tensorflow - 模型中无法识别 Input

问题描述

我在训练时对原始模型进行了一些修改我收到了这个错误：我们目前不支持使用在其第一层或子类模型中没有/设置的Sequential模型创建的分布策略。我是初学者，我现在该怎么办？input_shape/input_dim如何创建，应该在哪里添加？</p> input_shapeinput_dim
"""
 class SASRec(tf.keras.Model):
    def __init__(self, item_fea_col, blocks=1, num_heads=1,ffn_hidden_unit=128,max_relative_position=2,dropout=0., maxlen=40, norm_training=True, 
    causality=False, embed_reg=1e-6):
    
    super(SASRec, self).__init__()
    # sequence length
    self.maxlen = maxlen
    # item feature columns
    self.item_fea_col = item_fea_col
    # embed_dim
    self.embed_dim = self.item_fea_col['embed_dim']
    # d_model must be the same as embedding_dim, because of residual connection
    self.d_model = self.embed_dim
    self.max_relative_position = max_relative_position
    # item embedding
    self.item_embedding = Embedding(input_dim=self.item_fea_col['feat_num'],
                                    input_length=1,
                                    output_dim=self.item_fea_col['embed_dim'],
                                    mask_zero=True,
                                    embeddings_initializer='random_uniform',
                                    embeddings_regularizer=l2(embed_reg))
    # self.pos_embedding = Embedding(input_dim=self.maxlen,
    #                                input_length=1,
    #                                output_dim=self.embed_dim,
    #                                mask_zero=False,
    #                                embeddings_initializer='random_uniform',
    #                                embeddings_regularizer=l2(embed_reg))
    self.dropout = Dropout(dropout)
    # attention block
    self.encoder_layer = [EncoderLayer(self.d_model, num_heads, ffn_hidden_unit,
                                       dropout, norm_training, causality) for b in range(blocks)]

    self.GRU_layer1 = GRU(self.item_fea_col['embed_dim'], activation='tanh', use_bias=True, dropout=0.3,
                          return_sequences=True)
 
    self.Dense1 = Dense(300, use_bias=True)
    self.Dense2 = Dense(1, activation='sigmoid', use_bias=True)
    self.Dense3 = DenseA(self.d_model, ffn_hidden_unit)


def call(self, inputs, training=None):
    # inputs
    seq_inputs, pos_inputs, neg_inputs, gru_neg_inputs, gru_pos_inputs = inputs  # (None, maxlen), (None, 1), (None, 1)
    # mask
    mask = tf.expand_dims(tf.cast(tf.not_equal(seq_inputs, 0), dtype=tf.float32), axis=-1)  # (None, maxlen, 1)
    # seq info
    seq_embed = self.item_embedding(seq_inputs)  # (None, maxlen, dim)

    gru_neg_embed = self.item_embedding(gru_neg_inputs)
    gru_pos_embed = self.item_embedding(gru_pos_inputs)

    gru_inputs = seq_embed
    gru_outputs = self.GRU_layer1(gru_inputs)
    gru_outputs = self.Dense3(gru_outputs)

    gru_pos_logits = tf.reduce_sum(gru_outputs * gru_pos_embed, axis=-1)
    gru_neg_logits = tf.reduce_sum(gru_outputs * gru_neg_embed, axis=-1)

    gru_pos_logits = tf.math.log(tf.clip_by_value(gru_pos_logits, 1e-8, tf.reduce_max(gru_pos_logits)))
    gru_neg_logits = tf.math.log(tf.clip_by_value(gru_neg_logits, 1e-8, tf.reduce_max(gru_neg_logits)))

    aux_loss = tf.reduce_mean(- tf.math.log(tf.nn.sigmoid(gru_pos_logits)) -
                    tf.math.log(1 - tf.nn.sigmoid(gru_neg_logits))) / 2

    # pos encoding
    # pos_encoding = positional_encoding(seq_inputs, self.embed_dim)
    # pos_encoding = tf.expand_dims(self.pos_embedding(tf.range(self.maxlen)), axis=0)
    # seq_embed += pos_encoding
    seq_embed = self.dropout(seq_embed)
    att_outputs = seq_embed  # (None, maxlen, dim)
    att_outputs *= mask

    for block in self.encoder_layer:

        att_outputs = block([att_outputs, mask])  # (None, seq_len, dim)
        att_outputs *= mask

    cos_similarity = tf.keras.losses.cosine_similarity(seq_embed, gru_outputs)
    a = cos_similarity + 1

    # user_info = tf.reduce_mean(att_outputs, axis=1)  # (None, dim)

    # item info
    pos_info = self.item_embedding(pos_inputs)  # (None, 1, dim)
    neg_info = self.item_embedding(neg_inputs)  # (None, 1/100, dim)

    inputs1 = tf.concat([seq_embed, gru_outputs, gru_pos_embed], axis=-1)
    out = self.Dense1(inputs1)
    sim = self.Dense2(out)
    outputs = sim * gru_outputs + (1 - sim) * att_outputs
    user_info = tf.expand_dims(outputs[:, -1], axis=1)  # (None, 1, dim)

    pos_logits = tf.reduce_sum(user_info * pos_info, axis=-1)  # (None, 1)
    neg_logits = tf.reduce_sum(user_info * neg_info, axis=-1)  # (None, 1)

    # loss
    d = tf.reduce_mean(a)

    losses = tf.reduce_mean(- tf.math.log(tf.nn.sigmoid(pos_logits)) -
                            tf.math.log(1 - tf.nn.sigmoid(neg_logits))) / 2 + d * aux_loss
    self.add_loss(losses)
    logits = tf.concat([pos_logits, neg_logits], axis=-1)
    return logits

def summary(self):
    seq_inputs = Input(shape=(self.maxlen,), dtype=tf.int32)
    pos_inputs = Input(shape=(1,), dtype=tf.int32)
    neg_inputs = Input(shape=(1,), dtype=tf.int32)
    gru_neg_inputs = Input(shape=(self.maxlen,), dtype=tf.int32)
    gru_pos_inputs = Input(shape=(self.maxlen,), dtype=tf.int32)

    tf.keras.Model(inputs=[seq_inputs, pos_inputs, neg_inputs, gru_neg_inputs, gru_pos_inputs],
                   outputs=self.call([seq_inputs, pos_inputs, neg_inputs, gru_neg_inputs, gru_pos_inputs])).summary()
"""
标签： tensorflowkerasmodelsequential
tensorflow - 模型中无法识别 Input_shape

问题描述

解决方案

推荐阅读
