timmfrv2.py

import torch.nn as nn
import timm


class TimmFRWrapperV2(nn.Module):
    """
    Wraps timm model
    """

    def __init__(self, model_name="edgenext_x_small", featdim=512, batchnorm=False):
        super().__init__()
        self.featdim = featdim
        self.model_name = model_name

        self.model = timm.create_model(self.model_name)
        self.model.reset_classifier(self.featdim)

    def forward(self, x):
        x = self.model(x)
        return x


class LoRaLin(nn.Module):
    def __init__(self, in_features, out_features, rank, bias=True):
        super(LoRaLin, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.rank = rank
        self.linear1 = nn.Linear(in_features, rank, bias=False)
        self.linear2 = nn.Linear(rank, out_features, bias=bias)

    def forward(self, input):
        x = self.linear1(input)
        x = self.linear2(x)
        return x


def replace_linear_with_lowrank_recursive_2(model, rank_ratio=0.2):
    for name, module in model.named_children():
        if isinstance(module, nn.Linear) and "head" not in name:
            in_features = module.in_features
            out_features = module.out_features
            rank = max(2, int(min(in_features, out_features) * rank_ratio))
            bias = False
            if module.bias is not None:
                bias = True
            lowrank_module = LoRaLin(in_features, out_features, rank, bias)

            setattr(model, name, lowrank_module)
        else:
            replace_linear_with_lowrank_recursive_2(module, rank_ratio)


def replace_linear_with_lowrank_2(model, rank_ratio=0.2):
    replace_linear_with_lowrank_recursive_2(model, rank_ratio)
    return model


model_configs = {
    "edgeface_base": {
        "repo": "idiap/EdgeFace-Base",
        "filename": "edgeface_base.pt",
        "timm_model": "edgenext_base",
        "post_setup": lambda x: x,
    },
    "edgeface_s_gamma_05": {
        "repo": "idiap/EdgeFace-S-GAMMA",
        "filename": "edgeface_s_gamma_05.pt",
        "timm_model": "edgenext_small",
        "post_setup": lambda x: replace_linear_with_lowrank_2(x, rank_ratio=0.5),
    },
    "edgeface_xs_gamma_06": {
        "repo": "idiap/EdgeFace-XS-GAMMA",
        "filename": "edgeface_xs_gamma_06.pt",
        "timm_model": "edgenext_x_small",
        "post_setup": lambda x: replace_linear_with_lowrank_2(x, rank_ratio=0.6),
    },
    "edgeface_xxs": {
        "repo": "idiap/EdgeFace-XXS",
        "filename": "edgeface_xxs.pt",
        "timm_model": "edgenext_xx_small",
        "post_setup": lambda x: x,
    },
}