Question

我正在研究TensorFlow 2.0和Transformer，键入时标题出现错误

value = Embedding(tf.shape(vocals).numpy()[0], d_model=512)

vocals是形状为（100，45）的张量。该层的代码是：

def positional_encoding(length, d_model):
    encoded_vec = tf.Variable([pos/tf.pow(10000, 2*i/d_model) for pos in range(length) for i in range(d_model)],
                             dtype=tf.float32)
    encoded_vec[::2] = tf.sin(encoded_vec[::2])
    encoded_vec[1::2] = tf.cos(encoded_vec[1::2])

    return encoded_vec.reshape([sentence_length, dim])

class Embedding(tf.keras.layers.Layer):
    def __init__(self, vocab_size, d_model, dropout=0.1):
        super().__init__()
        self.d_model = d_model
        self.token_embedding = tf.keras.layers.Embedding(vocab_size, d_model)
        self.positional_encoding = positional_encoding(vocab_size, d_model)
        self.dropout = tf.keras.layers.Dropout(dropout)

    def call(self, x):
        seq_len = tf.shape(x)[1]
        x = self.token_embedding(x)
        x *= tf.math.sqrt(tf.cast(self.d_model, tf.float32))
        x += self.positional_encoding[:, :seq_len, :]
        x = self.dropout(x)
        return x

Answer 1

更改

encoded_vec = tf.Variable([pos/tf.pow(10000, 2*i/d_model) for pos in range(length) for i in range(d_model)],
                             dtype=tf.float32)

到

encoded_vec = np.array([pos/10000 ** (2*i/d_model) for pos in range(length) for i in range(d_model)],
                          dtype=tf.float32)

Answer 2

我对以下消息有同样的问题

<块引用>

类型错误：无法将 1e-07 转换为 dtype uint8 的 EagerTensor

尝试将 vocals 转换为所需的数据类型 np.float32，因为它询问 Cannot convert 0.0 to EagerTensor of dtype int32 我认为您的 vocals 数据类型在哪里{ {1}}。

TypeError：无法将0.0转换为dtype int32的EagerTensor

2 个答案: