model.py

from keras.layers import *
from keras.models import Model
from keras import backend as K
from keras.engine.topology import Layer, InputSpec
from keras import initializers   
from keras.optimizers import *
from const import *
import keras

class ComputeMasking(keras.layers.Layer):
    def __init__(self, maskvalue=0,**kwargs):
        self.maskvalue=maskvalue
        super(ComputeMasking, self).__init__(**kwargs)

    def call(self, inputs, **kwargs):
        mask = K.not_equal(inputs, self.maskvalue)
        return K.cast(mask, K.floatx())*(-99)

    def compute_output_shape(self, input_shape):
        return input_shape

def get_model(Otraining,hidden=HIDDEN,dropout=DROP):

    userembedding_layer = Embedding(Otraining.shape[0]+3, hidden, trainable=True)
    itemembedding_layer = Embedding(Otraining.shape[1]+3, hidden, trainable=True)

    userid_input = Input(shape=(1,), dtype='int32')
    itemid_input = Input(shape=(1,), dtype='int32')
    
    ui_input = Input(shape=(HIS_LEN,), dtype='int32')
    neighbor_embedding_input = Input(shape=(HIS_LEN,NEIGHBOR_LEN,hidden), dtype='float32')
    mask_neighbor = Lambda(lambda x:K.cast(K.cast(K.sum(x,axis=-1),'bool'),'float32'))(neighbor_embedding_input)
    
    neighbor_embeddings = TimeDistributed(TimeDistributed(Dense(hidden)))(neighbor_embedding_input)

    uiemb = Dense(hidden,activation='sigmoid')(itemembedding_layer(ui_input))
    uiembrepeat = Lambda(lambda x :K.repeat_elements(K.expand_dims(x,axis=2),NEIGHBOR_LEN,axis=2))(uiemb) 
    attention_gat = Reshape((HIS_LEN,NEIGHBOR_LEN))(LeakyReLU()(TimeDistributed(TimeDistributed(Dense(1)))(concatenate([uiembrepeat,neighbor_embeddings]))))
    attention_gat = Lambda(lambda x:x[0]+(1-x[1])*(-99))([attention_gat,mask_neighbor])
    agg_neighbor_embeddings = Lambda(lambda x:K.sum(K.repeat_elements(K.expand_dims(x[0],axis=3),hidden,axis=3)*x[1],axis=-2))([attention_gat,neighbor_embeddings])
    
    uiemb_agg = Dense(hidden)(concatenate([agg_neighbor_embeddings,uiemb]))  
    uemb = Dense(hidden,activation='sigmoid')(Flatten()(userembedding_layer(userid_input)))
    uemb = Dropout(dropout)(uemb)
    iemb = Dense(hidden,activation='sigmoid')(Flatten()(itemembedding_layer(itemid_input)))
    iemb = Dropout(dropout)(iemb)
    
    masker = ComputeMasking(Otraining.shape[1]+2)(ui_input)
    uembrepeat = Lambda(lambda x :K.repeat_elements(K.expand_dims(x,axis=1),HIS_LEN,axis=1))(uemb) 

    attention = Flatten()(LeakyReLU()(Dense(1)(concatenate([uembrepeat,uiemb_agg]))))
    attention = add([attention,masker])
    attention_weight = Activation('softmax')(attention)
    uemb_g = Dot((1, 1))([uiemb, attention_weight])
    uemb_g = Dense(hidden)(concatenate([uemb_g, uemb]))

    out = Dense(1,activation='sigmoid')(concatenate([uemb_g, iemb]))
    model = Model([userid_input,itemid_input,ui_input,neighbor_embedding_input],out)
    model.compile(loss=['mse'], optimizer=SGD(lr=LR,clipnorm=CLIP), metrics=['mse'])
    return model,userembedding_layer,itemembedding_layer