如何解决tanh 函数返回“nan”值; tanh 不稳定问题
这是一个简单的代码,用于测试来自 Deepmind 的新 TanH 包的 distrax 功能。
输出(带有意外值)
WARNING:absl:No GPU/TPU found,falling back to cpu. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
[nan nan nan nan nan nan nan nan]
[nan nan nan nan nan nan nan nan]
import distrax
import jax
import jax.numpy as jnp
BATCH_SIZE = 8
means = jnp.zeros((BATCH_SIZE,))
log_stds = jnp.full_like(means,10)
base_dist = distrax.normal(
loc=means,scale=jnp.exp(log_stds))
dist = distrax.Transformed(distribution=base_dist,bijector=distrax.Tanh()) ----> problem with this line
@jax.jit
def test(seed):
samples = dist.sample(seed=seed)
return dist.log_prob(samples)
print(test(seed=jax.random.PRNGKey(1)))
print(test(seed=jax.random.PRNGKey(2)))
其他上下文
有什么解决办法吗?
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。