nuscenes_gs25600_solid.py

_base_ = [
    './_base_/misc.py',
    './_base_/model.py',
    './_base_/surroundocc.py'
]

# =========== data config ==============
input_shape = (1600, 864)
data_aug_conf = {
    "resize_lim": (1.0, 1.0),
    "final_dim": input_shape[::-1],
    "bot_pct_lim": (0.0, 0.0),
    "rot_lim": (0.0, 0.0),
    "H": 900,
    "W": 1600,
    "rand_flip": True,
}
val_dataset_config = dict(
    data_aug_conf=data_aug_conf
)
train_dataset_config = dict(
    data_aug_conf=data_aug_conf
)
# =========== misc config ==============
optimizer = dict(
    optimizer = dict(
        type="AdamW", lr=2e-4, weight_decay=0.01,
    ),
    paramwise_cfg=dict(
        custom_keys={
            'img_backbone': dict(lr_mult=0.1)}
    )
)
grad_max_norm = 35
# ========= model config ===============
loss = dict(
    type='MultiLoss',
    loss_cfgs=[
        dict(
            type='OccupancyLoss',
            weight=1.0,
            empty_label=17,
            num_classes=18,
            use_focal_loss=False,
            use_dice_loss=False,
            balance_cls_weight=True,
            multi_loss_weights=dict(
                loss_voxel_ce_weight=10.0,
                loss_voxel_lovasz_weight=1.0),
            use_sem_geo_scal_loss=False,
            use_lovasz_loss=True,
            lovasz_ignore=17,
            manual_class_weight=[
                1.01552756, 1.06897009, 1.30013094, 1.07253735, 0.94637502, 1.10087012,
                1.26960524, 1.06258364, 1.189019,   1.06217292, 1.00595144, 0.85706115,
                1.03923299, 0.90867526, 0.8936431,  0.85486129, 0.8527829,  0.5       ])
        ])

loss_input_convertion = dict(
    pred_occ="pred_occ",
    sampled_xyz="sampled_xyz",
    sampled_label="sampled_label",
    occ_mask="occ_mask"
)
# ========= model config ===============
embed_dims = 128
num_decoder = 4
num_single_frame_decoder = 1
pc_range = [-50.0, -50.0, -5.0, 50.0, 50.0, 3.0]
scale_range = [0.08, 0.64]
xyz_coordinate = 'cartesian'
phi_activation = 'sigmoid'
include_opa = True
load_from = 'ckpts/r101_dcn_fcos3d_pretrain.pth'
semantics = True
semantic_dim = 17

model = dict(
    img_backbone_out_indices=[0, 1, 2, 3],
    img_backbone=dict(
        _delete_=True,
        type='ResNet',
        depth=101,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        norm_cfg=dict(type='BN2d', requires_grad=False),
        norm_eval=True,
        style='caffe',
        with_cp = True,
        dcn=dict(type='DCNv2', deform_groups=1, fallback_on_stride=False), # original DCNv2 will print log when perform load_state_dict
        stage_with_dcn=(False, False, True, True)),
    img_neck=dict(
        start_level=1),
    lifter=dict(
        type='GaussianLifter',
        num_anchor=25600,
        embed_dims=embed_dims,
        anchor_grad=True,
        feat_grad=False,
        phi_activation=phi_activation,
        semantics=semantics,
        semantic_dim=semantic_dim,
        include_opa=include_opa,
    ),
    encoder=dict(
        type='GaussianOccEncoder',
        anchor_encoder=dict(
            type='SparseGaussian3DEncoder',
            embed_dims=embed_dims, 
            include_opa=include_opa,
            semantics=semantics,
            semantic_dim=semantic_dim
        ),
        norm_layer=dict(type="LN", normalized_shape=embed_dims),
        ffn=dict(
            type="AsymmetricFFN",
            in_channels=embed_dims * 2,
            embed_dims=embed_dims,
            feedforward_channels=embed_dims * 4,
        ),
        deformable_model=dict(
            embed_dims=embed_dims,
            kps_generator=dict(
                embed_dims=embed_dims,
                phi_activation=phi_activation,
                xyz_coordinate=xyz_coordinate,
                num_learnable_pts=2,
                pc_range=pc_range,
                scale_range=scale_range
            ),
        ),
        refine_layer=dict(
            type='SparseGaussian3DRefinementModule',
            embed_dims=embed_dims,
            pc_range=pc_range,
            scale_range=scale_range,
            restrict_xyz=True,
            unit_xyz=[4.0, 4.0, 1.0],
            refine_manual=[0, 1, 2],
            phi_activation=phi_activation,
            semantics=semantics,
            semantic_dim=semantic_dim,
            include_opa=include_opa,
            xyz_coordinate=xyz_coordinate,
            semantics_activation='softplus',
        ),
        spconv_layer=dict(
            _delete_=True,
            type="SparseConv3D",
            in_channels=embed_dims,
            embed_channels=embed_dims,
            pc_range=pc_range,
            grid_size=[0.5, 0.5, 0.5],
            phi_activation=phi_activation,
            xyz_coordinate=xyz_coordinate,
            use_out_proj=True,
        ),
        num_decoder=num_decoder,
        num_single_frame_decoder=num_single_frame_decoder,
        operation_order=[
            "deformable",
            "ffn",
            "norm",
            "refine",
        ] * num_single_frame_decoder + [
            "spconv",
            "norm",
            "deformable",
            "ffn",
            "norm",
            "refine",
        ] * (num_decoder - num_single_frame_decoder),
    ),
    head=dict(
        type='GaussianHead',
        apply_loss_type='random_1',
        num_classes=semantic_dim + 1,
        empty_args=dict(
            _delete_=True,
            mean=[0, 0, -1.0],
            scale=[100, 100, 8.0],
        ),
        with_empty=True,
        cuda_kwargs=dict(
            _delete_=True,
            scale_multiplier=3,
            H=200, W=200, D=16,
            pc_min=[-50.0, -50.0, -5.0],
            grid_size=0.5),
    )
)