欢迎您访问程序员文章站本站旨在为大家提供分享程序员计算机编程知识!
您现在的位置是: 首页

tensorflow踩坑记录

程序员文章站 2024-02-10 08:54:04
...
Input 'b' of 'MatMul' Op has type float32 that does not match type int32 of argument 'a'
tf.reset_default_graph
with cnn_graph.as_default():
    with tf.name_scope("placeholders"):
        inputs = tf.placeholder(dtype=tf.int32, shape=(None, max_len), name="inputs")
        targets = tf.placeholder(dtype=tf.float32, shape=(None, 1), name="targets")
    
    # embeddings
    with tf.name_scope("embeddings"):
        # 用pre-trained词向量来作为embedding层
        embedding_matrix = tf.Variable([vocab_size,embedding_size], trainable=True, name="embedding_matrix")
        embed = tf.nn.embedding_lookup(embedding_matrix, inputs, name="embed")
        # 相加词向量得到句子向量
        sum_embed = tf.reduce_sum(embed, axis=1, name="sum_embed")
    # model
    with tf.name_scope("model"):
        # 隐层权重
        W1 = tf.Variable(tf.random_normal(shape=(embedding_size, hidden_size), stddev=0.1), name="W1")
        b1 = tf.Variable(tf.zeros(shape=(hidden_size), name="b1"))
        
        # 输出层权重
        W2 = tf.Variable(tf.random_normal(shape=(hidden_size, 1), stddev=0.1), name="W2")
        b2 = tf.Variable(tf.zeros(shape=(1), name="b2"))
        
        # 结果
        z1 = tf.add(tf.matmul(sum_embed, W1), b1)
        a1 = tf.nn.relu(z1)
        
        logits = tf.add(tf.matmul(a1, W2), b2)
        outputs = tf.nn.sigmoid(logits, name="outputs")

计算z1 tf.matmul时报错,int32和float32不匹配报错。

一开始以为是变量类型不匹配导致的,把sum_embed.shape print出来以后发现和预期的输出不一样,仔细检查后发现是

embedding_matrix定义时出错了,改为:

embedding_matrix = tf.Variable(tf.random_normal(shape=(vocab_size, embedding_size), stddev=0.1), trainable=True, name="embedding_matrix")

就好了,和tf.get_variable混淆了,后续有时间会补上tf.variable_scope,tf.name_scope,tf.get_variable的用法

相关标签: tensorflow