TF Probability and TF Decision Forests

The libraries that fix what neural networks don't do well

Standard neural networks output a single point estimate. Probabilistic models output distributions, capturing uncertainty. Critical for medical diagnosis (need confidence intervals, not just "tumor/not"), financial forecasting, safety-critical robotics.

TensorFlow Probability (TFP) provides distributions, bijectors (for normalizing flows), MCMC samplers, and DenseVariational layers for Bayesian neural networks.

TensorFlow Decision Forests (TF-DF) brings random forests and gradient boosted trees into the Keras API. For tabular data (the most common data type in enterprise ML), GBT typically outperforms neural networks unless you have millions of rows. TF-DF gives you that power with Keras conventions, TFX integration, and TF Serving compatibility — no toolkit-switching.

Code

TFP — distributions, Bayesian NN·python

import tensorflow as tf
import tensorflow_probability as tfp

tfd = tfp.distributions

# Distributions
normal = tfd.Normal(loc=0., scale=1.)
samples = normal.sample(100)
log_prob = normal.log_prob(0.5)

# Mixture of Gaussians
mixture = tfd.MixtureSameFamily(
    mixture_distribution=tfd.Categorical(probs=[0.3, 0.7]),
    components_distribution=tfd.Normal(loc=[-2., 2.], scale=[0.5, 0.5]),
)

# Bayesian neural network for uncertainty
model = tf.keras.Sequential([
    tfp.layers.DenseVariational(
        units=64,
        make_prior_fn=prior,
        make_posterior_fn=posterior,
    ),
    tf.keras.layers.Dense(1),
])

TF Decision Forests — tabular's best friend·python

import tensorflow_decision_forests as tfdf
import pandas as pd

# pip install tensorflow_decision_forests
df = pd.read_csv("adult_income.csv")
train_df, test_df = df[:800], df[800:]

train_ds = tfdf.keras.pd_dataframe_to_tf_dataset(train_df, label="income")
test_ds  = tfdf.keras.pd_dataframe_to_tf_dataset(test_df,  label="income")

# Random Forest — no LR tuning, no feature normalization needed
model = tfdf.keras.RandomForestModel()
model.fit(train_ds)

model.compile(metrics=["accuracy"])
print(model.evaluate(test_ds, return_dict=True))
model.summary()                 # built-in feature importance

# Save as TF SavedModel for deployment
model.save("decision_forest_model/")

TF Probability and TF Decision Forests

The libraries that fix what neural networks don't do well

Code

External links

Progress

Comments 0