"""Experimental EfficientNet-B0 variants with tied repeat-block kernels. The reference ``models.py`` implementation remains untouched while the live recovery registrations that hash it are running. This isolated module reuses the reference stage geometry, preserves its stable layer names, and builds the two requested tied-weight variants. Convolutional and SE kernels are shared; each repeat keeps distinct BatchNorm state and residual/drop-path operations. """ from __future__ import annotations import math import tensorflow as tf from imagenet_research.models import ( STAGE_SPECS, VARIANTS, _conv_bn_relu6, _mbconv, round_filters, round_repeats, ) def _make_repeat_kernel_set( *, channels: int, kernel: int, expansion: int, se_ratio: float, name: str, ) -> dict[str, tf.keras.layers.Layer]: """Create one stride-1 MBConv kernel set with baseline-compatible names.""" expanded = channels * expansion squeeze = max(1, int(channels * se_ratio)) layers = {} if expansion != 1: layers["expand"] = tf.keras.layers.Conv2D( expanded, 1, padding="same", use_bias=False, kernel_initializer="he_normal", name=f"{name}_expand_conv", ) layers["depthwise"] = tf.keras.layers.DepthwiseConv2D( kernel, strides=1, padding="same", use_bias=False, depthwise_initializer="he_normal", name=f"{name}_depthwise", ) layers["se_reduce"] = tf.keras.layers.Conv2D( squeeze, 1, activation="relu", kernel_initializer="he_normal", name=f"{name}_se_reduce", ) layers["se_expand"] = tf.keras.layers.Conv2D( expanded, 1, activation="sigmoid", kernel_initializer="glorot_uniform", name=f"{name}_se_expand", ) layers["project"] = tf.keras.layers.Conv2D( channels, 1, padding="same", use_bias=False, kernel_initializer="he_normal", name=f"{name}_project_conv", ) return layers def _shared_repeat( inputs, *, filters: int, kernel: int, expansion: int, se_ratio: float, drop_rate: float, call_name: str, kernels: dict[str, tf.keras.layers.Layer], ): """Apply a shared MBConv kernel set with call-specific BatchNorm layers.""" residual = inputs x = inputs if expansion != 1: x = kernels["expand"](x) x = tf.keras.layers.BatchNormalization( momentum=0.99, epsilon=1e-3, name=f"{call_name}_expand_bn" )(x) x = tf.keras.layers.ReLU(max_value=6.0, name=f"{call_name}_expand_relu6")(x) x = kernels["depthwise"](x) x = tf.keras.layers.BatchNormalization( momentum=0.99, epsilon=1e-3, name=f"{call_name}_depthwise_bn" )(x) x = tf.keras.layers.ReLU(max_value=6.0, name=f"{call_name}_depthwise_relu6")(x) squeeze = tf.keras.layers.GlobalAveragePooling2D( keepdims=True, name=f"{call_name}_se_pool" )(x) squeeze = kernels["se_reduce"](squeeze) squeeze = kernels["se_expand"](squeeze) x = tf.keras.layers.Multiply(name=f"{call_name}_se_scale")([x, squeeze]) x = kernels["project"](x) x = tf.keras.layers.BatchNormalization( momentum=0.99, epsilon=1e-3, name=f"{call_name}_project_bn" )(x) if drop_rate > 0: x = tf.keras.layers.Dropout( drop_rate, noise_shape=(None, 1, 1, 1), name=f"{call_name}_drop_path", )(x) return tf.keras.layers.Add(name=f"{call_name}_add")([residual, x]) def build_shared_weight_efficientnet( variant: str = "b0", num_classes: int = 1_000, *, tie_mode: str = "single", tied_stages: tuple[int, ...] | None = None, stage6_blocks: int | None = None, ) -> tf.keras.Model: """Build an EfficientNet whose repeated stride-1 blocks share kernels. ``tie_mode='single'`` maps all repeated blocks in selected stages to one shared kernel set, yielding ``[B1, A, A, ...]``. ``tie_mode='pair'`` maps repeats alternately to two sets, yielding ``[B1, A, B, A, B, ...]``. The preset pair experiment uses only stage 6 and five total stage-6 blocks. """ if variant not in VARIANTS: raise ValueError(f"Unknown EfficientNet scale {variant!r}") if tie_mode not in ("single", "pair"): raise ValueError("tie_mode must be 'single' or 'pair'") if tied_stages is None: tied_stages = (4, 5, 6) if tie_mode == "single" else (6,) tied_stages = tuple(sorted(set(int(stage) for stage in tied_stages))) if any(stage < 1 or stage > 7 for stage in tied_stages): raise ValueError("tied_stages must be stage numbers from 1 through 7") if stage6_blocks is not None and stage6_blocks < 2: raise ValueError("stage6_blocks must be at least two") scale = VARIANTS[variant] image = tf.keras.Input( shape=(scale.image_size, scale.image_size, 3), name="image" ) x = _conv_bn_relu6( image, round_filters(32, scale.width), 3, 2, "stem" ) repeats_by_stage = [round_repeats(spec[3], scale.depth) for spec in STAGE_SPECS] if stage6_blocks is not None: repeats_by_stage[5] = int(stage6_blocks) total_blocks = sum(repeats_by_stage) block_index = 0 filters_in = round_filters(32, scale.width) stage_shared_kernels: dict[int, list[dict[str, tf.keras.layers.Layer]]] = {} for stage_index, (kernel, expansion, base_filters, _, first_stride) in enumerate(STAGE_SPECS): stage_number = stage_index + 1 filters_out = round_filters(base_filters, scale.width) repeats = repeats_by_stage[stage_index] for block in range(repeats): stride = first_stride if block == 0 else 1 drop_rate = 0.1 * block_index / max(total_blocks - 1, 1) call_name = f"stage{stage_number}_block{block + 1}" if stage_number in tied_stages and block > 0: if stride != 1 or filters_in != filters_out: raise ValueError( f"Stage {stage_number} repeat {block + 1} is not a shape-compatible residual block" ) slots = 1 if tie_mode == "single" else 2 if stage_number not in stage_shared_kernels: stage_shared_kernels[stage_number] = [ _make_repeat_kernel_set( channels=filters_out, kernel=kernel, expansion=expansion, se_ratio=0.25, name=f"stage{stage_number}_block{slot + 2}", ) for slot in range(slots) ] slot = 0 if tie_mode == "single" else (block - 1) % 2 x = _shared_repeat( x, filters=filters_out, kernel=kernel, expansion=expansion, se_ratio=0.25, drop_rate=drop_rate, call_name=call_name, kernels=stage_shared_kernels[stage_number][slot], ) else: x = _mbconv( x, filters_in=filters_in, filters_out=filters_out, kernel=kernel, stride=stride, expansion=expansion, se_ratio=0.25, drop_rate=drop_rate, name=call_name, ) filters_in = filters_out block_index += 1 head = _conv_bn_relu6( x, round_filters(1280, scale.width), 1, 1, "head" ) pooled = tf.keras.layers.GlobalAveragePooling2D(name="global_pool")(head) pooled = tf.keras.layers.Dropout(scale.dropout, name="classifier_dropout")(pooled) logits = tf.keras.layers.Dense(num_classes, dtype="float32", name="logits")(pooled) suffix = "shared_single" if tie_mode == "single" else "shared_pair" return tf.keras.Model(image, logits, name=f"efficientnet_{variant}_{suffix}_relu6") def build_parameter_compression_variant(num_classes: int = 1_000) -> tf.keras.Model: """Arm 1: tie repeated kernels in stages 4, 5, and 6.""" return build_shared_weight_efficientnet( "b0", num_classes, tie_mode="single", tied_stages=(4, 5, 6) ) def build_stage6_pair_variant(num_classes: int = 1_000) -> tf.keras.Model: """Arm 2: stage 6 becomes ``[B1, A, B, A, B]`` using two repeat sets.""" return build_shared_weight_efficientnet( "b0", num_classes, tie_mode="pair", tied_stages=(6,), stage6_blocks=5 ) def build_stage6_pair_four_variant(num_classes: int = 1_000) -> tf.keras.Model: """Kernel-budget control: stage 6 becomes ``[B1, A, B, A]``.""" return build_shared_weight_efficientnet( "b0", num_classes, tie_mode="pair", tied_stages=(6,), stage6_blocks=4 ) def kernel_sharing_report(model: tf.keras.Model) -> dict: """Summarize unique parameter ownership in a built shared-weight model.""" shared = [] for layer in model.layers: if not isinstance(layer, (tf.keras.layers.Conv2D, tf.keras.layers.DepthwiseConv2D)): continue nodes = len(getattr(layer, "_inbound_nodes", ())) if nodes > 1: parameter_count = sum(int(math.prod(weight.shape)) for weight in layer.trainable_weights) shared.append({"layer": layer.name, "calls": nodes, "kernel_and_bias_parameters": parameter_count}) trainable_parameter_count = sum( int(math.prod(weight.shape)) for weight in model.trainable_weights ) non_trainable_parameter_count = sum( int(math.prod(weight.shape)) for weight in model.non_trainable_weights ) return { "model": model.name, "unique_trainable_parameters": trainable_parameter_count, "non_trainable_parameters": non_trainable_parameter_count, "total_model_parameters": trainable_parameter_count + non_trainable_parameter_count, "reused_kernel_layers": shared, "reused_kernel_layer_count": len(shared), "batch_norm_rule": "BatchNorm objects are created per repeat call and are not shared", }