如何解决TensorFlow中int64转string不支持的报错？

我想基于TensorFlow Recommenders搭建一个小型推荐系统，按照官方指引开发，但在数据处理环节遇到报错。我从远程数据库获取数据，相关代码如下：

数据库读取引擎代码 #

import sqlalchemy as engines
import tensorflow as tf
import pandas as pd

class BaseEngine():
    """BaseEngine class"""
    def __init__(self, url: str):
        self.engines = engines.create_engine(url)
    
    # 获取数据表
    def get_table(self, name):
        data = pd.read_sql_query("SELECT * from {}".format(name), self.engines)
        # data.to_csv('data/{}'.format(name))
        return data
    
    # 提取多特征
    def multi_features(self, tablename: str):
        data = tf.data.Dataset.from_tensor_slices(dict(self.get_table(tablename)))
        return data
        
    # 提取特征
    def feature(self, tablename: str):
        data = tf.data.Dataset.from_tensor_slices(dict(self.get_table(tablename)))
        return data

主程序代码 #

from random import shuffle
from myengine.engines.ormengines.mysql_engine import MysqlEngine
import tensorflow as tf
import numpy as np

engine = MysqlEngine('sqlite:///base.db')

ratings = engine.feature('ratings')
movies = engine.feature('movies')

for x in ratings.take(1).as_numpy_iterator():
    print(x)
    
for x in movies.take(1).as_numpy_iterator():
    print(x)


# 数据拆分
tf.random.set_seed(42)
shuffled = ratings.shuffle(100_000, seed=42, reshuffle_each_iteration=False)

train = shuffled.take(80_000)
test = shuffled.skip(80_000).take(20_000)


def iter_x(e):
    return {
        'userId': e['userId'],
        'movieId': e['movieId']
    }

def iter_y(e):
    return {
        'title': e['title'],
        'movieId': e['movieId']
    }

def mapping(x):
    return x['userId']

# 数据分批与特征映射
movies = movies.batch(1_000).map(iter_y)
ratings = ratings.batch(1_000_000).map(iter_x)
print(ratings)
print(movies)

user_ids_vocabulary = tf.keras.layers.StringLookup(mask_token=None)
user_ids_vocabulary.adapt(ratings.map(mapping))

报错信息 #

2022-08-19 22:58:28.366801: W tensorflow/core/framework/op_kernel.cc:1722] OP_REQUIRES failed at cast_op.cc:121 : UNIMPLEMENTED: Cast int64 to string is not supported
Traceback (most recent call last):
  File "C:\Users\guera\OneDrive\Documents\recommender\project\app.py", line 48, in <module>
    user_ids_vocabulary.adapt(ratings.map(mapping))
  File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\layers\preprocessing\string_lookup.py", line 396, in adapt
    super().adapt(data, batch_size=batch_size, steps=steps)
  File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\engine\base_preprocessing_layer.py", line 249, in adapt
    self._adapt_function(iterator)
  File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\tensorflow\python\util\traceback_utils.py", line 153, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\tensorflow\python\eager\execute.py", line 54, in quick_execute
    tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
tensorflow.python.framework.errors_impl.UnimplementedError: Graph execution error:

Detected at node 'Cast' defined at (most recent call last):
    File "C:\Users\guera\OneDrive\Documents\recommender\project\app.py", line 48, in <module>     
      user_ids_vocabulary.adapt(ratings.map(mapping))
    File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\layers\preprocessing\string_lookup.py", line 396, in adapt
      super().adapt(data, batch_size=batch_size, steps=steps)
    File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\engine\base_preprocessing_layer.py", line 249, in adapt
      self._adapt_function(iterator)
    File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\engine\base_preprocessing_layer.py", line 118, in adapt_step
      self.update_state(data)
    File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\layers\preprocessing\index_lookup.py", line 531, in update_state
      data = utils.ensure_tensor(data, dtype=self.vocabulary_dtype)
    File "C:\Users\guera\AppData\Local\Programs\Python\Python310\lib\site-packages\keras\layers\preprocessing\preprocessing_utils.py", line 33, in ensure_tensor
      inputs = tf.cast(inputs, dtype)
Node: 'Cast'
Cast int64 to string is not supported
         [[{{node Cast}}]] [Op:__inference_adapt_step_117]

解决方案 #

报错核心是**Cast int64 to string is not supported**：你使用StringLookup层处理userId，但userId是int64数值类型，TensorFlow不支持直接将整数强制转为字符串类型。

方法一：将userId转换为字符串类型 #

可以在数据读取阶段或Dataset处理阶段把userId转成字符串：

1. 在数据库读取时转换：

def get_table(self, name):
    data = pd.read_sql_query("SELECT * from {}".format(name), self.engines)
    # 针对ratings表转换userId类型
    if name == 'ratings':
        data['userId'] = data['userId'].astype(str)
    return data

2. 或者在Dataset的map操作中转换：

def mapping(x):
    return tf.strings.as_string(x['userId'])

方法二：改用IntegerLookup层处理数值ID #

如果userId本身是整数类型，没必要转成字符串，直接用专门处理整数的IntegerLookup层更高效：

# 替换原有的StringLookup为IntegerLookup
user_ids_vocabulary = tf.keras.layers.IntegerLookup(mask_token=None)
user_ids_vocabulary.adapt(ratings.map(mapping))

推荐使用方法二，因为IntegerLookup是为整数ID设计的预处理层，更贴合数据类型，性能也更好。

内容的提问来源于stack exchange，提问作者Amoungui Serge