PTv1, shapenet, part segmentation

[Jae-Seung-Jeon]

Hi, Thanks to your contributes.
I want to train PTv1, dataset=shapenet, for part segmentation.
Could you give me a running script for this please?
Thank you :)

[Gofinge]

Sure! The following ShapeNet config for PTv1 was from an older version of our codebase, and you need to migrate it to the released codebase. Also, I did not tune PTv1 on the ShapeNet part segmentation dataset, and you might need to tune the augmentation and scheduler setting to achieve better performance.

_base_ = [
    '../_base_/datasets/shapenet_part.py',
    '../_base_/schedulers/multi-step_sgd.py',
    '../_base_/tests/part_segmentation.py',
    '../_base_/default_runtime.py'
]

batch_size = 32
batch_size_val = 8
metric = "cat_mIoU"
# enable_amp = True

train_gpu = [2,3]

epochs = 100
start_epoch = 0
optimizer = dict(type='SGD', lr=0.5, momentum=0.9, weight_decay=0.0001, nesterov=True)
scheduler = dict(type='MultiStepLR', milestones=[epochs * 0.6, epochs * 0.8], steps_per_epoch=1, gamma=0.1)

model = dict(
    type='PointTransformerV2-PartSeg50',
    num_shape_classes=16,
    in_channels=6,
    num_classes=50
)


# dataset settings
dataset_type = "ShapeNetPartDataset"
data_root = "data/shapenetcore_partanno_segmentation_benchmark_v0_normal"
cache_data = False
names = ["Airplane_{}".format(i) for i in range(4)] + \
        ["Bag_{}".format(i) for i in range(2)] + \
        ["Cap_{}".format(i) for i in range(2)] + \
        ["Car_{}".format(i) for i in range(4)] + \
        ["Chair_{}".format(i) for i in range(4)] + \
        ["Earphone_{}".format(i) for i in range(3)] + \
        ["Guitar_{}".format(i) for i in range(3)] + \
        ["Knife_{}".format(i) for i in range(2)] + \
        ["Lamp_{}".format(i) for i in range(4)] + \
        ["Laptop_{}".format(i) for i in range(2)] + \
        ["Motorbike_{}".format(i) for i in range(6)] + \
        ["Mug_{}".format(i) for i in range(2)] + \
        ["Pistol_{}".format(i) for i in range(3)] + \
        ["Rocket_{}".format(i) for i in range(3)] + \
        ["Skateboard_{}".format(i) for i in range(3)] + \
        ["Table_{}".format(i) for i in range(3)]


data = dict(
    num_classes=50,
    ignore_label=-1,  # dummy ignore
    names=names,
    train=dict(
        type=dataset_type,
        split=["train", "val"],
        data_root=data_root,
        transform=[
            dict(type="NormalizeCoord"),
            # dict(type="CenterShift", apply_z=True),
            # dict(type="RandomRotate", angle=[-1, 1], axis='z', center=[0, 0, 0], p=0.5),
            # dict(type="RandomRotate", angle=[-1 / 24, 1 / 24], axis='x', p=0.5),
            # dict(type="RandomRotate", angle=[-1 / 24, 1 / 24], axis='y', p=0.5),
            # dict(type="RandomScale", scale=[0.9, 1.1]),
            # dict(type="RandomFlip", p=0.5),
            # dict(type="RandomJitter", sigma=0.005, clip=0.02),
            # dict(type="RandomShift", shift=[0.2, 0.2, 0.2]),
            # dict(type="ElasticDistortion", distortion_params=[[0.2, 0.4], [0.8, 1.6]]),

            # dict(type="Voxelize", voxel_size=0.01, hash_type='fnv', mode='train'),
            # dict(type="SphereCrop", point_max=2500, mode='random'),
            dict(type="ShufflePoint"),
            dict(type="ToTensor"),
            dict(type="Collect", keys=("coord", "cls_token", "label"), feat_keys=("coord", "norm"))
        ],
        loop=2,
        test_mode=False,
    ),

    val=dict(
        type=dataset_type,
        split="test",
        data_root=data_root,
        transform=[
            dict(type="NormalizeCoord"),
            dict(type="ToTensor"),
            dict(type="Collect", keys=("coord", "cls_token", "label"), feat_keys=("coord", "norm"))
        ],
        loop=1,
        test_mode=False,
    ),

    test=dict(
        type=dataset_type,
        split="test",
        data_root=data_root,
        transform=[
            dict(type="NormalizeCoord"),
            # dict(type="CenterShift", apply_z=True),
        ],
        loop=1,
        test_mode=True,
        test_cfg=dict(
            post_transform=[
                dict(type="ToTensor"),
                dict(type="Collect", keys=("coord", "cls_token"), feat_keys=("coord", "norm"))
            ],
            aug_transform=[
                [dict(type="RandomShift2", shift=((0, 0), (0, 0), (0, 0)))]
            ]
        )
    ),
)

criteria = [
    dict(type="CrossEntropyLoss",
         loss_weight=1.0,
         ignore_index=data["ignore_label"])
]

[Jae-Seung-Jeon]

Thank you @Gofinge !!
I'm not sure I did it right,
I revised [/config/base/datasets/shapenet_part.py ] to the one you gave me.

And, I want to know the command by [sh scripts/train.sh ...] or [python tools/train.py ...]
I tried many different ways, but I failed. TT
Could you give me some advice..?

[Gofinge]

Hi, /config/base/datasets/shapenet_part.py is just a base config, not a task config. It doesn't contain the necessary components to start a task. You can modify it from a task config located in the dataset folder. For example, you can copy the ptv1 config in the ScanNet folder to the ShapeNetPart folder (mkdir configs/shapenet_part)， then follow the old config which I provided to replace the config in the copied one.

[zongshun21]

I love people who have achieved so much and helped others, and the world will be wonderful because of you.