adelelsayed1991

Upload folder using huggingface_hub

abd02e7 verified 11 months ago

13.2 kB

import torch
import torch.nn as nn
import torch.nn.functional as F
import math

from models.mae import MaskedAutoEncoder
from models.densenet import DenseNet

class AttentionPool(nn.Module):
    def __init__(self, dim=768, embed_dim=2048, num_heads=8):
        super().__init__()
        self.query = nn.Parameter(torch.randn(1, 1, dim))
        self.attn = nn.MultiheadAttention(embed_dim=dim, num_heads=num_heads, batch_first=True)
        self.proj = nn.Linear(dim, embed_dim)

    def forward(self, x):  # x: (B, 576, 768)
        B = x.size(0)
        q = self.query.expand(B, -1, -1)   # (B, 1, 768)
        attn_out, _ = self.attn(q, x, x)   # (B, 1, 768)
        return self.proj(attn_out.squeeze(1))  # (B, 2048)

class CrossAttentionBlock(nn.Module):
    """
    Cross-attention: Query tokens attend to Key/Value tokens from another modality.
    """
    def __init__(self, dim_q, dim_kv, num_heads=8, dropout=0.1, proj_dim=None):
        super().__init__()
        self.proj_dim = proj_dim or dim_q
        self.num_heads = num_heads
        self.head_dim = self.proj_dim // num_heads
        self.scale = self.head_dim ** -0.5

        self.q_proj = nn.Linear(dim_q, self.proj_dim)
        self.k_proj = nn.Linear(dim_kv, self.proj_dim)
        self.v_proj = nn.Linear(dim_kv, self.proj_dim)
        self.out_proj = nn.Linear(self.proj_dim, dim_q)

        self.dropout = nn.Dropout(dropout)
        self.norm_q = nn.LayerNorm(dim_q)
        self.norm_kv = nn.LayerNorm(dim_kv)

    def forward(self, query, key_value):
        B, N_q, _ = query.shape
        N_kv = key_value.shape[1]

        q = self.norm_q(query)
        kv = self.norm_kv(key_value)

        Q = self.q_proj(q).view(B, N_q, self.num_heads, self.head_dim).transpose(1, 2)
        K = self.k_proj(kv).view(B, N_kv, self.num_heads, self.head_dim).transpose(1, 2)
        V = self.v_proj(kv).view(B, N_kv, self.num_heads, self.head_dim).transpose(1, 2)

        attn = (Q @ K.transpose(-2, -1)) * self.scale
        attn = F.softmax(attn, dim=-1)
        attn = self.dropout(attn)

        out = (attn @ V).transpose(1, 2).reshape(B, N_q, self.proj_dim)
        out = self.out_proj(out)

        return query + self.dropout(out)


class BidirectionalCrossAttention(nn.Module):
    """
    Bidirectional: MAE attends to DenseNet AND DenseNet attends to MAE.
    """
    def __init__(self, mae_dim=768, dense_dim=2048, num_heads=8, dropout=0.1, proj_dim=512):
        super().__init__()

        # MAE queries DenseNet
        self.mae_cross = CrossAttentionBlock(mae_dim, dense_dim, num_heads, dropout, proj_dim)
        # DenseNet queries MAE
        self.dense_cross = CrossAttentionBlock(dense_dim, mae_dim, num_heads, dropout, proj_dim)

        # FFN blocks
        self.mae_ffn = nn.Sequential(
            nn.LayerNorm(mae_dim),
            nn.Linear(mae_dim, mae_dim * 4),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(mae_dim * 4, mae_dim),
            nn.Dropout(dropout)
        )
        self.dense_ffn = nn.Sequential(
            nn.LayerNorm(dense_dim),
            nn.Linear(dense_dim, dense_dim * 2),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(dense_dim * 2, dense_dim),
            nn.Dropout(dropout)
        )

    def forward(self, mae_tokens, dense_tokens):
        # Cross attention
        mae_out = self.mae_cross(mae_tokens, dense_tokens)
        dense_out = self.dense_cross(dense_tokens, mae_tokens)

        # FFN with residual
        mae_out = mae_out + self.mae_ffn(mae_out)
        dense_out = dense_out + self.dense_ffn(dense_out)

        return mae_out, dense_out
class LearnedLogitEnsemble(nn.Module):
    def __init__(self, num_heads=7, num_classes=14, temperature_init=1.0, use_gate=False):
        super().__init__()
        self.num_classes = num_classes
        self.num_heads = num_heads

        # 1. Per-head temperature (very important!)
        self.log_temps = nn.Parameter(torch.ones(num_heads) * math.log(temperature_init))

        # 2. Learned head weights via tiny gating network (best version)
        # Input = concatenated logits (or probs) → predicts soft weights
        gate_input_dim = num_classes * num_heads   # concatenating raw logits works best
        self.use_gate = use_gate

        if use_gate:
            self.gate = nn.Sequential(
                nn.Linear(gate_input_dim, 256),
                nn.GELU(),
                nn.LayerNorm(256),
                nn.Dropout(0.1),
                nn.Linear(256, num_heads),
            )
        else:
            # Simpler: just learn fixed weights + L2 regularization later
            self.raw_weights = nn.Parameter(torch.ones(num_heads))

    def forward(self, logits_list):
        """
        logits_list: list/tuple of 7 tensors, each (B, 14)
        """
        B = logits_list[0].size(0)
        device = logits_list[0].device

        # Step 1: Temperature scaling per head
        scaled_logits = []
        for i, logits in enumerate(logits_list):
            T = torch.exp(self.log_temps[i])           # >0 guaranteed
            scaled_logits.append(logits / (T + 1e-8))

        # Stack → (B, num_heads, num_classes)
        stacked = torch.stack(scaled_logits, dim=1)     # (B, 7, 14)

        if self.use_gate:
            # Step 2: Dynamic gating (sample-wise & class-wise aware)
            gate_in = stacked.flatten(1)               # (B, 7*14)
            raw_gate = self.gate(gate_in)              # (B, 7)
            weights = torch.softmax(raw_gate, dim=-1).unsqueeze(-1)  # (B,7,1)
        else:
            # Step 2: Fixed learned weights (still strong!)
            weights = torch.softmax(self.raw_weights, dim=0)        # (7,)
            weights = weights.view(1, self.num_heads, 1).to(device) # (1,7,1)

        # Step 3: Weighted average in logit space
        fused_logits = (stacked * weights).sum(dim=1)               # (B, 14)

        return fused_logits
class XRAYClassifier(nn.Module):
    def __init__(self, num_classes=14, c=1, mask_ratio=0, dropout=0.25, img_size=384,
                 encoder_dim=768, mlp_dim=3072, decoder_dim=512, encoder_depth=12,
                 encoder_head=8, decoder_depth=8, decoder_head=8, patch_size=8):
        super().__init__()

        # ---- MAE branch (frozen) ----
        self.mae = MaskedAutoEncoder(
            c=c, mask_ratio=0, dropout=dropout, img_size=img_size,
            encoder_dim=encoder_dim, mlp_dim=mlp_dim, decoder_dim=decoder_dim,
            encoder_depth=encoder_depth, encoder_head=encoder_head,
            decoder_depth=decoder_depth, decoder_head=decoder_head, patch_size=patch_size
        )
        for p in self.mae.parameters():
            p.requires_grad = False

        self.token_ln = nn.LayerNorm(encoder_dim)
        self.attn_selfpool_mae=AttentionPool(encoder_dim,1024)

        # ---- DenseNet branch (pretrained by you) ----
        # If your DenseNet supports 1 channel, set c=1 and remove the input duplication at forward.
        self.dense = DenseNet(c=2, k=64, num_classes=num_classes)

        self.dn_feat_dim = 2048

        # ---- Cross-Attention Fusion (NEW) ----
        self.cross_attn_layers = nn.ModuleList([
            BidirectionalCrossAttention(
                mae_dim=encoder_dim,      # 768
                dense_dim=self.dn_feat_dim,  # 2048
                num_heads=8,
                dropout=0.1,
                proj_dim=512
            )
            for _ in range(12)
        ])

        self.attn_pool_mae=AttentionPool(encoder_dim,1024)

        self.classifier_mae=nn.Sequential(
            nn.Linear(1024, 512),
            nn.GELU(),
            nn.Dropout(0.1),
            nn.Linear(512, num_classes),
        )

        self.attn_pool_dense=AttentionPool(self.dn_feat_dim,1024)

        self.classifier_attn=nn.Sequential(
            nn.Linear(2048, 1024),
            nn.GELU(),
            nn.Dropout(0.2),
            nn.Linear(1024, 512),
            nn.GELU(),
            nn.Dropout(0.1),
            nn.Linear(512, num_classes),
        )
        #FPN
        self.lateral5 = nn.Conv2d(2048, 256, kernel_size=1, stride=1, padding=0)  # feat4: 2048 ✅
        self.lateral4 = nn.Conv2d(2048, 256, kernel_size=1, stride=1, padding=0)  # feat3: 2048 (CHANGED)
        self.lateral3 = nn.Conv2d(1024, 256, kernel_size=1, stride=1, padding=0)  # feat2: 1024 ✅
        self.lateral2 = nn.Conv2d(512, 256, kernel_size=1, stride=1, padding=0)   # feat1: 512 (CHANGED)
        self.output5 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
        self.output4 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
        self.output3 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
        self.output2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
        self.upsample = nn.Upsample(scale_factor=2, mode='nearest')

        self._classify_out5 = nn.Linear(256, num_classes)
        self._classify_out4 = nn.Linear(256, num_classes)
        self._classify_out3 = nn.Linear(256, num_classes)
        self._classify_out2 = nn.Linear(256, num_classes)

        self.learned_logit_ensemble = LearnedLogitEnsemble(num_classes=num_classes)

    def forward(self, x):
        mae_tokens, _, _, _ = self.mae.encoder(x)
        mae_tokens = self.token_ln(mae_tokens)
        #self.generate_kmeans_mask(self.kmeans,mae_tokens,5)
        doublex=torch.cat([x,x],dim=1)  # [B, 2, 384, 384]
        # ---- DenseNet path - Extract multi-scale features ----
        xdense = self.dense.initialconv(doublex)  # [B, 128, 192, 192]

        # Layer 1 + ECA (BEFORE transition)
        feat1 = self.dense.layer1(xdense)
        feat1 = self.dense.dropout1(feat1)
        feat1 = self.dense.eca1(feat1)            # [B, 512, 192, 192] ← Keep this!
        xdense1 = self.dense.trans1(feat1)        # [B, 256, 96, 96]

        # Layer 2 + ECA (BEFORE transition)
        feat2 = self.dense.layer2(xdense1)
        feat2 = self.dense.dropout2(feat2)
        feat2 = self.dense.eca2(feat2)            # [B, 1024, 96, 96] ← Keep this!
        xdense2 = self.dense.trans2(feat2)        # [B, 512, 48, 48]

        # Layer 3 + ECA (BEFORE transition)
        feat3 = self.dense.layer3(xdense2)
        feat3 = self.dense.dropout3(feat3)
        feat3 = self.dense.eca3(feat3)            # [B, 2048, 48, 48] ← Keep this!
        xdense3 = self.dense.trans3(feat3)        # [B, 1024, 24, 24]

        # Layer 4 (no transition)
        feat4 = self.dense.layer4(xdense3)
        feat4 = self.dense.dropout4(feat4)
        feat4 = self.dense.eca4(feat4)            # [B, 2048, 24, 24]
        xdense4 = feat4

        # Global pooling for DenseNet classifier
        xdense_pooled = self.dense.global_average_pool(xdense4)
        xdense_pooled = xdense_pooled.view(xdense_pooled.size(0), -1)
        xdense_pooled = self.dense.dropout(xdense_pooled)
        classifier_xdense = self.dense.classifier(xdense_pooled)

        # Dense tokens for cross-attention
        dense_tokens = xdense4.flatten(2).transpose(1, 2)  # [B, 576, 2048]

        # ---- FPN with CORRECT multi-scale features ----
        c4 = self.lateral5(feat4)   # [B, 2048, 24, 24]  → [B, 256, 24, 24]
        c3 = self.lateral4(feat3)   # [B, 2048, 48, 48]  → [B, 256, 48, 48]
        c2 = self.lateral3(feat2)   # [B, 1024, 96, 96]  → [B, 256, 96, 96]
        c1 = self.lateral2(feat1)   # [B, 512, 192, 192] → [B, 256, 192, 192]

        # Top-down pathway
        p4 = c4                         # 24×24
        p4 = self.output5(p4)

        p3 = self.upsample(p4) + c3     # 48×48 + 48×48 ✅
        p3 = self.output4(p3)

        p2 = self.upsample(p3) + c2     # 96×96 + 96×96 ✅
        p2 = self.output3(p2)

        p1 = self.upsample(p2) + c1     # 192×192 + 192×192 ✅
        p1 = self.output2(p1)

        # Classification heads
        out4 = self._classify_out5(p4.mean([2, 3]))
        out3 = self._classify_out4(p3.mean([2, 3]))
        out2 = self._classify_out3(p2.mean([2, 3]))
        out1 = self._classify_out2(p1.mean([2, 3]))

        # ---- MAE path ----


        mae_tokens_pooled = self.attn_selfpool_mae(mae_tokens)
        classifier_mae = self.classifier_mae(mae_tokens_pooled)

        # ---- Cross attention ----
        for cross_layer in self.cross_attn_layers:
            mae_cross, dense_cross = cross_layer(mae_tokens, dense_tokens)

        mae_cross = self.attn_pool_mae(mae_cross)
        dense_cross = self.attn_pool_dense(dense_cross)
        out = torch.cat([mae_cross, dense_cross], dim=1)
        classifier_attn = self.classifier_attn(out)

        # ---- Ensemble ----
        merged_classifier = self.learned_logit_ensemble([
            classifier_mae,
            classifier_xdense,
            classifier_attn,
            out4, out3, out2, out1  # 7 heads
        ])

        return merged_classifier

	import torch
	import torch.nn as nn
	import torch.nn.functional as F
	import math

	from models.mae import MaskedAutoEncoder
	from models.densenet import DenseNet

	class AttentionPool(nn.Module):
	def __init__(self, dim=768, embed_dim=2048, num_heads=8):
	super().__init__()
	self.query = nn.Parameter(torch.randn(1, 1, dim))
	self.attn = nn.MultiheadAttention(embed_dim=dim, num_heads=num_heads, batch_first=True)
	self.proj = nn.Linear(dim, embed_dim)

	def forward(self, x): # x: (B, 576, 768)
	B = x.size(0)
	q = self.query.expand(B, -1, -1) # (B, 1, 768)
	attn_out, _ = self.attn(q, x, x) # (B, 1, 768)
	return self.proj(attn_out.squeeze(1)) # (B, 2048)

	class CrossAttentionBlock(nn.Module):
	"""
	Cross-attention: Query tokens attend to Key/Value tokens from another modality.
	"""
	def __init__(self, dim_q, dim_kv, num_heads=8, dropout=0.1, proj_dim=None):
	super().__init__()
	self.proj_dim = proj_dim or dim_q
	self.num_heads = num_heads
	self.head_dim = self.proj_dim // num_heads
	self.scale = self.head_dim ** -0.5

	self.q_proj = nn.Linear(dim_q, self.proj_dim)
	self.k_proj = nn.Linear(dim_kv, self.proj_dim)
	self.v_proj = nn.Linear(dim_kv, self.proj_dim)
	self.out_proj = nn.Linear(self.proj_dim, dim_q)

	self.dropout = nn.Dropout(dropout)
	self.norm_q = nn.LayerNorm(dim_q)
	self.norm_kv = nn.LayerNorm(dim_kv)

	def forward(self, query, key_value):
	B, N_q, _ = query.shape
	N_kv = key_value.shape[1]

	q = self.norm_q(query)
	kv = self.norm_kv(key_value)

	Q = self.q_proj(q).view(B, N_q, self.num_heads, self.head_dim).transpose(1, 2)
	K = self.k_proj(kv).view(B, N_kv, self.num_heads, self.head_dim).transpose(1, 2)
	V = self.v_proj(kv).view(B, N_kv, self.num_heads, self.head_dim).transpose(1, 2)

	attn = (Q @ K.transpose(-2, -1)) * self.scale
	attn = F.softmax(attn, dim=-1)
	attn = self.dropout(attn)

	out = (attn @ V).transpose(1, 2).reshape(B, N_q, self.proj_dim)
	out = self.out_proj(out)

	return query + self.dropout(out)


	class BidirectionalCrossAttention(nn.Module):
	"""
	Bidirectional: MAE attends to DenseNet AND DenseNet attends to MAE.
	"""
	def __init__(self, mae_dim=768, dense_dim=2048, num_heads=8, dropout=0.1, proj_dim=512):
	super().__init__()

	# MAE queries DenseNet
	self.mae_cross = CrossAttentionBlock(mae_dim, dense_dim, num_heads, dropout, proj_dim)
	# DenseNet queries MAE
	self.dense_cross = CrossAttentionBlock(dense_dim, mae_dim, num_heads, dropout, proj_dim)

	# FFN blocks
	self.mae_ffn = nn.Sequential(
	nn.LayerNorm(mae_dim),
	nn.Linear(mae_dim, mae_dim * 4),
	nn.GELU(),
	nn.Dropout(dropout),
	nn.Linear(mae_dim * 4, mae_dim),
	nn.Dropout(dropout)
	)
	self.dense_ffn = nn.Sequential(
	nn.LayerNorm(dense_dim),
	nn.Linear(dense_dim, dense_dim * 2),
	nn.GELU(),
	nn.Dropout(dropout),
	nn.Linear(dense_dim * 2, dense_dim),
	nn.Dropout(dropout)
	)

	def forward(self, mae_tokens, dense_tokens):
	# Cross attention
	mae_out = self.mae_cross(mae_tokens, dense_tokens)
	dense_out = self.dense_cross(dense_tokens, mae_tokens)

	# FFN with residual
	mae_out = mae_out + self.mae_ffn(mae_out)
	dense_out = dense_out + self.dense_ffn(dense_out)

	return mae_out, dense_out
	class LearnedLogitEnsemble(nn.Module):
	def __init__(self, num_heads=7, num_classes=14, temperature_init=1.0, use_gate=False):
	super().__init__()
	self.num_classes = num_classes
	self.num_heads = num_heads

	# 1. Per-head temperature (very important!)
	self.log_temps = nn.Parameter(torch.ones(num_heads) * math.log(temperature_init))

	# 2. Learned head weights via tiny gating network (best version)
	# Input = concatenated logits (or probs) → predicts soft weights
	gate_input_dim = num_classes * num_heads # concatenating raw logits works best
	self.use_gate = use_gate

	if use_gate:
	self.gate = nn.Sequential(
	nn.Linear(gate_input_dim, 256),
	nn.GELU(),
	nn.LayerNorm(256),
	nn.Dropout(0.1),
	nn.Linear(256, num_heads),
	)
	else:
	# Simpler: just learn fixed weights + L2 regularization later
	self.raw_weights = nn.Parameter(torch.ones(num_heads))

	def forward(self, logits_list):
	"""
	logits_list: list/tuple of 7 tensors, each (B, 14)
	"""
	B = logits_list[0].size(0)
	device = logits_list[0].device

	# Step 1: Temperature scaling per head
	scaled_logits = []
	for i, logits in enumerate(logits_list):
	T = torch.exp(self.log_temps[i]) # >0 guaranteed
	scaled_logits.append(logits / (T + 1e-8))

	# Stack → (B, num_heads, num_classes)
	stacked = torch.stack(scaled_logits, dim=1) # (B, 7, 14)

	if self.use_gate:
	# Step 2: Dynamic gating (sample-wise & class-wise aware)
	gate_in = stacked.flatten(1) # (B, 7*14)
	raw_gate = self.gate(gate_in) # (B, 7)
	weights = torch.softmax(raw_gate, dim=-1).unsqueeze(-1) # (B,7,1)
	else:
	# Step 2: Fixed learned weights (still strong!)
	weights = torch.softmax(self.raw_weights, dim=0) # (7,)
	weights = weights.view(1, self.num_heads, 1).to(device) # (1,7,1)

	# Step 3: Weighted average in logit space
	fused_logits = (stacked * weights).sum(dim=1) # (B, 14)

	return fused_logits
	class XRAYClassifier(nn.Module):
	def __init__(self, num_classes=14, c=1, mask_ratio=0, dropout=0.25, img_size=384,
	encoder_dim=768, mlp_dim=3072, decoder_dim=512, encoder_depth=12,
	encoder_head=8, decoder_depth=8, decoder_head=8, patch_size=8):
	super().__init__()

	# ---- MAE branch (frozen) ----
	self.mae = MaskedAutoEncoder(
	c=c, mask_ratio=0, dropout=dropout, img_size=img_size,
	encoder_dim=encoder_dim, mlp_dim=mlp_dim, decoder_dim=decoder_dim,
	encoder_depth=encoder_depth, encoder_head=encoder_head,
	decoder_depth=decoder_depth, decoder_head=decoder_head, patch_size=patch_size
	)
	for p in self.mae.parameters():
	p.requires_grad = False

	self.token_ln = nn.LayerNorm(encoder_dim)
	self.attn_selfpool_mae=AttentionPool(encoder_dim,1024)

	# ---- DenseNet branch (pretrained by you) ----
	# If your DenseNet supports 1 channel, set c=1 and remove the input duplication at forward.
	self.dense = DenseNet(c=2, k=64, num_classes=num_classes)

	self.dn_feat_dim = 2048

	# ---- Cross-Attention Fusion (NEW) ----
	self.cross_attn_layers = nn.ModuleList([
	BidirectionalCrossAttention(
	mae_dim=encoder_dim, # 768
	dense_dim=self.dn_feat_dim, # 2048
	num_heads=8,
	dropout=0.1,
	proj_dim=512
	)
	for _ in range(12)
	])

	self.attn_pool_mae=AttentionPool(encoder_dim,1024)

	self.classifier_mae=nn.Sequential(
	nn.Linear(1024, 512),
	nn.GELU(),
	nn.Dropout(0.1),
	nn.Linear(512, num_classes),
	)

	self.attn_pool_dense=AttentionPool(self.dn_feat_dim,1024)

	self.classifier_attn=nn.Sequential(
	nn.Linear(2048, 1024),
	nn.GELU(),
	nn.Dropout(0.2),
	nn.Linear(1024, 512),
	nn.GELU(),
	nn.Dropout(0.1),
	nn.Linear(512, num_classes),
	)
	#FPN
	self.lateral5 = nn.Conv2d(2048, 256, kernel_size=1, stride=1, padding=0) # feat4: 2048 ✅
	self.lateral4 = nn.Conv2d(2048, 256, kernel_size=1, stride=1, padding=0) # feat3: 2048 (CHANGED)
	self.lateral3 = nn.Conv2d(1024, 256, kernel_size=1, stride=1, padding=0) # feat2: 1024 ✅
	self.lateral2 = nn.Conv2d(512, 256, kernel_size=1, stride=1, padding=0) # feat1: 512 (CHANGED)
	self.output5 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
	self.output4 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
	self.output3 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
	self.output2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
	self.upsample = nn.Upsample(scale_factor=2, mode='nearest')

	self._classify_out5 = nn.Linear(256, num_classes)
	self._classify_out4 = nn.Linear(256, num_classes)
	self._classify_out3 = nn.Linear(256, num_classes)
	self._classify_out2 = nn.Linear(256, num_classes)

	self.learned_logit_ensemble = LearnedLogitEnsemble(num_classes=num_classes)

	def forward(self, x):
	mae_tokens, _, _, _ = self.mae.encoder(x)
	mae_tokens = self.token_ln(mae_tokens)
	#self.generate_kmeans_mask(self.kmeans,mae_tokens,5)
	doublex=torch.cat([x,x],dim=1) # [B, 2, 384, 384]
	# ---- DenseNet path - Extract multi-scale features ----
	xdense = self.dense.initialconv(doublex) # [B, 128, 192, 192]

	# Layer 1 + ECA (BEFORE transition)
	feat1 = self.dense.layer1(xdense)
	feat1 = self.dense.dropout1(feat1)
	feat1 = self.dense.eca1(feat1) # [B, 512, 192, 192] ← Keep this!
	xdense1 = self.dense.trans1(feat1) # [B, 256, 96, 96]

	# Layer 2 + ECA (BEFORE transition)
	feat2 = self.dense.layer2(xdense1)
	feat2 = self.dense.dropout2(feat2)
	feat2 = self.dense.eca2(feat2) # [B, 1024, 96, 96] ← Keep this!
	xdense2 = self.dense.trans2(feat2) # [B, 512, 48, 48]

	# Layer 3 + ECA (BEFORE transition)
	feat3 = self.dense.layer3(xdense2)
	feat3 = self.dense.dropout3(feat3)
	feat3 = self.dense.eca3(feat3) # [B, 2048, 48, 48] ← Keep this!
	xdense3 = self.dense.trans3(feat3) # [B, 1024, 24, 24]

	# Layer 4 (no transition)
	feat4 = self.dense.layer4(xdense3)
	feat4 = self.dense.dropout4(feat4)
	feat4 = self.dense.eca4(feat4) # [B, 2048, 24, 24]
	xdense4 = feat4

	# Global pooling for DenseNet classifier
	xdense_pooled = self.dense.global_average_pool(xdense4)
	xdense_pooled = xdense_pooled.view(xdense_pooled.size(0), -1)
	xdense_pooled = self.dense.dropout(xdense_pooled)
	classifier_xdense = self.dense.classifier(xdense_pooled)

	# Dense tokens for cross-attention
	dense_tokens = xdense4.flatten(2).transpose(1, 2) # [B, 576, 2048]

	# ---- FPN with CORRECT multi-scale features ----
	c4 = self.lateral5(feat4) # [B, 2048, 24, 24] → [B, 256, 24, 24]
	c3 = self.lateral4(feat3) # [B, 2048, 48, 48] → [B, 256, 48, 48]
	c2 = self.lateral3(feat2) # [B, 1024, 96, 96] → [B, 256, 96, 96]
	c1 = self.lateral2(feat1) # [B, 512, 192, 192] → [B, 256, 192, 192]

	# Top-down pathway
	p4 = c4 # 24×24
	p4 = self.output5(p4)

	p3 = self.upsample(p4) + c3 # 48×48 + 48×48 ✅
	p3 = self.output4(p3)

	p2 = self.upsample(p3) + c2 # 96×96 + 96×96 ✅
	p2 = self.output3(p2)

	p1 = self.upsample(p2) + c1 # 192×192 + 192×192 ✅
	p1 = self.output2(p1)

	# Classification heads
	out4 = self._classify_out5(p4.mean([2, 3]))
	out3 = self._classify_out4(p3.mean([2, 3]))
	out2 = self._classify_out3(p2.mean([2, 3]))
	out1 = self._classify_out2(p1.mean([2, 3]))

	# ---- MAE path ----


	mae_tokens_pooled = self.attn_selfpool_mae(mae_tokens)
	classifier_mae = self.classifier_mae(mae_tokens_pooled)

	# ---- Cross attention ----
	for cross_layer in self.cross_attn_layers:
	mae_cross, dense_cross = cross_layer(mae_tokens, dense_tokens)

	mae_cross = self.attn_pool_mae(mae_cross)
	dense_cross = self.attn_pool_dense(dense_cross)
	out = torch.cat([mae_cross, dense_cross], dim=1)
	classifier_attn = self.classifier_attn(out)

	# ---- Ensemble ----
	merged_classifier = self.learned_logit_ensemble([
	classifier_mae,
	classifier_xdense,
	classifier_attn,
	out4, out3, out2, out1 # 7 heads
	])

	return merged_classifier