ahmh.py

# -*- coding: utf-8 -*-
"""AHMH.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1JsnyjP7BFWuFxbhW48dw0fudfjvGX6vL
"""

# === Custom Hierarchical Convolution Layer ===
class HierarchicalConvolution(Layer):
    def __init__(self, filters, kernel_sizes, group_factors=(1, 2, 4), pool_sizes=None, **kwargs):
        super(HierarchicalConvolution, self).__init__(**kwargs)
        self.filters = filters
        self.kernel_sizes = kernel_sizes
        self.group_factors = group_factors
        self.pool_sizes = pool_sizes

    def build(self, input_shape):
        num_resolutions = len(self.kernel_sizes)
        self.conv_layers = []
        for i in range(num_resolutions):
            groups = input_shape[-1] // self.group_factors[i]
            conv_layer = Conv2D(self.filters, self.kernel_sizes[i], activation='relu', padding='same', groups=groups)
            if self.pool_sizes and i < len(self.pool_sizes):
                pool_size = self.pool_sizes[i]
                pooling = GlobalAveragePooling2D() if pool_size is None else AveragePooling2D(pool_size=pool_size)
                conv_layer = Sequential([conv_layer, pooling])
            self.conv_layers.append(conv_layer)
        super(HierarchicalConvolution, self).build(input_shape)

    def call(self, inputs):
        conv_outputs = [conv_layer(inputs) for conv_layer in self.conv_layers]
        return Concatenate(axis=-1)(conv_outputs)

    def compute_output_shape(self, input_shape):
        output_shape = list(input_shape)
        output_shape[-1] = sum([conv.output_shape[-1] for conv in self.conv_layers])
        return tuple(output_shape)

    def get_config(self):
        config = super(HierarchicalConvolution, self).get_config()
        config.update({
            'filters': self.filters,
            'kernel_sizes': self.kernel_sizes,
            'group_factors': self.group_factors,
            'pool_sizes': self.pool_sizes
        })
        return config