Datasets
Models
Diff of
/layers/aggregator.py
[000000]
..
[c0da92]
Switch to unified view
| a | b/layers/aggregator.py | ||
|---|---|---|---|
| 1 | # -*- coding: utf-8 -*- |
||
| 2 | |||
| 3 | from keras.engine.topology import Layer |
||
| 4 | from keras import backend as K |
||
| 5 | |||
| 6 | # class AvgAggregator(Layer): |
||
| 7 | # def __init__(self, activation: str ='relu', initializer='glorot_normal', regularizer=None, |
||
| 8 | # **kwargs): |
||
| 9 | # super(AvgAggregator, self).__init__(**kwargs) |
||
| 10 | # if activation == 'relu': |
||
| 11 | # self.activation = K.relu |
||
| 12 | # elif activation == 'tanh': |
||
| 13 | # self.activation = K.tanh |
||
| 14 | # else: |
||
| 15 | # raise ValueError(f'`activation` not understood: {activation}') |
||
| 16 | # self.initializer = initializer |
||
| 17 | # self.regularizer = regularizer |
||
| 18 | # def build(self, input_shape): |
||
| 19 | # ent_embed_dim = input_shape[0][-1] |
||
| 20 | # self.w = self.add_weight(name=self.name+'_w', shape=(ent_embed_dim, ent_embed_dim), |
||
| 21 | # initializer=self.initializer, regularizer=self.regularizer) |
||
| 22 | # self.b = self.add_weight(name=self.name+'_b', shape=(ent_embed_dim,), initializer='zeros') |
||
| 23 | # super(SumAggregator, self).build(input_shape) |
||
| 24 | |||
| 25 | |||
| 26 | |||
| 27 | class SumAggregator(Layer): |
||
| 28 | def __init__(self, activation: str ='relu', initializer='glorot_normal', regularizer=None, |
||
| 29 | **kwargs): |
||
| 30 | super(SumAggregator, self).__init__(**kwargs) |
||
| 31 | if activation == 'relu': |
||
| 32 | self.activation = K.relu |
||
| 33 | elif activation == 'tanh': |
||
| 34 | self.activation = K.tanh |
||
| 35 | else: |
||
| 36 | raise ValueError(f'`activation` not understood: {activation}') |
||
| 37 | self.initializer = initializer |
||
| 38 | self.regularizer = regularizer |
||
| 39 | |||
| 40 | def build(self, input_shape): |
||
| 41 | ent_embed_dim = input_shape[0][-1] |
||
| 42 | self.w = self.add_weight(name=self.name+'_w', shape=(ent_embed_dim, ent_embed_dim), |
||
| 43 | initializer=self.initializer, regularizer=self.regularizer) |
||
| 44 | self.b = self.add_weight(name=self.name+'_b', shape=(ent_embed_dim,), initializer='zeros') |
||
| 45 | super(SumAggregator, self).build(input_shape) |
||
| 46 | |||
| 47 | def call(self, inputs, **kwargs): |
||
| 48 | entity, neighbor = inputs |
||
| 49 | return self.activation(K.dot((entity + neighbor), self.w) + self.b) |
||
| 50 | |||
| 51 | def compute_output_shape(self, input_shape): |
||
| 52 | return input_shape[0] |
||
| 53 | |||
| 54 | |||
| 55 | class ConcatAggregator(Layer): |
||
| 56 | def __init__(self, activation: str = 'relu', initializer='glorot_normal', regularizer=None, |
||
| 57 | **kwargs): |
||
| 58 | super(ConcatAggregator, self).__init__(**kwargs) |
||
| 59 | if activation == 'relu': |
||
| 60 | self.activation = K.relu |
||
| 61 | elif activation == 'tanh': |
||
| 62 | self.activation = K.tanh |
||
| 63 | else: |
||
| 64 | raise ValueError(f'`activation` not understood: {activation}') |
||
| 65 | self.initializer = initializer |
||
| 66 | self.regularizer = regularizer |
||
| 67 | |||
| 68 | def build(self, input_shape): |
||
| 69 | ent_embed_dim = input_shape[0][-1] |
||
| 70 | neighbor_embed_dim = input_shape[1][-1] |
||
| 71 | self.w = self.add_weight(name=self.name + '_w', |
||
| 72 | shape=(ent_embed_dim+neighbor_embed_dim, ent_embed_dim), |
||
| 73 | initializer=self.initializer, regularizer=self.regularizer) |
||
| 74 | self.b = self.add_weight(name=self.name + '_b', shape=(ent_embed_dim,), |
||
| 75 | initializer='zeros') |
||
| 76 | super(ConcatAggregator, self).build(input_shape) |
||
| 77 | |||
| 78 | def call(self, inputs, **kwargs): |
||
| 79 | entity, neighbor = inputs |
||
| 80 | return self.activation(K.dot(K.concatenate([entity, neighbor]), self.w) + self.b) |
||
| 81 | |||
| 82 | def compute_output_shape(self, input_shape): |
||
| 83 | return input_shape[0] |
||
| 84 | |||
| 85 | |||
| 86 | class NeighAggregator(Layer): |
||
| 87 | def __init__(self, activation: str = 'relu', initializer='glorot_normal', regularizer=None, |
||
| 88 | **kwargs): |
||
| 89 | super(NeighAggregator, self).__init__() |
||
| 90 | if activation == 'relu': |
||
| 91 | self.activation = K.relu |
||
| 92 | elif activation == 'tanh': |
||
| 93 | self.activation = K.tanh |
||
| 94 | else: |
||
| 95 | raise ValueError(f'`activation` not understood: {activation}') |
||
| 96 | self.initializer = initializer |
||
| 97 | self.regularizer = regularizer |
||
| 98 | |||
| 99 | def build(self, input_shape): |
||
| 100 | ent_embed_dim = input_shape[0][-1] |
||
| 101 | neighbor_embed_dim = input_shape[1][-1] |
||
| 102 | self.w = self.add_weight(name=self.name + '_w', |
||
| 103 | shape=(neighbor_embed_dim, ent_embed_dim), |
||
| 104 | initializer=self.initializer, regularizer=self.regularizer) |
||
| 105 | self.b = self.add_weight(name=self.name + '_b', shape=(ent_embed_dim,), |
||
| 106 | initializer='zeros') |
||
| 107 | super(NeighAggregator, self).build(input_shape) |
||
| 108 | |||
| 109 | def call(self, inputs, **kwargs): |
||
| 110 | entity, neighbor = inputs |
||
| 111 | return self.activation(K.dot(neighbor, self.w) + self.b) |
||
| 112 | |||
| 113 | def compute_output_shape(self, input_shape): |
||
| 114 | return input_shape[0] |