310 lines
10 KiB
Python
310 lines
10 KiB
Python
import torch
|
|
from torch import nn
|
|
from torchvision import models
|
|
|
|
|
|
# 无融合模型
|
|
class MOAFNoFusion(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
shuff = models.shufflenet_v2_x0_5(weights="DEFAULT")
|
|
self.features = nn.Sequential(
|
|
shuff.conv1, shuff.maxpool,
|
|
shuff.stage2, shuff.stage3,
|
|
shuff.stage4, shuff.conv5
|
|
)
|
|
self.avgpool = nn.AdaptiveAvgPool2d(1)
|
|
self.regressor = nn.Sequential(
|
|
nn.Flatten(),
|
|
nn.Linear(1024, 128),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(128, 1)
|
|
)
|
|
|
|
def forward(self, image, params):
|
|
x = self.features(image)
|
|
x = self.avgpool(x)
|
|
x = self.regressor(x)
|
|
return x
|
|
|
|
|
|
# 参数嵌入模型
|
|
class ParamEmbedding(nn.Module):
|
|
def __init__(self, in_dim=3, hidden_dim=64, out_dim=128):
|
|
super().__init__()
|
|
self.embedding = nn.Sequential(
|
|
nn.Linear(in_dim, hidden_dim),
|
|
nn.ReLU(),
|
|
nn.Linear(hidden_dim, out_dim),
|
|
nn.ReLU(),
|
|
nn.LayerNorm(out_dim)
|
|
)
|
|
|
|
def forward(self, params):
|
|
# min-max 归一化参数
|
|
normalized_params = torch.stack([
|
|
(params[:, 0] - 10.0) / 90.0,
|
|
params[:, 1] / 1.25,
|
|
(params[:, 2] - 1.0) / 0.5
|
|
], dim=1)
|
|
return self.embedding(normalized_params)
|
|
|
|
|
|
# FiLM 融合块
|
|
class FiLMBlock(nn.Module):
|
|
def __init__(self, param_emb_dim=128, feat_channels=128):
|
|
super().__init__()
|
|
self.gamma_gen = nn.Linear(param_emb_dim, feat_channels)
|
|
self.beta_gen = nn.Linear(param_emb_dim, feat_channels)
|
|
|
|
def forward(self, feature_map, param_emb):
|
|
gamma = self.gamma_gen(param_emb).unsqueeze(-1).unsqueeze(-1)
|
|
beta = self.beta_gen(param_emb).unsqueeze(-1).unsqueeze(-1)
|
|
return feature_map * (1.0 + gamma) + beta
|
|
|
|
|
|
# 通道交叉注意力融合块
|
|
class ChannelCrossAttention(nn.Module):
|
|
def __init__(self, param_emb_dim=128, feat_channels=128):
|
|
super().__init__()
|
|
self.param_emb_dim = param_emb_dim
|
|
self.feat_channels = feat_channels
|
|
self.hidden_dim = max(16, self.feat_channels // 4)
|
|
|
|
self.global_pool = nn.AdaptiveAvgPool2d(1)
|
|
|
|
self.bottle_neck = nn.Sequential(
|
|
nn.Linear(self.param_emb_dim + self.feat_channels, self.hidden_dim),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(self.hidden_dim, self.feat_channels),
|
|
nn.Sigmoid()
|
|
)
|
|
|
|
def forward(self, feature_map, param_emb):
|
|
b, c, h, w = feature_map.shape
|
|
pooled = self.global_pool(feature_map).view(b, c)
|
|
cat = torch.cat([pooled, param_emb], dim=1)
|
|
weights = self.bottle_neck(cat)
|
|
weights4d = weights.view(b, c, 1, 1)
|
|
return feature_map * (1.0 + weights4d)
|
|
|
|
|
|
# SE块
|
|
class SEBlock(nn.Module):
|
|
def __init__(self, feat_channels=128):
|
|
super().__init__()
|
|
self.feat_channels = feat_channels
|
|
self.hidden_dim = max(16, self.feat_channels // 4)
|
|
|
|
self.global_pool = nn.AdaptiveAvgPool2d(1)
|
|
|
|
self.bottle_neck = nn.Sequential(
|
|
nn.Linear(self.feat_channels, self.hidden_dim),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(self.hidden_dim, self.feat_channels),
|
|
nn.Sigmoid()
|
|
)
|
|
|
|
def forward(self, feature_map):
|
|
b, c, h, w = feature_map.shape
|
|
pooled = self.global_pool(feature_map).view(b, c)
|
|
weights = self.bottle_neck(pooled)
|
|
weights4d = weights.view(b, c, 1, 1)
|
|
return feature_map * (1.0 + weights4d)
|
|
|
|
|
|
# 仅返回特征图的恒等变换
|
|
class FusionIdentity(nn.Module):
|
|
def forward(self, feature_map, param_emb):
|
|
return feature_map
|
|
|
|
|
|
# 使用 FiLM 融合的模型
|
|
class MOAFWithFiLM(nn.Module):
|
|
def __init__(self, fusion_level=None):
|
|
super().__init__()
|
|
shuff = models.shufflenet_v2_x0_5(weights="DEFAULT")
|
|
self.cbrm = nn.Sequential(shuff.conv1, shuff.maxpool)
|
|
self.stage2 = shuff.stage2
|
|
self.stage3 = shuff.stage3
|
|
self.stage4 = shuff.stage4
|
|
self.cbr2 = shuff.conv5
|
|
|
|
self.param_embedding = ParamEmbedding()
|
|
|
|
if fusion_level is None:
|
|
fusion_level = [2]
|
|
|
|
self.film_block0 = FiLMBlock(feat_channels=48) if 0 in fusion_level else FusionIdentity()
|
|
self.film_block1 = FiLMBlock(feat_channels=96) if 1 in fusion_level else FusionIdentity()
|
|
self.film_block2 = FiLMBlock(feat_channels=192) if 2 in fusion_level else FusionIdentity()
|
|
|
|
self.avgpool = nn.AdaptiveAvgPool2d(1)
|
|
self.regressor = nn.Sequential(
|
|
nn.Flatten(),
|
|
nn.Linear(1024, 128),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(128, 1)
|
|
)
|
|
|
|
def forward(self, image, params):
|
|
x = self.cbrm(image)
|
|
x = self.stage2(x)
|
|
param_emb = self.param_embedding(params)
|
|
x = self.film_block0(x, param_emb)
|
|
x = self.stage3(x)
|
|
x = self.film_block1(x, param_emb)
|
|
x = self.stage4(x)
|
|
x = self.film_block2(x, param_emb)
|
|
x = self.cbr2(x)
|
|
x = self.avgpool(x)
|
|
x = self.regressor(x)
|
|
return x
|
|
|
|
|
|
# 使用交叉注意力融合的模型
|
|
class MOAFWithChannelCrossAttention(nn.Module):
|
|
def __init__(self, fusion_level=None):
|
|
super().__init__()
|
|
shuff = models.shufflenet_v2_x0_5(weights="DEFAULT")
|
|
self.cbrm = nn.Sequential(shuff.conv1, shuff.maxpool)
|
|
self.stage2 = shuff.stage2
|
|
self.stage3 = shuff.stage3
|
|
self.stage4 = shuff.stage4
|
|
self.cbr2 = shuff.conv5
|
|
|
|
self.param_embedding = ParamEmbedding()
|
|
|
|
if fusion_level is None:
|
|
fusion_level = [2]
|
|
|
|
self.cca_block0 = ChannelCrossAttention(feat_channels=48) if 0 in fusion_level else FusionIdentity()
|
|
self.cca_block1 = ChannelCrossAttention(feat_channels=96) if 1 in fusion_level else FusionIdentity()
|
|
self.cca_block2 = ChannelCrossAttention(feat_channels=192) if 2 in fusion_level else FusionIdentity()
|
|
|
|
self.avgpool = nn.AdaptiveAvgPool2d(1)
|
|
self.regressor = nn.Sequential(
|
|
nn.Flatten(),
|
|
nn.Linear(1024, 128),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(128, 1)
|
|
)
|
|
|
|
def forward(self, image, params):
|
|
x = self.cbrm(image)
|
|
x = self.stage2(x)
|
|
param_emb = self.param_embedding(params)
|
|
x = self.cca_block0(x, param_emb)
|
|
x = self.stage3(x)
|
|
x = self.cca_block1(x, param_emb)
|
|
x = self.stage4(x)
|
|
x = self.cca_block2(x, param_emb)
|
|
x = self.cbr2(x)
|
|
x = self.avgpool(x)
|
|
x = self.regressor(x)
|
|
return x
|
|
|
|
|
|
# 使用 SE 块但无融合的模型
|
|
class MOAFWithSE(nn.Module):
|
|
def __init__(self, fusion_level=None):
|
|
super().__init__()
|
|
shuff = models.shufflenet_v2_x0_5(weights="DEFAULT")
|
|
self.cbrm = nn.Sequential(shuff.conv1, shuff.maxpool)
|
|
self.stage2 = shuff.stage2
|
|
self.stage3 = shuff.stage3
|
|
self.stage4 = shuff.stage4
|
|
self.cbr2 = shuff.conv5
|
|
|
|
if fusion_level is None:
|
|
fusion_level = [0]
|
|
|
|
self.se_block0 = SEBlock(feat_channels=48) if 0 in fusion_level else nn.Identity()
|
|
self.se_block1 = SEBlock(feat_channels=96) if 1 in fusion_level else nn.Identity()
|
|
self.se_block2 = SEBlock(feat_channels=192) if 2 in fusion_level else nn.Identity()
|
|
|
|
self.avgpool = nn.AdaptiveAvgPool2d(1)
|
|
self.regressor = nn.Sequential(
|
|
nn.Flatten(),
|
|
nn.Linear(1024, 128),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(128, 1)
|
|
)
|
|
|
|
def forward(self, image, params):
|
|
x = self.cbrm(image)
|
|
x = self.stage2(x)
|
|
x = self.se_block0(x)
|
|
x = self.stage3(x)
|
|
x = self.se_block1(x)
|
|
x = self.stage4(x)
|
|
x = self.se_block2(x)
|
|
x = self.cbr2(x)
|
|
x = self.avgpool(x)
|
|
x = self.regressor(x)
|
|
return x
|
|
|
|
# 多回归头模型
|
|
class MOAFWithMMLP(nn.Module):
|
|
def __init__(self, num_lenses=7):
|
|
super().__init__()
|
|
shuff = models.shufflenet_v2_x0_5(weights="DEFAULT")
|
|
self.features = nn.Sequential(
|
|
shuff.conv1, shuff.maxpool,
|
|
shuff.stage2, shuff.stage3,
|
|
shuff.stage4, shuff.conv5
|
|
)
|
|
self.avgpool = nn.AdaptiveAvgPool2d(1)
|
|
self.num_lenses = num_lenses
|
|
self.regressors = nn.ModuleList([
|
|
self._create_regressor_head() for _ in range(num_lenses)
|
|
])
|
|
|
|
# 注册物镜参数基准张量（不可学习）
|
|
lens_params_base = torch.tensor([
|
|
[10, 0.25, 1.0000], # obj1
|
|
[10, 0.30, 1.0000], # obj2
|
|
[20, 0.70, 1.0000], # obj3
|
|
[20, 0.80, 1.0000], # obj4
|
|
[40, 0.65, 1.0000], # obj5
|
|
[100, 0.80, 1.0000], # obj6
|
|
[100, 1.25, 1.4730] # obj7
|
|
], dtype=torch.float32)
|
|
self.register_buffer('lens_params_base', lens_params_base)
|
|
|
|
def _create_regressor_head(self):
|
|
return nn.Sequential(
|
|
nn.Flatten(),
|
|
nn.Linear(1024, 128),
|
|
nn.ReLU(inplace=True),
|
|
nn.Linear(128, 1)
|
|
)
|
|
|
|
def _find_lens_id(self, params):
|
|
# 扩展维度以进行广播计算
|
|
params_expanded = params.unsqueeze(1) # [batch_size, 1, 3]
|
|
base_expanded = self.lens_params_base.unsqueeze(0) # [1, 7, 3]
|
|
|
|
# 计算欧氏距离（完全向量化）
|
|
distances = torch.sqrt(torch.sum((params_expanded - base_expanded) ** 2, dim=2)) # [batch_size, 7]
|
|
|
|
# 找到最小距离的索引
|
|
lens_ids = torch.argmin(distances, dim=1) # [batch_size]
|
|
return lens_ids
|
|
|
|
def forward(self, image, params):
|
|
x = self.features(image)
|
|
x = self.avgpool(x)
|
|
|
|
lens_ids = self._find_lens_id(params)
|
|
lens_ids = torch.clamp(lens_ids, 0, self.num_lenses - 1)
|
|
batch_size = x.size(0)
|
|
all_outputs = []
|
|
for i in range(batch_size):
|
|
head_idx = lens_ids[i]
|
|
output = self.regressors[head_idx](x[i].unsqueeze(0))
|
|
all_outputs.append(output)
|
|
|
|
return torch.cat(all_outputs, dim=0)
|
|
|