# |
# V
#
# ******** Fully connected layer
# |
# V
#
# ABC Per-instrument classification neurons (outputs)
#
# We vary the following parameters to find the best accuracy.
# - MFCC computation time intervals: 5ms, 10ms, 20ms, 25ms (matching various of the intervals in papers above)
# - MFCC intervals in a sliding window presented to the neural network: 3, 4, 5
# (i.e. 15ms up to 125ms when multiplied by the time intervals).
# - Training batch size
instruments = InstrumentLoader(samplesDirPath, [])
Log("Max, min MFCC rows across all instruments: ", instruments.maxMfccRows,
instruments.minMfccRows)
Log("Number of instruments by length in MFCC rows:")
for k, v in sorted(instruments.mfccLenToSamplesMap.items()):
suffix = ''
if len(v) == 1:
suffix = '(' + os.path.basename(v[0].wavPath) + ')'
Log(" ", k, ": ", len(v), suffix)
if instruments.minWavHz < wavMinAllowedHz:
print(
"ERROR: One or more wav files found with rate in Hz less than configured minimum. Min found:",
instruments.minWavHz, " allowed min:", wavMinAllowedHz)
exit(1)
import keras.models
from KerasTensorFlowAnalyzer import KerasTensorFlowAnalyzer
import MfccComparisonAnalyzer
import os
from SoundStreamAnalyzer import SoundStreamAnalyzer
import sys
if len(sys.argv) < 5:
print('Usage:')
print(' <wav-file-path> <instruments-folder-path> <model-file-path> <model-params-json-path>')
exit(1)
wavFilePath = sys.argv[1]
instrumentsFolderPath = sys.argv[2]
modelFilePath = sys.argv[3]
modelParamsPath = sys.argv[4]
trainedModel = keras.models.load_model(modelFilePath)
# See SoundModelParams.py
f = open(modelParamsPath)
modelParams = json.load(f)
orderedResultInstrumentLabels = modelParams["instruments"]
print("Ordered labels:", orderedResultInstrumentLabels)
instruments = InstrumentLoader(instrumentsFolderPath, orderedResultInstrumentLabels)
analyzers = [ KerasTensorFlowAnalyzer(trainedModel, modelParams) ] + list(MfccComparisonAnalyzer.constructFromInstruments(instruments))
analyzer = SoundStreamAnalyzer(wavFilePath, instruments, analyzers, 0.9)
analyzer.getMatches()
