Spaces:
Sleeping
Sleeping
File size: 2,085 Bytes
281df87 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 |
import torch
import torch.nn as nn
from ldm.modules.attention import BasicTransformerBlock
from ldm.modules.diffusionmodules.util import checkpoint, FourierEmbedder
import torch.nn.functional as F
from ..attention import SelfAttention, FeedForward
from .convnext import convnext_tiny
class PositionNet(nn.Module):
def __init__(self, resize_input=448, out_dim=768):
super().__init__()
self.resize_input = resize_input
self.down_factor = 32 # determined by the res50 backbone
self.out_dim = out_dim
assert self.resize_input % self.down_factor == 0
self.convnext_tiny_backbone = convnext_tiny(pretrained=True)
self.num_tokens = (self.resize_input // self.down_factor) ** 2
convnext_feature_dim = 768
self.pos_embedding = nn.Parameter(torch.empty(1, self.num_tokens, convnext_feature_dim).normal_(std=0.02)) # from BERT
self.linears = nn.Sequential(
nn.Linear( convnext_feature_dim, 512),
nn.SiLU(),
nn.Linear( 512, 512),
nn.SiLU(),
nn.Linear(512, out_dim),
)
self.null_feature = torch.nn.Parameter(torch.zeros([convnext_feature_dim]))
def forward(self, hed_edge, mask):
B = hed_edge.shape[0]
# token from edge map
hed_edge = torch.nn.functional.interpolate(hed_edge, self.resize_input)
hed_edge_feature = self.convnext_tiny_backbone(hed_edge)
objs = hed_edge_feature.reshape(B, -1, self.num_tokens)
objs = objs.permute(0, 2, 1) # N*Num_tokens*dim
# expand null token
null_objs = self.null_feature.view(1,1,-1)
null_objs = null_objs.repeat(B,self.num_tokens,1)
# mask replacing
mask = mask.view(-1,1,1)
objs = objs*mask + null_objs*(1-mask)
# add pos
objs = objs + self.pos_embedding
# fuse them
objs = self.linears(objs)
assert objs.shape == torch.Size([B,self.num_tokens,self.out_dim])
return objs
|