keras JointDistribution minimal example not working

未關閉
  • 哪裡不對？
#2,000 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

還沒有人認領這個 Issue。

評估

難度
4/5
預估耗時
3-5 天
新手友好度
30/100
Issue 類型
缺陷
描述清晰度
基本清楚
活躍度
停滯
技術堆疊
jupyter-notebook, python
領域
machine-learning

研究方向

使用 JointDistributionNamed 範例重現連結的 Colab，並檢查 tensorflow_probability/python/layers/distribution_layer.py 中 traceback 所示的 DistributionLambda 呼叫。比較預期的 distribution 回傳值與觀察到的 dict 錯誤，並確認是否需要一條可正常運作的 model.fit 路徑，或者需要記錄一項已文件化的限制。

由索引模型根據 Issue 內容生成。

描述

Hi,
I am trying to train a keras model using a mix of JointDistribution and other layers. My goal is to be able to build the model, compile and use model.fit on data.
Somehow, I keep getting errors and can't get it working.
I could not find a minimal example with it running and I think I am missing something.
Here is the minimal example I tried:

import tensorflow as tf
import tensorflow_probability as tfp
import tf_keras as tfk  
import numpy as np

tfd = tfp.distributions
tfpl = tfp.layers

# Generate synthetic data
np.random.seed(42)
x = np.random.randn(2000, 1).astype(np.float32)
true_w, true_b = 2.0, -1.0
y = true_w * x + true_b + np.random.randn(2000, 1).astype(np.float32) * 0.5

# Define the JointDistributionNamed model
def make_joint(x):
    return tfd.JointDistributionNamed({
        "weight": tfd.Normal(loc=0., scale=1.),
        "bias": tfd.Normal(loc=0., scale=1.),
        "obs": lambda weight, bias: tfd.Independent(
            tfd.Normal(loc=weight * x + bias, scale=0.5),
            reinterpreted_batch_ndims=1
        )
    })

# Keras model using DistributionLambda
inputs = tfk.Input(shape=(1,))
dense = tfk.layers.Dense(2)(inputs)  # Not used for parameterization here, just as a placeholder

def posterior_fn(params):
    # For demonstration, use fixed x from outer scope
    jd = make_joint(x)
    return jd

outputs = tfpl.DistributionLambda(posterior_fn)(dense)
model = tfk.Model(inputs=inputs, outputs=outputs)

# Custom loss: negative log-likelihood of observed y under the joint
def nll(y_true, y_pred):
    # y_pred is a JointDistributionNamed, so evaluate log_prob of obs
    return -y_pred.log_prob({"obs": y_true})

model.compile(optimizer=tf.optimizers.Adam(0.01), loss=nll)
model.fit(x, y, epochs=5, batch_size=32)

And I got this error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
[<ipython-input-1-e7f1317b9e9b>](https://localhost:8080/#) in <cell line: 0>()
     33     return jd
     34 
---> 35 outputs = tfpl.DistributionLambda(posterior_fn)(dense)
     36 model = tfk.Model(inputs=inputs, outputs=outputs)
     37 

3 frames
[/usr/local/lib/python3.11/dist-packages/tensorflow_probability/python/layers/distribution_layer.py](https://localhost:8080/#) in _fn(*fargs, **fkwargs)
    183       # TODO(b/126056144): Remove silent handle once we identify how/why Keras
    184       # is losing the distribution handle for activity_regularizer.
--> 185       value._tfp_distribution = distribution  # pylint: disable=protected-access
    186       # TODO(b/120153609): Keras is incorrectly presuming everything is a
    187       # `tf.Tensor`. Closing this bug entails ensuring Keras only accesses

AttributeError: Exception encountered when calling layer "distribution_lambda" (type DistributionLambda).

'dict' object has no attribute '_tfp_distribution'

Call arguments received by layer "distribution_lambda" (type DistributionLambda):
  • inputs=tf.Tensor(shape=(None, 2), dtype=float32)
  • args=<class 'inspect._empty'>
  • kwargs={'training': 'None'}

I am posting it here because I feel this might be a bug.

You can reproduce it via this colab link

Please note I also tried with tfd.JointDistributionSequential and no luck either.

主要語言
Jupyter Notebook
星號
4.4k
分支
1.1k
30 天內沒有已合併 PR

環境準備

  • 沒有 Dockerfile 或 Docker Compose 檔案
  • 沒有 Pull Request 範本
  • 閱讀貢獻指南

從這裡開始

  1. 先讀完整個 Issue，再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫，在一個分支上完成修改。
  4. 送出 Pull Request，並在描述裡引用這個 Issue 編號。

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 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。