def getSingleFeatureLineFromFile(file, decisions, shot, leave_out_class=None):
"""
This is a less troublesome but slow method to get a featureLine.
"""
beatList, context = getContextAndBeatListFromFile(file)
blockList = coalesceBeats(beatList)
Features.initializeContextVars(context)
lastShotId, context, blockList = applyDecisionsToBeatscript(context, blockList,
decisions)
featureLine = getFeatureLine(context, blockList[len(decisions)], shot, lastShotId,
leave_out_class)
return featureLine
def getFeatureLinesFromFileAndModify(file, shot, leave_out=-1):
"""
Creates featureLines by blowing up the data with duplicating SAYS in some
cases. This rapidly increases training times.
"""
context, beatList = getContextAndBeatListFromFile(file)
originalList = deepcopy(beatList)
for i in range(0, 2):
for beat in originalList:
beatList.append(beat)
if beat.type == SAYS:
if random.randint(0, 1):
beatList.append(beat)
return createFeatureLines(context, beatList, shot, leave_out)
def onlineFeatureLineCreator(filename, use_classified_shot = False, use_history = True):
# load context and complete beatlist from file
context, beatList = getContextAndBeatListFromFile(filename)
Features.initializeContextVars(context)
blockList = coalesceBeats(beatList)
context["BygoneBlocks"] = []
for block in blockList:
shot_true = block[-1].shot
# get current feature line and true shot class
featureLine = getFeatureLine(context, block, True, -1)
features = np.array(featureLine[:-1], dtype=np.float64)
shot_classified = yield features, shot_true
# update block and lastShotId
if use_classified_shot:
for beat in block:
beat.shot = shot_classified
if use_history:
context["BygoneBlocks"].append(block)
def testAllButFile(file, files, scaler, return_queue, fake_decisions=False, C=None,
gamma=None, leave_out_class=None):
"""
This function trains with all files in files except file, which is used for
testing. The performance of the test is returned.
"""
training_set = [f for f in files if f != file]
training_data, training_data_classes = getDataMatrix(training_set,
leave_out_class=leave_out_class, shot=True)
training_data = scaler.transform(training_data, training_data_classes)
trained_svm = trainSVM(training_data, training_data_classes, C=C, gamma=gamma)
context, beatList = getContextAndBeatListFromFile(file)
blockList = coalesceBeats(beatList)
part_blockList = []
decisions = []
correct_classification_count = 0
medium_shot_count = 0
metric_sum = 0
correct_histogram = [0, 0, 0, 0, 0, 0, 0]
guessed_histogram = [0, 0, 0, 0, 0, 0, 0]
#correct_histogram = [0, 0]
#guessed_histogram = [0, 0]
#last_block = None
for block in blockList:
# prepare block-list and decision-list
part_blockList.append(block)
if fake_decisions:
decisions = []
for i in range(len(part_blockList)-1):
decisions.append(part_blockList[i][-1].shot)
svm_distribution, svm_classification = calculateDistributionAndClassification(
trained_svm, deepcopy(context), part_blockList, decisions, scaler,
shot_or_cut=True, leave_out_class=leave_out_class)
if not fake_decisions:
decisions.append(svm_classification)
guessed_histogram[svm_classification] += 1
correct_histogram[block[-1].shot] += 1
#is_shot = True
#if last_block:
# is_shot = block[-1].shotId != last_block[-1].shotId
#correct_histogram[int(is_shot)] += 1
if svm_classification == block[-1].shot:
#if boost_classification == int(is_shot):
correct_classification_count += 1
if block[-1].shot == 2: medium_shot_count += 1
if len(part_blockList) >= 2:
previous_correct_class = part_blockList[-2][-1].shot
if len(decisions) >= 2:
previous_guessed_class = decisions[-2]
else: previous_guessed_class = previous_correct_class
else:
previous_correct_class = part_blockList[-1][-1].shot
if len(decisions) >= 1:
previous_guessed_class = decisions[-1]
else: previous_guessed_class = previous_correct_class
metric_sum += pointMetric(svm_classification, block[-1].shot,
previous_guessed_class, previous_correct_class)
#last_block = block
performance = float(correct_classification_count)/len(blockList)
medium_shot_performance = float(medium_shot_count)/len(blockList)
return_queue.put((
correct_histogram, guessed_histogram, performance, medium_shot_performance,
float(metric_sum) / len(blockList)))
return_queue.close()
def XValidation(files, fake_decisions = False):
"""
Since the decisions of the classifiers during classifying a beatscript are used this
is not a classical cross-validation. Instead the training is done with all but one
Training files and the remaining beatscript is tested based on the classification from
that data. This process is repeated with all files.
In this case the decision history is faked by using the original classes from the
testfile.
This function tests the performance for decisions using a SVM, each with faked History.
"""
reference_data, _ = getDataMatrix(TRAIN_FILES)
scaler = preprocessing.Scaler()
scaler.fit(reference_data)
correct_histogram = [0, 0, 0, 0, 0, 0, 0]
guessed_histogram = [0, 0, 0, 0, 0, 0, 0]
performances = []
allover_point_sum = 0.0
medium_shot_performances = []
for file in files:
print("X-Validation: ca. " +
str(int(round(float(files.index(file)) / len(files) * 100))) +
"% fertig.")
training_set = [f for f in files if f != file]
training_data, training_data_classes = getDataMatrix(training_set)
training_data = scaler.transform(training_data, training_data_classes)
print("Trainingsdaten erzeugt. Trainiere Classifier...")
print_lock = Lock()
svm_queue = Queue(maxsize=1)
svm_learning_process = Process(target=trainSVM,
args=(training_data, training_data_classes, svm_queue, print_lock))
svm_learning_process.start()
context, beatList = getContextAndBeatListFromFile(file)
blockList = coalesceBeats(beatList)
part_blockList = []
decisions = []
correct_classification_count = 0
medium_shot_count = 0
metric_sum = 0
trained_svm = svm_queue.get()
svm_learning_process.join()
print("Training finished for: " + file)
for block in blockList:
# prepare blocklist and decision-list
part_blockList.append(block)
if fake_decisions:
decisions = []
for i in range(len(part_blockList)-1):
decisions.append(part_blockList[i][-1].shot)
svm_queue = Queue(maxsize=1)
svm_classification_process = Process(
target=calculateDistributionAndClassification,
args=(
trained_svm, deepcopy(context), part_blockList, decisions, scaler, True,
svm_queue))
svm_classification_process.start()
svm_distribution, svm_classification = svm_queue.get()
svm_classification_process.join()
if not fake_decisions:
decisions.append(svm_classification)
print("SVM Classification:\t" + SHOT_NAMES[svm_classification])
guessed_histogram[svm_classification] += 1
print("Correct Class:\t\t" + SHOT_NAMES[block[-1].shot])
if len(part_blockList)>= 2:
previous_correct_class = part_blockList[-2][-1].shot
if len(decisions) >= 2:
previous_guessed_class = decisions[-2]
else: previous_guessed_class = previous_correct_class
else:
previous_correct_class = part_blockList[-1][-1].shot
if len(decisions) >= 1:
previous_guessed_class = decisions[-1]
else: previous_guessed_class = previous_correct_class
metric_value = pointMetric(svm_classification, block[-1].shot,
previous_guessed_class, previous_correct_class)
print("Wrongness:\t\t\t" + str(metric_value))
metric_sum += metric_value
correct_histogram[block[-1].shot] += 1
if svm_classification == block[-1].shot:
correct_classification_count += 1
if block[-1].shot == 2: medium_shot_count += 1
print("------------------------------------")
print("File Performance: " + str(
float(correct_classification_count) / len(blockList) * 100) + "%")
print(
"File Wrongness: " + str(float(metric_sum) / len(blockList)) + " Points ( 0 - 5 )")
performances.append(float(correct_classification_count) / len(blockList))
medium_shot_performances.append(float(medium_shot_count) / len(blockList))
allover_point_sum += float(metric_sum) / len(blockList)
print("__________________________________________")
performance_sum = 0
performance_best = 0
performance_last = 1
for p in medium_shot_performances:
performance_sum += p
if p > performance_best: performance_best = p
if p < performance_last: performance_last = p
print("MS-Performance:\t" + str(performance_sum / len(performances) * 100.0) + "%\t(" +
str(performance_last) + " - " + str(performance_best) + ")")
performance_sum = 0
performance_best = 0
performance_last = 1
for p in performances:
performance_sum += p
if p > performance_best: performance_best = p
if p < performance_last: performance_last = p
print("Performance:\t" + str(performance_sum / len(performances) * 100.0) + "%\t(" +
str(performance_last) + " - " + str(performance_best) + ")")
print("Wrongness:\t" + str(allover_point_sum / len(performances)))
return allover_point_sum / len(performances)
def getFeatureLinesFromFile(file, shot, leave_out_class=None):
"""
Returns a list of featureLines converted from the beatscript given in file.
"""
context, beatList = getContextAndBeatListFromFile(file)
return createFeatureLines(context, beatList, shot, leave_out_class)