File size: 4,841 Bytes
961cf0c | 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 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 | #!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Author: André Pacheco
E-mail: pacheco.comp@gmail.com
This file implements the Context Guided Cell (GCell)
and a full MetaNet + ResNet-50 model.
Paper:
Fusing Metadata and Dermoscopy Images for Skin Disease Diagnosis
IEEE Journal of Biomedical and Health Informatics, 2020
https://ieeexplore.ieee.org/document/9098645
"""
import timm
import torch
import torch.nn as nn
import torch.nn.functional as F
# =====================================================
# MetaNet block (Context Guided Cell)
# =====================================================
class MetaNet(nn.Module):
"""
Metadata-driven channel attention (MetaNet / GCell)
metadata (B, meta_dim) → (B, C, 1, 1)
feat_maps (B, C, H, W) → gated feature maps
"""
def __init__(self, in_channels: int, middle_channels: int, out_channels: int):
super().__init__()
self.metanet = nn.Sequential(
nn.Conv2d(in_channels, middle_channels, kernel_size=1),
nn.ReLU(inplace=True),
nn.Conv2d(middle_channels, out_channels, kernel_size=1),
nn.Sigmoid()
)
def forward(self, feat_maps: torch.Tensor, metadata: torch.Tensor) -> torch.Tensor:
"""
feat_maps: (B, C, H, W)
metadata: (B, meta_dim)
"""
m = metadata.unsqueeze(-1).unsqueeze(-1) # (B, meta_dim, 1, 1)
attn = self.metanet(m) # (B, C, 1, 1)
return feat_maps * attn
# =====================================================
# MetaNet + ResNet-50 model
# =====================================================
class MetaNetModel(nn.Module):
"""
MetaNet + ResNet-50 (faithful to IEEE JBHI paper)
"""
def __init__(
self,
meta_dim: int,
num_classes: int = 6,
dropout_fraction: float = 0.3,
image_encoder: str = "resnet50",
pretrained: bool = True,
unfreeze_weights: bool = False
):
super().__init__()
self.meta_dim = meta_dim
self.num_classes = num_classes
self.dropout_fraction = dropout_fraction
self.image_encoder = image_encoder
self.pretrained = pretrained
self.unfreeze_weights = unfreeze_weights
# =====================================================
# 1) CNN backbone (conv features only)
# =====================================================
self.backbone = timm.create_model(
self.image_encoder,
pretrained=self.pretrained,
num_classes=0, # ❗ remove FC
global_pool="" # ❗ remove GAP → retorna (B,C,H,W)
)
self.feat_dim = self.backbone.num_features # 2048 for resnet50
if not self.unfreeze_weights:
for p in self.backbone.parameters():
p.requires_grad = False
# =====================================================
# 2) MetaNet attention (metadata → channel gates)
# =====================================================
self.metanet = MetaNet(
in_channels=self.meta_dim,
middle_channels=128,
out_channels=self.feat_dim
)
# =====================================================
# 3) Classifier (after GAP)
# =====================================================
self.classifier = self.fc_mlp_module(self.feat_dim)
# -----------------------------------------------------
# MLP classifier (stronger than single FC)
# -----------------------------------------------------
def fc_mlp_module(self, input_dim: int) -> nn.Module:
return nn.Sequential(
nn.Linear(input_dim, input_dim),
nn.LayerNorm(input_dim),
nn.ReLU(inplace=True),
nn.Dropout(self.dropout_fraction),
nn.Linear(input_dim, input_dim // 2),
nn.LayerNorm(input_dim // 2),
nn.ReLU(inplace=True),
nn.Dropout(self.dropout_fraction),
nn.Linear(input_dim // 2, self.num_classes)
)
# =====================================================
# Forward
# =====================================================
def forward(self, image: torch.Tensor, metadata: torch.Tensor) -> torch.Tensor:
"""
image: (B, 3, 224, 224)
metadata: (B, meta_dim)
"""
# 1) CNN feature maps
feat_maps = self.backbone(image) # (B, 2048, H, W)
# 2) Metadata-guided channel attention
feat_maps = self.metanet(feat_maps, metadata)
# 3) Global Average Pooling
pooled = F.adaptive_avg_pool2d(feat_maps, 1).flatten(1) # (B, 2048)
# 4) Classification
logits = self.classifier(pooled)
return logits
|