Similar to ../exerc01.
-
WAV file:
MARS_20161221_000046_SongSession_16kHz_HPF5Hz.wav -
Selection file:
../exerc01/MARS_20161221_000046_SongSession_16kHz_HPF5HzNorm_labels.csv.
$ ecoz2 sgn extract --segments ../exerc01/MARS_20161221_000046_SongSession_16kHz_HPF5HzNorm_labels.csv \
--wav ${SOURCE_WAV} \
--out-dir data/signals
Loading .../MARS_20161221_000046_SongSession_16kHz_HPF5Hz.wav
num_samples: 266117287 sample_rate: 16000 bits_per_sample: 16 sample_format = Int
duration: 16632 sample_period: 0.0000625
parsed ranges = []
M 60 instances
D 175 instances
EG 3 instances
E1 27 instances
Bd 50 instances
B 13 instances
I2 714 instances
F 340 instances
Bm 608 instances
P 171 instances
G2 307 instances
I3 324 instances
I4 46 instances
A 512 instances
C 550 instances
? 110 instances
I 471 instances
H 141 instances
Bu 43 instances
E 713 instances
G 76 instances
C1 16 instances
5470 total extracted instances
$ rm -rf data/signals/\?
Eg., data/signals/B/00123.wav, indicating that this is the selection
123 from the original file.
Note: no train/test split here as in exerc01.
ecoz2 lpc -P 36 -W 45 -O 15 -m 10 data/signals
Eg., data/predictors/B/00123.prd, corresponding to the signal above.
tt-list.csv will contain all the available predictor filenames with
~80% per class marked as "TRAIN" and ~20% as "TEST":
$ echo "tt,class,selection" > tt-list.csv
$ for class in `ls data/predictors/`; do
ecoz2 util split --train-fraction 0.8 --file-ext .prd --files data/predictors/${class} >> tt-list.csv
done
The totals:
$ grep TRAIN tt-list.csv| wc -l
4277
$ grep TEST tt-list.csv| wc -l
1080
Using all TRAIN instances:
$ ecoz2 vq learn --prediction-order 36 --epsilon 0.0005 --predictors tt-list.csv
vq_learn: base_codebook_opt=None prediction_order=Some(36), epsilon=0.0005 codebook_class_name=_ predictor_filenames: 4277
Codebook generation:
prediction_order=36 class='_' epsilon=0.0005
393993 training vectors (ε=0.0005)
Report: data/codebooks/_/eps_0.0005.rpt
...
Quantize all vectors (TRAIN and TEST) using a number of the various codebook sizes:
$ for M in 0512 1024 2048 4096; do
ecoz2 vq quantize --codebook data/codebooks/_/eps_0.0005_M_${M}.cbook data/predictors
done
Resulting sequences get generated under data/sequences/M<M>/.
Eg.: data/sequences/M512/B/00123.seq
Running ./hmm-exercise.sh
Classification done using a new option --c12n (for "classification")
to generate report
c12n/TEST/N17__M4096_t3__a0.3_I1.csv
with all details about rankings:
ecoz2 hmm classify \
--c12n c12n/TEST/N17__M4096_t3__a0.3_I1.csv \
--models data/hmms/N17__M4096_t3__a0.3_I1 \
-M=4096 --tt=TEST --sequences tt-list.csv
Each data row in the report corresponds to a classification case. With one row as an example:
data/sequences/M4096/A/00007.seq, A, *, 1, A, C, I2, I3, I4, I, E, D, G2, F, C1, P, Bd, Bm, H, Bu, G, E1, B, M
The columns are:
seq_filename: the sequence filename (data/sequences/M4096/A/00007.seq), whose structure indicates the particular selection number (00007)seq_class_name: the label associated with the sequence (A)correct: a star*when the classification is correct, or!when incorrect (just as a quick visual element)rank: the rank of the model corresponding to the sequence (that is, with same class name), with 1 indicating a correct classification, 2 indicating it was 2nd place, etc.r1, ...: columns with the complete ranking of the models in order of decreasing probability.
Based on this report, and along with the original signal and labelling files, a python script was used to generate the plots below, which show the results associated with the signal for certain intervals.
c12n.plot.py --cover \
--signal ../MARS_20161221_000046_SongSession_16kHz_HPF5Hz.wav \
--segments ../exerc01/MARS_20161221_000046_SongSession_16kHz_HPF5HzNorm_labels.csv \
--c12n c12n/TEST/N17__M4096_t3__a0.3_I1.csv \
--out-prefix c12n/TEST/
This shows the location of the selections (delimited by gray vertical lines) on both the rank subplot at the top and the spectrogram subplot at the bottom. For each selection, the label on the x axis in the rank subplot indicates the selection number and the color and height of the bar reflect the rank of the most probable model: full green bar for rank 1 (meaning correct classification), and other colors and less tall bars for other ranks. For misclassifications, the label also shows the best model.
The above uses --cover to show the relevant interval in the original
signal containing the selections in the input file.
Alternatively, to put together the selections next to each other, the
--concat option can be used:
c12n.plot.py --concat \
--signal ../MARS_20161221_000046_SongSession_16kHz_HPF5Hz.wav \
--segments ../exerc01/MARS_20161221_000046_SongSession_16kHz_HPF5HzNorm_labels.csv \
--c12n c12n/TEST/N17__M4096_t3__a0.3_I1.csv \
--out-prefix c12n/TEST/
In this case, due to the large number of selections, the classification labels and the vertical lines delimiting the segments are not displayed.
To restrict the visualization to a particular class or rank, the options
--class and --rank can be used.
For example:
c12n.plot.py --concat --class Bd \
--signal ../MARS_20161221_000046_SongSession_16kHz_HPF5Hz.wav \
--segments ../exerc01/MARS_20161221_000046_SongSession_16kHz_HPF5HzNorm_labels.csv \
--c12n c12n/TEST/N17__M4096_t3__a0.3_I1.csv \
--out-prefix c12n/TEST/
A couple of more classes:
c12n.plot.py --concat --class A \
--signal ../MARS_20161221_000046_SongSession_16kHz_HPF5Hz.wav \
--segments ../exerc01/MARS_20161221_000046_SongSession_16kHz_HPF5HzNorm_labels.csv \
--c12n c12n/TEST/N17__M4096_t3__a0.3_I1.csv \
--out-prefix c12n/TEST/
c12n.plot.py --concat --class F \
--signal ../MARS_20161221_000046_SongSession_16kHz_HPF5Hz.wav \
--segments ../exerc01/MARS_20161221_000046_SongSession_16kHz_HPF5HzNorm_labels.csv \
--c12n c12n/TEST/N17__M4096_t3__a0.3_I1.csv \
--out-prefix c12n/TEST/
For possible reference, c12n/TRAIN/ contains some similar plots but for training instances.