File size: 881 Bytes
2cda886 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 | import torch.nn as nn
from torchvision.models import efficientnet_v2_l, efficientnet_v2_s
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
class EfficientNetEncoder(ModelMixin, ConfigMixin):
@register_to_config
def __init__(self, c_latent=16, c_cond=1280, effnet="efficientnet_v2_s"):
super().__init__()
if effnet == "efficientnet_v2_s":
self.backbone = efficientnet_v2_s(weights="DEFAULT").features
else:
self.backbone = efficientnet_v2_l(weights="DEFAULT").features
self.mapper = nn.Sequential(
nn.Conv2d(c_cond, c_latent, kernel_size=1, bias=False),
nn.BatchNorm2d(c_latent), # then normalize them to have mean 0 and std 1
)
def forward(self, x):
return self.mapper(self.backbone(x))
|