MultiTaskGaussianProcessRegressionModel issue with tf.function

未关闭
  • 哪里不对？
#1,981 6 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
3/5
预计耗时
1-2 天
新手友好度
35/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
停滞
技术栈
python, tensorflow
领域
machine-learning

调研方向

首先使用 tf.function 和形状为 [None, 1] 的 input_signature 重现该示例。检查 tensorflow_probability/python/experimental/distributions/multitask_gaussian_process_regression_model.py，尤其是 _compute_flattened_covariance 以及报告的维度处理，然后验证该示例对于未知维度不再引发 UnboundLocalError。

由索引模型根据 Issue 内容生成。

描述

Hi, I have encountered an error when using tf.function decorator on a function calling MultiTaskGaussianProcessRegressionModel from the experimental package. The error occurs specifically when passing input_signature of unknown dimension to tf.function. Below are the details. Thanks!

Code
import tensorflow as tf
import tensorflow_probability as tfp

tfde = tfp.experimental.distributions
tfk = tfp.math.psd_kernels
tfke = tfp.experimental.psd_kernels

base_kernel = tfk.ExponentiatedQuadratic(
    amplitude=tf.convert_to_tensor(0.6, tf.float64),
    length_scale=tf.convert_to_tensor(0.5, tf.float64),
)
kernel = tfke.Independent(num_tasks=2, base_kernel=base_kernel)
observations = tf.constant([[0., 1.],[-0.5, -1.0]], tf.float64)
observation_index_points = tf.constant([[0],[1.0]], tf.float64)

@tf.function(input_signature=[tf.TensorSpec(shape=[None, 1], dtype=tf.float64)])
def predict(index_points):
    gp = tfde.MultiTaskGaussianProcessRegressionModel(
        kernel=kernel,
        observations=observations,
        observation_index_points=observation_index_points,
        index_points=index_points,
    )
    return gp.mean()

predict(tf.constant([[0.2],[0.4],[0.6]], tf.float64))

The code above works:

  • if we remove tf.function, or
  • if we remove input_signature inside with tf.function

We get the following error if we set shape=[None, 1] inside tf.TensorSpec:

UnboundLocalError: in user code:

    File "/tmp/ipykernel_299029/1064227357.py", line 24, in predict  *
        return gp.mean()
    File "/usr/.local/lib/python3.12/site-packages/tensorflow_probability/python/distributions/distribution.py", line 1536, in mean  **
        return self._mean(**kwargs)
    File "/usr/.local/lib/python3.12/site-packages/tensorflow_probability/python/experimental/distributions/multitask_gaussian_process_regression_model.py", line 852, in _mean
        self._get_flattened_marginal_distribution(
    File "/usr/.local/lib/python3.12/site-packages/tensorflow_probability/python/experimental/distributions/multitask_gaussian_process_regression_model.py", line 831, in _get_flattened_marginal_distribution
        covariance = self._compute_flattened_covariance(index_points)
    File "/usr/.local/lib/python3.12/site-packages/tensorflow_probability/python/experimental/distributions/multitask_gaussian_process_regression_model.py", line 811, in _compute_flattened_covariance
        cholinv_kzx = observation_scale.solve(kxz, adjoint_arg=True)

    UnboundLocalError: cannot access local variable 'dim' where it is not associated with a value
主要语言
Jupyter Notebook
星标
4.4k
派生
1.1k
30 天内没有已合并 PR

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

tensorflow/probability 的其他 Issue

查看 tensorflow/probability 的全部 Issue

相似的 Issue

更多 Machine Learning Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。