir='ann_dir/val', pipeline=test_pipeline))


* CMID-Swin
```python
# model settings
# checkpoint_file = "./pretrained/CMID_Swin-B_bk_200ep.pth"
checkpoint_file = "./pretrained/CMID-Swin-B-millionAID.pth"
norm_cfg = dict(type='SyncBN', requires_grad=True)
backbone_norm_cfg = dict(type='LN', requires_grad=True)
model = dict(
    type='EncoderDecoder',
    pretrained=None,
    backbone=dict(
        type='SwinTransformer',
        pretrain_img_size=224,
        embed_dims=128,
        patch_size=4,
        window_size=7,
        mlp_ratio=4,
        depths=[2, 2, 18, 2],
        num_heads=[4, 8, 16, 32],
        strides=(4, 2, 2, 2),
        out_indices=(0, 1, 2, 3),
        qkv_bias=True,
        qk_scale=None,
        patch_norm=True,
        drop_rate=0.,
        attn_drop_rate=0.,
        drop_path_rate=0.3,
        use_abs_pos_embed=False,
        act_cfg=dict(type='GELU'),
        init_cfg=dict(type='Pretrained', checkpoint=checkpoint_file),
        norm_cfg=backbone_norm_cfg),
    decode_head=dict(
        type='UPerHead',
        in_channels=[128, 256, 512, 1024],
        in_index=[0, 1, 2, 3],
        pool_scales=(1, 2, 3, 6),
        channels=512,
        dropout_ratio=0.1,
        num_classes=6,
        ignore_index=255,
        norm_cfg=norm_cfg,
        align_corners=False,
        loss_decode=dict(
            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0)),
    auxiliary_head=dict(
        type='FCNHead',
        in_channels=512,
        in_index=2,
        channels=256,
        num_convs=1,
        concat_input=False,
        dropout_ratio=0.1,
        num_classes=6,
        ignore_index=255,
        norm_cfg=norm_cfg,
        align_corners=False,
        loss_decode=dict(
            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
    # model training and testing settings
    train_cfg=dict(),
    test_cfg=dict(mode='whole'))

log_config = dict(
    interval=50,
    hooks=[
        dict(type='TextLoggerHook', by_epoch=False),
        dict(type='TensorboardLoggerHook')
    ])
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
resume_from = None
workflow = [('train', 1)]
cudnn_benchmark = True
dataset_type = 'PotsdamDataset'
data_root = 'data/potsdam'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size = (512, 512)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='LoadAnnotationsReduceIgnoreIndex',
        reduce_zero_label=True,
        ignore_index=6),
    dict(type='Resize', img_scale=(512, 512), ratio_range=(0.5, 2.0)),
    dict(type='RandomCrop', crop_size=(512, 512), cat_max_ratio=0.75),
    dict(type='RandomFlip', prob=0.5),
    dict(type='PhotoMetricDistortion'),
    dict(
        type='Normalize',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        to_rgb=True),
    dict(type='Pad', size=(512, 512), pad_val=0, seg_pad_val=255),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg'])
]
val_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(512, 512),
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(
                type='Normalize',
                mean=[123.675, 116.28, 103.53],
                std=[58.395, 57.12, 57.375],
                to_rgb=True),
            dict(type='Pad', size_divisor=512),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img'])
        ])
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=None,
        img_ratios=[1.0],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(
                type='Normalize',
                mean=[123.675, 116.28, 103.53],
                std=[58.395, 57.12, 57.375],
                to_rgb=True),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img'])
        ])
]
data = dict(
    samples_per_gpu=8,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='img_dir/train',
        ann_dir='ann_dir/train',
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='img_dir/val',
        ann_dir='ann_dir/val',
        pipeline=test_pipeline),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='img_dir/val',
        ann_dir='ann_dir/val',
        pipeline=test_pipeline))

optimizer_config = None

optimizer = dict(
    type='AdamW',
    lr=0.00006,
    betas=(0.9, 0.999),
    weight_decay=0.01,
    paramwise_cfg=dict(
        custom_keys={
            'absolute_pos_embed': dict(decay_mult=0.),
            'relative_position_bias_table': dict(decay_mult=0.),
            'norm': dict(decay_mult=0.)
        }))

lr_config = dict(
    policy='CosineAnnealing',
    warmup='linear',
    warmup_iters=100,
    min_lr=0.000001,
    by_epoch=False)

checkpoint_config = dict(by_epoch=True, interval=1)
evaluation = dict(
    interval=5, metric=['mIoU', 'mFscore'], pre_eval=True, by_epoch=True)
runner = dict(type='EpochBasedRunner', max_epochs=50)
gpu_ids = [0]
auto_resume = False

And here are the log json files

CMID-ResNet

{
"config": "./configs/cmid_swin.py",
"metric": {
    "aAcc": 0.8137000000000001,
    "mIoU": 0.6962,
    "mAcc": 0.8009999999999999,
    "IoU.impervious_surface": 0.747699966430664,
    "IoU.building": 0.6880999755859375,
    "IoU.low_vegetation": 0.6165000152587891,
    "IoU.tree": 0.6458000183105469,
    "IoU.car": 0.7827999877929688,
    "Acc.impervious_surface": 0.8461000061035157,
    "Acc.building": 0.9302999877929687,
    "Acc.low_vegetation": 0.725,
    "Acc.tree": 0.6923000335693359,
    "Acc.car": 0.8113999938964844
}
}

CMID-Swin

{
"config": "./configs/cmid_resnet.py",
"metric": {
    "aAcc": 0.8151999999999999,
    "mIoU": 0.7041,
    "mAcc": 0.8096,
    "IoU.impervious_surface": 0.7501999664306641,
    "IoU.building": 0.6973999786376953,
    "IoU.low_vegetation": 0.6097000122070313,
    "IoU.tree": 0.6477999877929688,
    "IoU.car": 0.8154000091552734,
    "Acc.impervious_surface": 0.8262000274658203,
    "Acc.building": 0.9519999694824218,
    "Acc.low_vegetation": 0.729800033569336,
    "Acc.tree": 0.6966000366210937,
    "Acc.car": 0.8431999969482422
}
}

pUmpKin-Co commented 9 months ago

Hi~ @sskim0126 Try the following for reproducing similar results.

Please make sure pre-trained model is correctly loaded (You can monitor this using log file provied by mmsegmentation).
Replace or update the files of mmseg according to same file structures as newled updated folder SemanticSegmentation/mmseg/core/optimizers
Replace the optimizer of the old config:

ResNet50

```python optimizer = dict(constructor='TimmConstructor',type='Adan', lr=0.0020, weight_decay=0.02) optimizer_config = dict(type='Fp16OptimizerHook', distributed=False, grad_clip=dict(max_norm=5.0, norm_type=2)) ```

SwinB

```python optimizer = dict(constructor='TimmConstructor',type='Adan', lr=0.0003, weight_decay=0.02, filter_bias_and_bn=False, paramwise_cfg=dict( custom_keys={ 'absolute_pos_embed': dict(decay_mult=0.), 'relative_position_bias_table': dict(decay_mult=0.), 'norm': dict(decay_mult=0.)})) optimizer_config = dict(type='Fp16OptimizerHook', distributed=False) ```

Then run again!

For your reference, the following is my evaluation result after training 50 epochs (you can achieve the best result by adding --aug-test with mmseg test.py script. The following result didn't add --aug-test since I lost the added results :( ):

CMID-Swin-B

```json {"aAcc": 0.9299, "mIoU": 0.8736, "mAcc": 0.9295, "mFscore": 0.9312, "mPrecision": 0.9332, "mRecall": 0.9295, "mFWIoU": 0.8716, "IoU.impervious_surface": 0.8999, "IoU.building": 0.9483, "IoU.low_vegetation": 0.7963, "IoU.tree": 0.794, "IoU.car": 0.9294, "Acc.impervious_surface": 0.9377, "Acc.building": 0.975, "Acc.low_vegetation": 0.908, "Acc.tree": 0.8749, "Acc.car": 0.9521, "Freq.impervious_surface": 0.3279, "Freq.building": 0.2603, "Freq.low_vegetation": 0.2156, "Freq.tree": 0.1797, "Freq.car": 0.0166, "Fscore.impervious_surface": 0.9473, "Fscore.building": 0.9734, "Fscore.low_vegetation": 0.8866, "Fscore.tree": 0.8851, "Fscore.car": 0.9634, "Precision.impervious_surface": 0.9571, "Precision.building": 0.9719, "Precision.low_vegetation": 0.8662, "Precision.tree": 0.8956, "Precision.car": 0.975, "Recall.impervious_surface": 0.9377, "Recall.building": 0.975, "Recall.low_vegetation": 0.908, "Recall.tree": 0.8749, "Recall.car": 0.9521} ```

CMID-Resnet50

```json {"aAcc": 0.9277, "mIoU": 0.8704, "mAcc": 0.9282, "mFscore": 0.9293, "mPrecision": 0.9307, "mRecall": 0.9282, "mFWIoU": 0.8678, "IoU.impervious_surface": 0.8951, "IoU.building": 0.9476, "IoU.low_vegetation": 0.7916, "IoU.tree": 0.7879, "IoU.car": 0.9298, "Acc.impervious_surface": 0.9368, "Acc.building": 0.9745, "Acc.low_vegetation": 0.9073, "Acc.tree": 0.865, "Acc.car": 0.9575, "Freq.impervious_surface": 0.3279, "Freq.building": 0.2603, "Freq.low_vegetation": 0.2156, "Freq.tree": 0.1797, "Freq.car": 0.0166, "Fscore.impervious_surface": 0.9446, "Fscore.building": 0.9731, "Fscore.low_vegetation": 0.8837, "Fscore.tree": 0.8813, "Fscore.car": 0.9636, "Precision.impervious_surface": 0.9526, "Precision.building": 0.9717, "Precision.low_vegetation": 0.8612, "Precision.tree": 0.8983, "Precision.car": 0.9698, "Recall.impervious_surface": 0.9368, "Recall.building": 0.9745, "Recall.low_vegetation": 0.9073, "Recall.tree": 0.865, "Recall.car": 0.9575} ```

Finally, I have updated the instructions for semantic segmention tasks.

Thanks!

pUmpKin-Co commented 8 months ago

Closed as long periods of inactivity, feel free to reopen if there is any problem.

NJU-LHRS / official-CMID

Question About Discrepancy Between My Semantic Segmentation Results and Expected Performance #17

dataset_type = 'PotsdamAllDataset'

dict(type='LoadAnnotations', reduce_zero_label=True),