marvinbertin / hiddenmarkovmodel_tensorflow

TensorFlow: Viterbi, Forward-Backward and Baum Welch with a Hidden Markov Model (HMM)

Hidden Markov Model in TensorFlow

##Jupyter Notebook: Check out the Notebook for Examples.

Viterbi Algorithm

• Efficient way of finding the most likely state sequence.
• Method is general statistical framework of compound decision theory.
• Maximizes a posteriori probability recursively.
• Assumed to have a finite-state discrete-time Markov process.

Forward-Backward Algorithm

• The goal of the forward-backward algorithm is to find the conditional distribution over hidden states given the data.
• It is used to find the most likely state for any point in time.
• It cannot, however, be used to find the most likely sequence of states (see Viterbi)

Baum Welch Algorithm

• Expectation Maximization Inference of unknown parameters of a Hidden Markov Model.

Viterbi

Unrolled graph

Belief Propagation

Backtrack

Baum Welch and Forward-Backward

Forward-Backward

Re-estimate

Unrolled

hiddenmarkovmodel_tensorflow's Issues

ExpandDims dim value error

My problem is that I give observation sequence 3~5 states, the program can run, but when I give observation sequence 2 states, it raised some errors.

Below is error,

ValueError: dim 1 not in the interval [-1, 0]. for 'Train_Baum_Welch/EM_step-0/Re_estimate_transition/Smooth_gamma/ExpandDims' (op: 'ExpandDims') with input shapes: [], [] and with computed input tensors: input[1] = <1>.

Is it caused by the tensor dimension? Could you please help me to solve the issue? Appreciate for your help.

Warm regards,
Jacky

running forward backward AttributeError

It seems in running forward backward algorithm there are errors:

`AttributeError Traceback (most recent call last)
in ()
2 model = HiddenMarkovModel_FB(trans, emi, p0)
3
----> 4 results = model.run_forward_backward(obs_seq)
5 result_list = ["Forward", "Backward", "Posterior"]
6

/Users/xxx/Desktop/hmm/forward_bakward.pyc in run_forward_backward(self, obs_seq)
207 with tf.Session() as sess:
208
--> 209 forward, backward, posterior = self.forward_backward(obs_seq)
210 sess.run(tf.initialize_all_variables())
211 return sess.run([forward, backward, posterior])

/Users/xxx/Desktop/hmm/forward_bakward.pyc in forward_backward(self, obs_seq)
188
189 # forward belief propagation
--> 190 self._forward(obs_prob_seq)
191
192 # backward belief propagation

/Users/xxx/Desktop/hmm/forward_bakward.pyc in _forward(self, obs_prob_seq)
119 prior_prob = tf.matmul(prev_prob, self.T)
120 # forward belief propagation
--> 121 forward_score = tf.mul(prior_prob, tf.cast(obs_prob_seq[step, :], tf.float64))
122 # Normalize score into a probability
123 forward_prob = tf.reshape(forward_score / tf.reduce_sum(forward_score), [-1])

AttributeError: 'module' object has no attribute 'mul'`