TypeError: ('Not JSON Serializable:', <tf.Tensor: shape=(), dtype=float32, numpy=123.0>) [英] TypeError: ('Not JSON Serializable:', <tf.Tensor: shape=(), dtype=float32, numpy=123.0>)

问题描述

我有一个 production 模型，它可以通过我自己编写的 ModelConfig 进行配置，它本身就是一个 class Config(dict).在那个 ModelConfig 中，我正在设置 hidden_​​size=123.以下是我的生产模型代码的简化:

I have a production model which is configurable over a ModelConfig I wrote myself which itself is a class Config(dict). In that ModelConfig I am setting hidden_size=123. The following s a simplification of my production model-code:

class Config(dict):
  # ..

  def as_dict(self) -> dict:
    return self._serialize()

class ModelConfig(Config):
  hidden_size: int = 123

现在，我有一个自定义的 keras 模型，其实现方式如下:

Now, I have a custom keras Model which is implemented like this:

class MyModel(keras.Model):

  def __init__(self, config: ModelConfig):
    self.config = config

    self.dense = layers.Dense(config.hidden_size)
    # ...

  def get_config():
    return self.config.as_dict()

除了调用 MyModel#save 外，一切正常:

Everything works except calling MyModel#save:

model.save(model_path, save_format='tf')

我收到一个 TypeError 说:

TypeError: ('Not JSON Serializable:', <tf.Tensor: shape=(), dtype=float32, numpy=123.0>)

我知道 ModelConfig#hidden_​​size 是这里成为 Tensor/EagerTensor 的属性.

I know it's ModelConfig#hidden_size that is being the property that's becoming a Tensor/EagerTensor here.

这里奇怪的是，在我的一个测试脚本中，一切都按预期工作.我创建它是为了查看 Config 是否会导致 save() 出现问题，但情况似乎并非如此:

The weird thing here is that in one of my test-scripts everything is working as intended. I created it in order to see if Config causes a problem on save() but this appears not to be the case:

所以这里的以下内容按预期工作:

import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

from config import Config


class ModelConfig(Config):
    hidden_size: int = 123


class MyLayer(layers.Layer):

    def __init__(self, config, other=None, **kwargs):
        super().__init__(**kwargs)
        self.config = config
        self.other = other if other is not None else None
        self.dense = layers.Dense(config.hidden_size + 1)

    def call(self, inputs, **kwargs):
        x = self.other(inputs, **kwargs) if self.other is not None else inputs
        return self.dense(x, **kwargs)


class MyModel(keras.Model):

    def __init__(self, config: ModelConfig, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.config = config
        self.other = MyLayer(config, other=MyLayer(config))
        self.dense = layers.Dense(config.hidden_size)

    def call(self, inputs, **kwargs):
        x = self.dense(inputs, **kwargs)
        return self.other(x, **kwargs)

    def get_config(self):
        return self.config.to_dict()

    @tf.function(
        input_signature=[tf.TensorSpec(shape=(None, 100))]
    )
    def infer(self, inputs):
        return self.call(inputs)


def main():
    fp = '/tmp/mymodel'
    config = ModelConfig()
    model = MyModel(config)
    model(np.random.rand(1, 100))
    model.save(fp, save_format='tf')
    model = tf.saved_model.load(fp)
    print(model.infer(np.random.rand(1, 100)))

    print('All done.')


if __name__ == '__main__':
    main()

这意味着问题不在于 Config 对象，但由于某种原因，在我的生产模型中，它的一个属性导致它在调用 save() 时失败.

This means the problem is not the Config object but for some reason in my production-model one of its properties is causing it to fail on calling save().

这可能是一个大问题，但有人知道在哪里可以找到问题吗?

It's probably a big shot but does anybody have an idea where to look for the issue here?

在一种情况下，该属性变成了一个无法序列化的 Tensor/EagerTensor(这是有道理的)，而在测试脚本的情况下，它仍然是一个按预期工作的 int..

In one case the property becomes a Tensor/EagerTensor which cannot be serialized (which makes sense) and in the test-script case it stays an int which works as intended..

我也尝试了 tf.saved_model.save，结果相同.

I also tried tf.saved_model.save with the same result.

推荐答案

我猜是 tf.Tensor 不是 JSON Serializable.您可以尝试将 tf.Tensor 的对象转换为 numpy 数组，然后您应该可以保存配置以供以后重新加载.

I guess it is tf.Tensor that is not JSON Serializable. You can try to convert an object of tf.Tensor to a numpy array, then you should be able to save the config for later reloading.

这篇关于TypeError: ('Not JSON Serializable:', <tf.Tensor: shape=(), dtype=float32, numpy=123.0>)的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！