def createPrediction(trainedGMM, jsonFile, silenceClassNum):
"""
Create prediction with a 2s majority vote. Like in the Android app, 2s windows where all
amplitude values are below the threshold will be ignored, and no prediction will be made.
@param trainedGMM: already trained GMM
@param jsonFile: path the json file that contains the features and amplitude values. This
file has to have the features under the key "features" and the amplitude values under the
key "amps"
@param silenceClassNum: The class number to which silent sequences will be assigned
"""
n_classes = len(trainedGMM['clfs'])
jsonData = json.load(open(jsonFile, "rb"))
featureData = np.array(jsonData["features"])
amps = np.array(jsonData["amps"])
X_test = trainedGMM['scaler'].transform(featureData)
logLikelihood = np.zeros((n_classes, X_test.shape[0]))
""" Compute log-probability for each class for all points: """
for i in range(n_classes):
logLikelihood[i] = logProb(X_test, trainedGMM['clfs'][i].weights_,
trainedGMM['clfs'][i].means_, trainedGMM['clfs'][i].covars_)
""" Select the class with the highest log-probability: """
y_pred = np.argmax(logLikelihood, 0)
return majorityVoteSilence(y_pred, amps, silenceClassNum)
def makePrediction(trainedGMM, evalFeatures, evalAmps, silenceClassNum):
"""
@param trainedGMM:
@param evalFeatures: not scaled!!
@param evalAmps: Amplitude values
@param silenceClassNum:
@return:
"""
n_classes = len(trainedGMM['clfs'])
X_test = trainedGMM['scaler'].transform(evalFeatures)
logLikelihood = np.zeros((n_classes, X_test.shape[0]))
user_component_matrix = np.empty((n_classes, X_test.shape[0]), dtype=np.bool_)
""" Compute log-probability for each class for all points: """
for i in range(n_classes):
logLikelihood[i], tmp_comp_matrix = logProb(X_test, trainedGMM['clfs'][i].weights_,
trainedGMM['clfs'][i].means_, trainedGMM['clfs'][i].covars_, return_component_matrix=True)
user_component_matrix[i] = tmp_comp_matrix
""" Select the class with the highest log-probability: """
y_pred = np.argmax(logLikelihood, 0)
# Number of predictions where most likely component from user-centric model:
n_user_component = 0
for i in range(X_test.shape[0]):
if user_component_matrix[y_pred[i],i] == True:
n_user_component += 1
percent_user_component = round(100 * (n_user_component / float(X_test.shape[0])), 2)
print(str(percent_user_component) + "% of all predictions had component from user" +
" as most likely component")
return majorityVoteSilence(y_pred, evalAmps, silenceClassNum), percent_user_component
def makePrediction(trainedGMM, evalFeatures, evalAmps, silenceClassNum):
"""
@param trainedGMM:
@param evalFeatures: not scaled!!
@param evalAmps: Amplitude values
@param silenceClassNum:
@return:
"""
n_classes = len(trainedGMM['clfs'])
X_test = trainedGMM['scaler'].transform(evalFeatures)
logLikelihood = np.zeros((n_classes, X_test.shape[0]))
""" Compute log-probability for each class for all points: """
for i in range(n_classes):
logLikelihood[i] = logProb(X_test, trainedGMM['clfs'][i].weights_,
trainedGMM['clfs'][i].means_, trainedGMM['clfs'][i].covars_)
""" Select the class with the highest log-probability: """
y_pred = np.argmax(logLikelihood, 0)
return majorityVoteSilence(y_pred, evalAmps, silenceClassNum)
