`tfp.math.log1mexp` has a NaN gradient for small inputs (reaches `OrderedLogistic.log_prob`)

未关闭 适合新手
#2,041 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

@aviralgarg05 已经在做这个了。

开始于 2026年11月19日。

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
86/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
python

调研方向

从 tensorflow_probability/python/math/generic.py 中的 log1mexp 开始,其中一个 tf.where 在 log(-expm1(-x)) 和 log1p(-exp(-x)) 之间做选择。使用 issue 中的代码片段复现问题,然后在 generic_test.py 中添加一个梯度测试。完成的标准是:对于较小的 x,梯度是有限的;OrderedLogistic 在 loc=18 时的 log_prob 梯度约为 1.13e-07;前向计算结果保持不变。

由索引模型根据 Issue 内容生成。

描述

OrderedLogistic(...).log_prob gets a NaN gradient in float32 at loc = 18 (17 is still fine), and the NaN comes from tfp.math.log1mexp:

import tensorflow as tf
import tensorflow_probability as tfp
tfd = tfp.distributions

x = tf.constant([1e-9, 1e-7, 1e-3])
with tf.GradientTape() as tape:
  tape.watch(x)
  y = tfp.math.log1mexp(x)
print(tape.gradient(y, x).numpy())  # [nan 1.e+07 999.50006]  exact 1/expm1(x): [1e9, 1e7, 999.5]

loc = tf.Variable(18.)
with tf.GradientTape() as tape:
  lp = tfd.OrderedLogistic(cutpoints=[0., 1., 2.], loc=loc).log_prob(3)
print(tape.gradient(lp, loc).numpy())  # nan  exact sigmoid(-16) = 1.13e-07

log1mexp picks log(-expm1(-x)) or log1p(-exp(-x)) with one tf.where, so the gradient of the branch it didn't pick is still computed. For small x, exp(-x) rounds to 1, that branch's derivative is inf, and the tf.where backward turns 0 * inf into NaN. StoppingRatioLogistic, Kumaraswamy.log_cdf and Geometric.log_cdf go through the same path.

The double-where trick keeps the forward values as they are:

--- a/tensorflow_probability/python/math/generic.py
+++ b/tensorflow_probability/python/math/generic.py
@@ -703,10 +703,14 @@ def log1mexp(x, name=None):
     dtype = dtype_util.common_dtype([x], dtype_hint=tf.float32)
     x = tf.convert_to_tensor(x, dtype=dtype, name='x')
     x = tf.math.abs(x)
+    # This switching point is recommended in [1].
+    small = x < np.log(2)
+    # The branch not taken must not see small x: its derivative is infinite
+    # where exp(-x) rounds to 1, and would make the gradient NaN.
+    safe_x = tf.where(small, tf.ones_like(x), x)
     return tf.where(
-        # This switching point is recommended in [1].
-        x < np.log(2), tf.math.log(-tf.math.expm1(-x)),
-        tf.math.log1p(-tf.math.exp(-x)))
+        small, tf.math.log(-tf.math.expm1(-x)),
+        tf.math.log1p(-tf.math.exp(-safe_x)))

I ran into this while formalizing a few distribution formulas in Lean (TorchLean) and comparing them with TFP. 0.25.0 and the 2026-11-12 nightlies give the numbers above, on CPU. I have a gradient test for it in generic_test.py and can send the PR whenever you'd like :)

主要语言
Jupyter Notebook
星标
4.4k
派生
1.1k
PR 合并指标
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。