""" Attention U-Net for RockGlacierNet. Architecture: Oktay et al. (2018) "Attention U-Net: Learning Where to Look for the Pancreas" — standard U-Net encoder/decoder with soft attention gates on every skip connection. Input : (128, 128, 12) Output: (128, 128, 1) sigmoid — binary glacier mask Encoder filters : 64 → 128 → 256 → 512 Bottleneck : 512 Decoder filters : 256 → 128 → 64 → 32 """ import tensorflow as tf from tensorflow.keras.layers import ( Input, Conv2D, Conv2DTranspose, MaxPooling2D, BatchNormalization, Activation, concatenate, Multiply ) from tensorflow.keras.models import Model # ───────────────────────────────────────────────────────────────────────────── # Building blocks # ───────────────────────────────────────────────────────────────────────────── def _double_conv(x, filters, dropout=0.0, name=""): """Two Conv3×3 + BN + ReLU blocks (standard U-Net block).""" x = Conv2D(filters, 3, padding='same', name=f'{name}_c1')(x) x = BatchNormalization(name=f'{name}_bn1')(x) x = Activation('relu', name=f'{name}_r1')(x) if dropout > 0: x = tf.keras.layers.Dropout(dropout, name=f'{name}_dp')(x) x = Conv2D(filters, 3, padding='same', name=f'{name}_c2')(x) x = BatchNormalization(name=f'{name}_bn2')(x) x = Activation('relu', name=f'{name}_r2')(x) return x def _attention_gate(x, g, inter_filters, name=""): """ Soft attention gate. Parameters ---------- x : skip-connection features (H, W, C_x) — from encoder g : gating signal (H, W, C_g) — from decoder (same spatial size) inter_filters : intermediate channel size (typically C_x // 2) Returns ------- Attention-weighted skip features (H, W, C_x) """ # Project skip and gating to inter_filters theta = Conv2D(inter_filters, 1, padding='same', use_bias=False, name=f'{name}_theta')(x) phi = Conv2D(inter_filters, 1, padding='same', use_bias=False, name=f'{name}_phi')(g) # Add → ReLU → 1×1 conv → sigmoid = attention coefficient map f = Activation('relu', name=f'{name}_relu')(theta + phi) psi = Conv2D(1, 1, padding='same', use_bias=True, name=f'{name}_psi')(f) alpha = Activation('sigmoid', name=f'{name}_sigmoid')(psi) # Scale skip features by attention map (broadcast over channels) return Multiply(name=f'{name}_out')([x, alpha]) # ───────────────────────────────────────────────────────────────────────────── # Full model # ───────────────────────────────────────────────────────────────────────────── def build_attention_unet(input_shape=(128, 128, 12), num_classes=1): """ Build Attention U-Net. Parameters ---------- input_shape : tuple default (128, 128, 12) num_classes : int 1 for binary segmentation Returns ------- tf.keras.Model """ inputs = Input(input_shape, name='input') # ── Encoder ────────────────────────────────────────────────── # Each encoder level: double_conv → save skip → pool e1 = _double_conv(inputs, 64, dropout=0.1, name='enc1') # 128×128, 64ch p1 = MaxPooling2D(2, name='pool1')(e1) e2 = _double_conv(p1, 128, dropout=0.1, name='enc2') # 64×64, 128ch p2 = MaxPooling2D(2, name='pool2')(e2) e3 = _double_conv(p2, 256, dropout=0.2, name='enc3') # 32×32, 256ch p3 = MaxPooling2D(2, name='pool3')(e3) e4 = _double_conv(p3, 512, dropout=0.2, name='enc4') # 16×16, 512ch p4 = MaxPooling2D(2, name='pool4')(e4) # ── Bottleneck ─────────────────────────────────────────────── b = _double_conv(p4, 512, dropout=0.3, name='btn') # 8×8, 512ch # ── Decoder ────────────────────────────────────────────────── # Each decoder level: # 1. upsample bottleneck / previous decoder feature # 2. attention_gate(skip=encoder_skip, g=upsampled) # 3. concatenate(attended_skip, upsampled) # 4. double_conv # -- Level 4 -- u4 = Conv2DTranspose(256, 2, strides=2, padding='same', name='up4')(b) ag4 = _attention_gate(x=e4, g=u4, inter_filters=256, name='ag4') # e4:512, u4:256 d4 = _double_conv(concatenate([ag4, u4], name='cat4'), 256, dropout=0.2, name='dec4') # 16×16, 256ch # -- Level 3 -- u3 = Conv2DTranspose(128, 2, strides=2, padding='same', name='up3')(d4) ag3 = _attention_gate(x=e3, g=u3, inter_filters=128, name='ag3') # e3:256, u3:128 d3 = _double_conv(concatenate([ag3, u3], name='cat3'), 128, dropout=0.2, name='dec3') # 32×32, 128ch # -- Level 2 -- u2 = Conv2DTranspose(64, 2, strides=2, padding='same', name='up2')(d3) ag2 = _attention_gate(x=e2, g=u2, inter_filters=64, name='ag2') # e2:128, u2:64 d2 = _double_conv(concatenate([ag2, u2], name='cat2'), 64, dropout=0.1, name='dec2') # 64×64, 64ch # -- Level 1 -- u1 = Conv2DTranspose(32, 2, strides=2, padding='same', name='up1')(d2) ag1 = _attention_gate(x=e1, g=u1, inter_filters=32, name='ag1') # e1:64, u1:32 d1 = _double_conv(concatenate([ag1, u1], name='cat1'), 32, dropout=0.1, name='dec1') # 128×128, 32ch # ── Output head ────────────────────────────────────────────── outputs = Conv2D(num_classes, 1, activation='sigmoid', name='output')(d1) return Model(inputs, outputs, name='AttentionUNet')