12345678910111213141516171819202122232425262728293031323334353637383940414243 |
- README
- for codec2/unittest
- Created David Rowe 31 July 2012
- Training (experimental) sparse phase VQs:
- 1/ In ../src/phase.c phase_experiment() enable:
- print_sparse_pred_error()
- and 'make' c2sim (in src)
- 2/ Run over a training database:
- $ ./c2sim /xhome1/codec2/samples/train.spc --phaseexp > train_phtrain.txt
- a) check stats in Octave:
- octave> load ../src/train_phtrain.txt
- octave> std(nonzeros(train_phtrain(:,1:20)))
- octave> hist(nonzeros(train_phtrain(:,1:20)),20)
- 3/ Extract and convert to floats vector you wish to train for example
- first 20 (out of MAX_AMP == 80):
- $ ./extract ../src/train_phtrain.txt train_phtrain.flt 1 20
- 4/ Convert to rectangular:
- $ ./polar2rect train_phtrain.flt train_phtrainr.flt
- 5/ Run this program:
- $ ./vqtrainph train_phtrainr.flt 20 1024 vq.txt
- Ouput is vq.txt
- Tests
- -----
- + build up insmallest possible stesp
- + impl errors v alg errors
- + use actual phase data as codebook
- + test vq with rand phases first or known data
|