"# hmm_training-inference"

README

This project corresponds to problem 4 in the pdf.

Files Included In Code

• DataSet.py: A class that can read in specific format of training/testing data.

• HMM.py: The main HMM class that include prior probabilities and the needed functions for smoothing and training.

• Viterbi.py: Viterbi class that could predict state sequences from testing observations.

• TrainHMM.py: Train the different probability matrices using training data and plot observation likelihood wrt to iteration number.

• RunViterbi.py: Train the HMM then make predictions.

• map.png: The map for this project.

• Figure_1.png: Observation likelihood vs iteration number for 50 iterations.

• randomwalk.train.txt: 200 sequences of 200 timesteps for training. Sequences are separated with "."

• randomwalk.test.txt: 200 sequences of 200 timesteps for testing Viterbi.

• one_seq.test.txt: 1 sequence of 200 timesteps for testing Viterbi.

Instructions

To train the HMM only:

$python TrainHMM.py randomwalk.train.txt

To train the HMM and make predictions:

$python RunViterbi.py randomwalk.train.txt randomwalk.test.txt

To change iteration number or the number of sequences to read in, please follow Note 2. and Note 3.

Note

For the training and testing dataset, please follow the exact formats as the provided file.

ReadFile function requires two inputs (numseq, flag). Numseq indicates the number of sequence in the file, or the number of sequence you wish to read in. Flag can only be either "train" or "test"; this will decide where to store the read sequences.

The training process will seem to converge around 15-20 iterations, resulting in accuracy ~65%, but it will suddenly start optimizaing again from 35-40 iterations and converge around 50 iterations. This is the main reason why I did not implement a threshold. Accuracy is about 68% for 50 iterations, can reach 69% for 100 iterations. Please modifay HMM.py line 203, 205, 207 if anyone truly wish to test different number of iterations.

I understand there are some limitations as to how the user can run this program, but it serves the assignment. Thank you!

Below is a test record for all 200 test sequences using a 50-iter HMM. overall target percentage: 68.26% test seq 0 misses 66 test seq 1 misses 58 test seq 2 misses 90 test seq 3 misses 92 test seq 4 misses 47 test seq 5 misses 68 test seq 6 misses 42 test seq 7 misses 49 test seq 8 misses 59 test seq 9 misses 26 test seq 10 misses 47 test seq 11 misses 48 test seq 12 misses 85 test seq 13 misses 57 test seq 14 misses 31 test seq 15 misses 53 test seq 16 misses 46 test seq 17 misses 69 test seq 18 misses 64 test seq 19 misses 78 test seq 20 misses 69 test seq 21 misses 36 test seq 22 misses 66 test seq 23 misses 90 test seq 24 misses 48 test seq 25 misses 68 test seq 26 misses 50 test seq 27 misses 44 test seq 28 misses 106 test seq 29 misses 52 test seq 30 misses 82 test seq 31 misses 47 test seq 32 misses 70 test seq 33 misses 51 test seq 34 misses 66 test seq 35 misses 43 test seq 36 misses 42 test seq 37 misses 36 test seq 38 misses 40 test seq 39 misses 64 test seq 40 misses 96 test seq 41 misses 100 test seq 42 misses 35 test seq 43 misses 68 test seq 44 misses 38 test seq 45 misses 42 test seq 46 misses 51 test seq 47 misses 59 test seq 48 misses 80 test seq 49 misses 69 test seq 50 misses 63 test seq 51 misses 109 test seq 52 misses 69 test seq 53 misses 56 test seq 54 misses 40 test seq 55 misses 65 test seq 56 misses 78 test seq 57 misses 72 test seq 58 misses 94 test seq 59 misses 75 test seq 60 misses 64 test seq 61 misses 51 test seq 62 misses 76 test seq 63 misses 66 test seq 64 misses 66 test seq 65 misses 92 test seq 66 misses 59 test seq 67 misses 41 test seq 68 misses 48 test seq 69 misses 85 test seq 70 misses 49 test seq 71 misses 60 test seq 72 misses 60 test seq 73 misses 50 test seq 74 misses 40 test seq 75 misses 45 test seq 76 misses 21 test seq 77 misses 78 test seq 78 misses 41 test seq 79 misses 53 test seq 80 misses 55 test seq 81 misses 58 test seq 82 misses 78 test seq 83 misses 63 test seq 84 misses 61 test seq 85 misses 59 test seq 86 misses 66 test seq 87 misses 47 test seq 88 misses 81 test seq 89 misses 39 test seq 90 misses 123 test seq 91 misses 67 test seq 92 misses 39 test seq 93 misses 56 test seq 94 misses 63 test seq 95 misses 40 test seq 96 misses 69 test seq 97 misses 82 test seq 98 misses 53 test seq 99 misses 60 test seq 100 misses 80 test seq 101 misses 62 test seq 102 misses 79 test seq 103 misses 76 test seq 104 misses 53 test seq 105 misses 73 test seq 106 misses 40 test seq 107 misses 76 test seq 108 misses 40 test seq 109 misses 51 test seq 110 misses 60 test seq 111 misses 58 test seq 112 misses 72 test seq 113 misses 70 test seq 114 misses 26 test seq 115 misses 48 test seq 116 misses 52 test seq 117 misses 47 test seq 118 misses 66 test seq 119 misses 61 test seq 120 misses 73 test seq 121 misses 76 test seq 122 misses 48 test seq 123 misses 65 test seq 124 misses 65 test seq 125 misses 55 test seq 126 misses 82 test seq 127 misses 71 test seq 128 misses 34 test seq 129 misses 44 test seq 130 misses 42 test seq 131 misses 68 test seq 132 misses 88 test seq 133 misses 105 test seq 134 misses 99 test seq 135 misses 78 test seq 136 misses 112 test seq 137 misses 80 test seq 138 misses 52 test seq 139 misses 91 test seq 140 misses 65 test seq 141 misses 51 test seq 142 misses 79 test seq 143 misses 55 test seq 144 misses 33 test seq 145 misses 43 test seq 146 misses 68 test seq 147 misses 44 test seq 148 misses 98 test seq 149 misses 40 test seq 150 misses 48 test seq 151 misses 52 test seq 152 misses 76 test seq 153 misses 100 test seq 154 misses 59 test seq 155 misses 62 test seq 156 misses 120 test seq 157 misses 82 test seq 158 misses 86 test seq 159 misses 96 test seq 160 misses 76 test seq 161 misses 56 test seq 162 misses 73 test seq 163 misses 97 test seq 164 misses 59 test seq 165 misses 37 test seq 166 misses 45 test seq 167 misses 57 test seq 168 misses 58 test seq 169 misses 73 test seq 170 misses 70 test seq 171 misses 61 test seq 172 misses 61 test seq 173 misses 55 test seq 174 misses 67 test seq 175 misses 75 test seq 176 misses 85 test seq 177 misses 43 test seq 178 misses 104 test seq 179 misses 67 test seq 180 misses 63 test seq 181 misses 65 test seq 182 misses 43 test seq 183 misses 71 test seq 184 misses 66 test seq 185 misses 63 test seq 186 misses 103 test seq 187 misses 83 test seq 188 misses 68 test seq 189 misses 80 test seq 190 misses 58 test seq 191 misses 43 test seq 192 misses 55 test seq 193 misses 97 test seq 194 misses 59 test seq 195 misses 43 test seq 196 misses 49 test seq 197 misses 54 test seq 198 misses 84 test seq 199 misses 47