scenedino/models/prediction_heads/__init__.py

from .multi_view_head import (
    MultiViewHead,
    SimpleMultiViewHead,
    MultiViewHead2,
    MultiViewHead3,
    make_attn_layers,
)

from .resnetfc import ResnetFC
from .mlp import ImplicitNet, make_embedding_encoder


# TODO: reorganize this directory
def make_head(conf, d_in: int, d_out: int):
    head_type = conf.get("type", "resnet")

    if head_type == "mlp":
        head = ImplicitNet.from_conf(conf["args"], d_in, d_out)
    elif head_type == "resnet":
        head = ResnetFC.from_conf(conf["args"], d_in, d_out)
    elif head_type == "MultiViewHead":
        head = MultiViewHead.from_conf(conf["args"], d_in, d_out)
    elif head_type == "SimpleMultiViewHead":
        head = SimpleMultiViewHead.from_conf(conf["args"], d_in, d_out)
    elif head_type == "MultiViewHead2":
        head = MultiViewHead2.from_conf(conf["args"], d_in, d_out)
    elif head_type == "MultiViewHead3":
        head = MultiViewHead3.from_conf(conf["args"], d_in, d_out)
    ## For baseline comparison
    # elif head_type == "IBRNet":
    #     head = MultiViewHead.from_conf(conf["args"], d_in, d_out)
    # elif head_type == "NeuRay":
    #     head = MultiViewHead.from_conf(conf["args"], d_in, d_out)
    # elif head_type == "GeoNeRF":
    #     head = MultiViewHead.from_conf(conf["args"], d_in, d_out)
    # elif head_type == "PixelNeRF":
    #     head = MultiViewHead.from_conf(conf["args"], d_in, d_out)

    else:
        raise NotImplementedError("Unsupported Head type")
    if conf.get("freeze", False):
        for param in head.parameters():
            param.requires_grad = False
        print(f"__freezing the {conf['name']} for knowledge distillation.")
        # else:
        #     print("__No freezing heads during training.")
    return head