CODEWITHSUNDEEP
pythontensorflowtensorflow-probability
LOLOL
LOLOL

Reputation: 11

Exporting Tensorflow probability's Hidden Markov Model

I want to export HMM model because training it every time takes time. My method is to save all matrices in file. I want to know is there any tensorflow way I can do it? Also is it possible to export it with api to other languages like C++.

Upvotes: 0

Views: 347

Answers (2)

Brian Patton
Brian Patton

Reputation: 1076

tf.saved_model would be the recommended way to do this. Something like:

import tensorflow as tf
import tensorflow_probability as tfp

hmm = tfp.distributions.HiddenMarkovModel(
    initial_distribution=tfp.distributions.Categorical(logits=tf.Variable([0., 0])),
    transition_distribution=tfp.distributions.Categorical(logits=tf.Variable([[0., 0]] * 2)),
    observation_distribution=tfp.distributions.Normal(tf.Variable([0., 0]), 
                                                      tfp.util.TransformedVariable([1., 1], tfp.bijectors.Softplus(low=1e-3))),
    num_steps=10)

x = hmm.sample(100)

opt = tf.optimizers.Adam(0.01)

@tf.function
def one_step():
  with tf.GradientTape() as t:
    nll = -hmm.log_prob(x)
  grads = t.gradient(nll, hmm.trainable_variables)
  opt.apply_gradients(zip(grads, hmm.trainable_variables))

for _ in range(10):
  one_step()

class Foo(tf.Module):
  def __init__(self, hmm):
    self._hmm = hmm
  @tf.function(input_signature=[tf.TensorSpec.from_tensor(x)])
  def log_prob(self, x):
    return self._hmm.log_prob(x)

tf.saved_model.save(Foo(hmm), '/tmp/tf.model')
q = tf.saved_model.load('/tmp/tf.model')
q.log_prob(x)

Upvotes: 0

Poe Dator
Poe Dator

Reputation: 4893

you can iterate over and save the weights from the model variables by calling variables attribute of tfp.distributions.HiddenMarkovModel()

Upvotes: 1

Related Questions