def __init__(self):
super(DBRDPreprocessing, self).__init__(
MultiLineTokenizer(),
SnowballStemmer('english', ignore_stopwords=True),
set(
stopwords.words('english') + list(string.punctuation) +
["n't", "'t"]), [HTMLSymbolFilter()])
def main(_run, _config, _seed, _log):
"""
:param _run:
:param _config:
:param _seed:
:param _log:
:return:
"""
"""
Setting and loading parameters
"""
# Setting logger
args = _config
logger = _log
logger.info(args)
logger.info('It started at: %s' % datetime.now())
torch.manual_seed(_seed)
bugReportDatabase = BugReportDatabase.fromJson(args['bug_database'])
paddingSym = "</s>"
batchSize = args['batch_size']
device = torch.device('cuda' if args['cuda'] else "cpu")
if args['cuda']:
logger.info("Turning CUDA on")
else:
logger.info("Turning CUDA off")
# It is the folder where the preprocessed information will be stored.
cacheFolder = args['cache_folder']
# Setting the parameter to save and loading parameters
importantParameters = ['compare_aggregation', 'categorical']
parametersToSave = dict([(parName, args[parName])
for parName in importantParameters])
if args['load'] is not None:
mapLocation = (
lambda storage, loc: storage.cuda()) if args['cuda'] else 'cpu'
modelInfo = torch.load(args['load'], map_location=mapLocation)
modelState = modelInfo['model']
for paramName, paramValue in modelInfo['params'].items():
args[paramName] = paramValue
else:
modelState = None
preprocessors = PreprocessorList()
inputHandlers = []
categoricalOpt = args.get('categorical')
if categoricalOpt is not None and len(categoricalOpt) != 0:
categoricalEncoder, _, _ = processCategoricalParam(
categoricalOpt, bugReportDatabase, inputHandlers, preprocessors,
None, logger)
else:
categoricalEncoder = None
filterInputHandlers = []
compareAggOpt = args['compare_aggregation']
databasePath = args['bug_database']
# Loading word embedding
if compareAggOpt["lexicon"]:
emb = np.load(compareAggOpt["word_embedding"])
lexicon = Lexicon(unknownSymbol=None)
with codecs.open(compareAggOpt["lexicon"]) as f:
for l in f:
lexicon.put(l.strip())
lexicon.setUnknown("UUUKNNN")
paddingId = lexicon.getLexiconIndex(paddingSym)
embedding = Embedding(lexicon, emb, paddingIdx=paddingId)
logger.info("Lexicon size: %d" % (lexicon.getLen()))
logger.info("Word Embedding size: %d" % (embedding.getEmbeddingSize()))
elif compareAggOpt["word_embedding"]:
# todo: Allow use embeddings and other representation
lexicon, embedding = Embedding.fromFile(
compareAggOpt['word_embedding'],
'UUUKNNN',
hasHeader=False,
paddingSym=paddingSym)
logger.info("Lexicon size: %d" % (lexicon.getLen()))
logger.info("Word Embedding size: %d" % (embedding.getEmbeddingSize()))
paddingId = lexicon.getLexiconIndex(paddingSym)
else:
embedding = None
if compareAggOpt["norm_word_embedding"]:
embedding.zscoreNormalization()
# Tokenizer
if compareAggOpt['tokenizer'] == 'default':
logger.info("Use default tokenizer to tokenize summary information")
tokenizer = MultiLineTokenizer()
elif compareAggOpt['tokenizer'] == 'white_space':
logger.info(
"Use white space tokenizer to tokenize summary information")
tokenizer = WhitespaceTokenizer()
else:
raise ArgumentError(
"Tokenizer value %s is invalid. You should choose one of these: default and white_space"
% compareAggOpt['tokenizer'])
# Preparing input handlers, preprocessors and cache
minSeqSize = max(compareAggOpt['aggregate']["window"]
) if compareAggOpt['aggregate']["model"] == "cnn" else -1
bow = compareAggOpt.get('bow', False)
freq = compareAggOpt.get('frequency', False) and bow
logger.info("BoW={} and TF={}".format(bow, freq))
if compareAggOpt['extractor'] is not None:
# Use summary and description (concatenated) to address this problem
logger.info("Using Summary and Description information.")
# Loading Filters
extractorFilters = loadFilters(compareAggOpt['extractor']['filters'])
arguments = (databasePath, compareAggOpt['word_embedding'],
str(compareAggOpt['lexicon']), ' '.join(
sorted([
fil.__class__.__name__ for fil in extractorFilters
])), compareAggOpt['tokenizer'], str(bow), str(freq),
SABDEncoderPreprocessor.__name__)
inputHandlers.append(SABDInputHandler(paddingId, minSeqSize))
extractorCache = PreprocessingCache(cacheFolder, arguments)
if bow:
extractorPreprocessor = SABDBoWPreprocessor(
lexicon, bugReportDatabase, extractorFilters, tokenizer,
paddingId, freq, extractorCache)
else:
extractorPreprocessor = SABDEncoderPreprocessor(
lexicon, bugReportDatabase, extractorFilters, tokenizer,
paddingId, extractorCache)
preprocessors.append(extractorPreprocessor)
# Create model
model = SABD(embedding, categoricalEncoder, compareAggOpt['extractor'],
compareAggOpt['matching'], compareAggOpt['aggregate'],
compareAggOpt['classifier'], freq)
if args['loss'] == 'bce':
logger.info("Using BCE Loss: margin={}".format(args['margin']))
lossFn = BCELoss()
lossNoReduction = BCELoss(reduction='none')
cmp_collate = PairBugCollate(inputHandlers,
torch.float32,
unsqueeze_target=True)
elif args['loss'] == 'triplet':
logger.info("Using Triplet Loss: margin={}".format(args['margin']))
lossFn = TripletLoss(args['margin'])
lossNoReduction = TripletLoss(args['margin'], reduction='none')
cmp_collate = TripletBugCollate(inputHandlers)
model.to(device)
if modelState:
model.load_state_dict(modelState)
"""
Loading the training and validation. Also, it sets how the negative example will be generated.
"""
# load training
if args.get('pairs_training'):
negativePairGenOpt = args.get('neg_pair_generator', )
trainingFile = args.get('pairs_training')
offlineGeneration = not (negativePairGenOpt is None
or negativePairGenOpt['type'] == 'none')
masterIdByBugId = bugReportDatabase.getMasterIdByBugId()
randomAnchor = negativePairGenOpt['random_anchor']
if not offlineGeneration:
logger.info("Not generate dynamically the negative examples.")
negativePairGenerator = None
else:
pairGenType = negativePairGenOpt['type']
if pairGenType == 'random':
logger.info("Random Negative Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = RandomGenerator(
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
randomAnchor=randomAnchor)
elif pairGenType == 'non_negative':
logger.info("Non Negative Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = NonNegativeRandomGenerator(
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
elif pairGenType == 'misc_non_zero':
logger.info("Misc Non Zero Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = MiscNonZeroRandomGen(
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
bugIds,
trainingDataset.duplicateIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
elif pairGenType == 'product_component':
logger.info("Product Component Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = ProductComponentRandomGen(
bugReportDatabase,
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
elif pairGenType == 'random_k':
logger.info("Random K Negative Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = KRandomGenerator(
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['k'],
device,
randomAnchor=randomAnchor)
elif pairGenType == "pre":
logger.info("Pre-selected list generator")
negativePairGenerator = PreSelectedGenerator(
negativePairGenOpt['pre_list_file'],
preprocessors,
negativePairGenOpt['rate'],
masterIdByBugId,
negativePairGenOpt['preselected_length'],
randomAnchor=randomAnchor)
elif pairGenType == "positive_pre":
logger.info("Positive Pre-selected list generator")
negativePairGenerator = PositivePreSelectedGenerator(
negativePairGenOpt['pre_list_file'],
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
masterIdByBugId,
negativePairGenOpt['preselected_length'],
randomAnchor=randomAnchor)
elif pairGenType == "misc_non_zero_pre":
logger.info("Misc: non-zero and Pre-selected list generator")
negativePairGenerator1 = PreSelectedGenerator(
negativePairGenOpt['pre_list_file'],
preprocessors,
negativePairGenOpt['rate'],
masterIdByBugId,
negativePairGenOpt['preselected_length'],
randomAnchor=randomAnchor)
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
negativePairGenerator2 = NonNegativeRandomGenerator(
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
negativePairGenerator = MiscOfflineGenerator(
(negativePairGenerator1, negativePairGenerator2))
elif pairGenType == "misc_non_zero_positive_pre":
logger.info(
"Misc: non-zero and Positive Pre-selected list generator")
negativePairGenerator1 = PositivePreSelectedGenerator(
negativePairGenOpt['pre_list_file'],
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
masterIdByBugId,
negativePairGenOpt['preselected_length'],
randomAnchor=randomAnchor)
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
negativePairGenerator2 = NonNegativeRandomGenerator(
preprocessors,
cmp_collate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
negativePairGenerator = MiscOfflineGenerator(
(negativePairGenerator1, negativePairGenerator2))
else:
raise ArgumentError(
"Offline generator is invalid (%s). You should choose one of these: random, hard and pre"
% pairGenType)
if isinstance(lossFn, BCELoss):
training_reader = PairBugDatasetReader(
trainingFile,
preprocessors,
negativePairGenerator,
randomInvertPair=args['random_switch'])
elif isinstance(lossFn, TripletLoss):
training_reader = TripletBugDatasetReader(
trainingFile,
preprocessors,
negativePairGenerator,
randomInvertPair=args['random_switch'])
trainingLoader = DataLoader(training_reader,
batch_size=batchSize,
collate_fn=cmp_collate.collate,
shuffle=True)
logger.info("Training size: %s" % (len(trainingLoader.dataset)))
# load validation
if args.get('pairs_validation'):
if isinstance(lossFn, BCELoss):
validation_reader = PairBugDatasetReader(
args.get('pairs_validation'), preprocessors)
elif isinstance(lossFn, TripletLoss):
validation_reader = TripletBugDatasetReader(
args.get('pairs_validation'), preprocessors)
validationLoader = DataLoader(validation_reader,
batch_size=batchSize,
collate_fn=cmp_collate.collate)
logger.info("Validation size: %s" % (len(validationLoader.dataset)))
else:
validationLoader = None
"""
Training and evaluate the model.
"""
optimizer_opt = args.get('optimizer', 'adam')
if optimizer_opt == 'sgd':
logger.info('SGD')
optimizer = optim.SGD(model.parameters(),
lr=args['lr'],
weight_decay=args['l2'])
elif optimizer_opt == 'adam':
logger.info('Adam')
optimizer = optim.Adam(model.parameters(),
lr=args['lr'],
weight_decay=args['l2'])
# Recall rate
rankingScorer = GeneralScorer(
model, preprocessors, device,
PairBugCollate(inputHandlers, ignore_target=True),
args['ranking_batch_size'], args['ranking_n_workers'])
recallEstimationTrainOpt = args.get('recall_estimation_train')
if recallEstimationTrainOpt:
preselectListRankingTrain = PreselectListRanking(
recallEstimationTrainOpt, args['sample_size_rr_tr'])
recallEstimationOpt = args.get('recall_estimation')
if recallEstimationOpt:
preselectListRanking = PreselectListRanking(recallEstimationOpt,
args['sample_size_rr_val'])
# LR scheduler
lrSchedulerOpt = args.get('lr_scheduler', None)
if lrSchedulerOpt is None:
logger.info("Scheduler: Constant")
lrSched = None
elif lrSchedulerOpt["type"] == 'step':
logger.info("Scheduler: StepLR (step:%s, decay:%f)" %
(lrSchedulerOpt["step_size"], args["decay"]))
lrSched = StepLR(optimizer, lrSchedulerOpt["step_size"],
lrSchedulerOpt["decay"])
elif lrSchedulerOpt["type"] == 'exp':
logger.info("Scheduler: ExponentialLR (decay:%f)" %
(lrSchedulerOpt["decay"]))
lrSched = ExponentialLR(optimizer, lrSchedulerOpt["decay"])
elif lrSchedulerOpt["type"] == 'linear':
logger.info(
"Scheduler: Divide by (1 + epoch * decay) ---- (decay:%f)" %
(lrSchedulerOpt["decay"]))
lrDecay = lrSchedulerOpt["decay"]
lrSched = LambdaLR(optimizer, lambda epoch: 1 /
(1.0 + epoch * lrDecay))
else:
raise ArgumentError(
"LR Scheduler is invalid (%s). You should choose one of these: step, exp and linear "
% pairGenType)
# Set training functions
def trainingIteration(engine, batch):
engine.kk = 0
model.train()
optimizer.zero_grad()
x, y = cmp_collate.to(batch, device)
output = model(*x)
loss = lossFn(output, y)
loss.backward()
optimizer.step()
return loss, output, y
def scoreDistanceTrans(output):
if len(output) == 3:
_, y_pred, y = output
else:
y_pred, y = output
if lossFn == F.nll_loss:
return torch.exp(y_pred[:, 1]), y
elif isinstance(lossFn, (BCELoss)):
return y_pred, y
trainer = Engine(trainingIteration)
trainingMetrics = {'training_loss': AverageLoss(lossFn)}
if isinstance(lossFn, BCELoss):
trainingMetrics['training_dist_target'] = MeanScoreDistance(
output_transform=scoreDistanceTrans)
trainingMetrics['training_acc'] = AccuracyWrapper(
output_transform=thresholded_output_transform)
trainingMetrics['training_precision'] = PrecisionWrapper(
output_transform=thresholded_output_transform)
trainingMetrics['training_recall'] = RecallWrapper(
output_transform=thresholded_output_transform)
# Add metrics to trainer
for name, metric in trainingMetrics.items():
metric.attach(trainer, name)
# Set validation functions
def validationIteration(engine, batch):
if not hasattr(engine, 'kk'):
engine.kk = 0
model.eval()
with torch.no_grad():
x, y = cmp_collate.to(batch, device)
y_pred = model(*x)
return y_pred, y
validationMetrics = {
'validation_loss':
LossWrapper(lossFn,
output_transform=lambda x: (x[0], x[0][0])
if x[1] is None else x)
}
if isinstance(lossFn, BCELoss):
validationMetrics['validation_dist_target'] = MeanScoreDistance(
output_transform=scoreDistanceTrans)
validationMetrics['validation_acc'] = AccuracyWrapper(
output_transform=thresholded_output_transform)
validationMetrics['validation_precision'] = PrecisionWrapper(
output_transform=thresholded_output_transform)
validationMetrics['validation_recall'] = RecallWrapper(
output_transform=thresholded_output_transform)
evaluator = Engine(validationIteration)
# Add metrics to evaluator
for name, metric in validationMetrics.items():
metric.attach(evaluator, name)
# recommendation
recommendation_fn = generateRecommendationList
@trainer.on(Events.EPOCH_STARTED)
def onStartEpoch(engine):
epoch = engine.state.epoch
logger.info("Epoch: %d" % epoch)
if lrSched:
lrSched.step()
logger.info("LR: %s" % str(optimizer.param_groups[0]["lr"]))
@trainer.on(Events.EPOCH_COMPLETED)
def onEndEpoch(engine):
epoch = engine.state.epoch
logMetrics(_run, logger, engine.state.metrics, epoch)
# Evaluate Training
if validationLoader:
evaluator.run(validationLoader)
logMetrics(_run, logger, evaluator.state.metrics, epoch)
lastEpoch = args['epochs'] - epoch == 0
if recallEstimationTrainOpt and (epoch % args['rr_train_epoch'] == 0):
logRankingResult(_run,
logger,
preselectListRankingTrain,
rankingScorer,
bugReportDatabase,
None,
epoch,
"train",
recommendationListfn=recommendation_fn)
rankingScorer.free()
if recallEstimationOpt and (epoch % args['rr_val_epoch'] == 0):
logRankingResult(_run,
logger,
preselectListRanking,
rankingScorer,
bugReportDatabase,
args.get("ranking_result_file"),
epoch,
"validation",
recommendationListfn=recommendation_fn)
rankingScorer.free()
if not lastEpoch:
training_reader.sampleNewNegExamples(model, lossNoReduction)
if args.get('save'):
save_by_epoch = args['save_by_epoch']
if save_by_epoch and epoch in save_by_epoch:
file_name, file_extension = os.path.splitext(args['save'])
file_path = file_name + '_epoch_{}'.format(
epoch) + file_extension
else:
file_path = args['save']
modelInfo = {
'model': model.state_dict(),
'params': parametersToSave
}
logger.info("==> Saving Model: %s" % file_path)
torch.save(modelInfo, file_path)
if args.get('pairs_training'):
trainer.run(trainingLoader, max_epochs=args['epochs'])
elif args.get('pairs_validation'):
# Evaluate Training
evaluator.run(validationLoader)
logMetrics(_run, logger, evaluator.state.metrics, 0)
if recallEstimationOpt:
logRankingResult(_run,
logger,
preselectListRanking,
rankingScorer,
bugReportDatabase,
args.get("ranking_result_file"),
0,
"validation",
recommendationListfn=recommendation_fn)
# Test Dataset (accuracy, recall, precision, F1)
pair_test_dataset = args.get('pair_test_dataset')
if pair_test_dataset is not None and len(pair_test_dataset) > 0:
pairTestReader = PairBugDatasetReader(pair_test_dataset, preprocessors)
testLoader = DataLoader(pairTestReader,
batch_size=batchSize,
collate_fn=cmp_collate.collate)
if not isinstance(cmp_collate, PairBugCollate):
raise NotImplementedError(
'Evaluation of pairs using tanh was not implemented yet')
logger.info("Test size: %s" % (len(testLoader.dataset)))
testMetrics = {
'test_accuracy':
ignite.metrics.Accuracy(
output_transform=thresholded_output_transform),
'test_precision':
ignite.metrics.Precision(
output_transform=thresholded_output_transform),
'test_recall':
ignite.metrics.Recall(
output_transform=thresholded_output_transform),
'test_predictions':
PredictionCache(),
}
test_evaluator = Engine(validationIteration)
# Add metrics to evaluator
for name, metric in testMetrics.items():
metric.attach(test_evaluator, name)
test_evaluator.run(testLoader)
for metricName, metricValue in test_evaluator.state.metrics.items():
metric = testMetrics[metricName]
if isinstance(metric, ignite.metrics.Accuracy):
logger.info({
'type': 'metric',
'label': metricName,
'value': metricValue,
'epoch': None,
'correct': metric._num_correct,
'total': metric._num_examples
})
_run.log_scalar(metricName, metricValue)
elif isinstance(metric,
(ignite.metrics.Precision, ignite.metrics.Recall)):
logger.info({
'type': 'metric',
'label': metricName,
'value': metricValue,
'epoch': None,
'tp': metric._true_positives.item(),
'total_positive': metric._positives.item()
})
_run.log_scalar(metricName, metricValue)
elif isinstance(metric, ConfusionMatrix):
acc = cmAccuracy(metricValue)
prec = cmPrecision(metricValue, False)
recall = cmRecall(metricValue, False)
f1 = 2 * (prec * recall) / (prec + recall + 1e-15)
logger.info({
'type':
'metric',
'label':
metricName,
'accuracy':
np.float(acc),
'precision':
prec.cpu().numpy().tolist(),
'recall':
recall.cpu().numpy().tolist(),
'f1':
f1.cpu().numpy().tolist(),
'confusion_matrix':
metricValue.cpu().numpy().tolist(),
'epoch':
None
})
_run.log_scalar('test_f1', f1[1])
elif isinstance(metric, PredictionCache):
logger.info({
'type': 'metric',
'label': metricName,
'predictions': metric.predictions
})
# Calculate recall rate
recallRateOpt = args.get('recall_rate', {'type': 'none'})
if recallRateOpt['type'] != 'none':
if recallRateOpt['type'] == 'sun2011':
logger.info("Calculating recall rate: {}".format(
recallRateOpt['type']))
recallRateDataset = BugDataset(recallRateOpt['dataset'])
rankingClass = SunRanking(bugReportDatabase, recallRateDataset,
recallRateOpt['window'])
# We always group all bug reports by master in the results in the sun 2011 methodology
group_by_master = True
elif recallRateOpt['type'] == 'deshmukh':
logger.info("Calculating recall rate: {}".format(
recallRateOpt['type']))
recallRateDataset = BugDataset(recallRateOpt['dataset'])
rankingClass = DeshmukhRanking(bugReportDatabase,
recallRateDataset)
group_by_master = recallRateOpt['group_by_master']
else:
raise ArgumentError(
"recall_rate.type is invalid (%s). You should choose one of these: step, exp and linear "
% recallRateOpt['type'])
logRankingResult(_run,
logger,
rankingClass,
rankingScorer,
bugReportDatabase,
recallRateOpt["result_file"],
0,
None,
group_by_master,
recommendationListfn=recommendation_fn)
def processSumDescParam(sum_desc_opts, bugReportDatabase, inputHandlers, preprocessors, encoders, cacheFolder,
databasePath, logger, paddingSym):
# Use summary and description (concatenated) to address this problem
logger.info("Using summary and description information.")
# Loading word embedding
lexicon, embedding = load_embedding(sum_desc_opts, paddingSym)
logger.info("Lexicon size: %d" % (lexicon.getLen()))
logger.info("Word Embedding size: %d" % (embedding.getEmbeddingSize()))
paddingId = lexicon.getLexiconIndex(paddingSym)
# Loading Filters
filters = loadFilters(sum_desc_opts['filters'])
# Tokenizer
if sum_desc_opts['tokenizer'] == 'default':
logger.info("Use default tokenizer to tokenize summary+description information")
tokenizer = MultiLineTokenizer()
elif sum_desc_opts['tokenizer'] == 'white_space':
logger.info("Use white space tokenizer to tokenize summary+description information")
tokenizer = WhitespaceTokenizer()
else:
raise ArgumentError(
"Tokenizer value %s is invalid. You should choose one of these: default and white_space" %
sum_desc_opts['tokenizer'])
arguments = (
databasePath, sum_desc_opts['word_embedding'],
' '.join(sorted([fil.__class__.__name__ for fil in filters])),
sum_desc_opts['tokenizer'], "summary_description")
cacheSumDesc = PreprocessingCache(cacheFolder, arguments)
sumDescPreprocessor = SummaryDescriptionPreprocessor(lexicon, bugReportDatabase, filters, tokenizer, paddingId, cacheSumDesc)
preprocessors.append(sumDescPreprocessor)
if sum_desc_opts['encoder_type'] == 'cnn':
windowSizes = sum_desc_opts.get('window_sizes', [3])
nFilters = sum_desc_opts.get('nfilters', 100)
updateEmb = sum_desc_opts.get('update_embedding', False)
actFunc = loadActivationFunction(sum_desc_opts.get('activation', 'relu'))
batchNorm = sum_desc_opts.get('batch_normalization', False)
dropout = sum_desc_opts.get('dropout', 0.0)
sumDescEncoder = TextCNN(windowSizes, nFilters, embedding, updateEmb, actFunc, batchNorm, dropout)
encoders.append(sumDescEncoder)
inputHandlers.append(TextCNNInputHandler(paddingId, max(windowSizes)))
elif sum_desc_opts['encoder_type'] == 'cnn+dense':
windowSizes = sum_desc_opts.get('window_sizes', [3])
nFilters = sum_desc_opts.get('nfilters', 100)
updateEmb = sum_desc_opts.get('update_embedding', False)
actFunc = loadActivationFunction(sum_desc_opts.get('activation', 'relu'))
batchNorm = sum_desc_opts.get('batch_normalization', False)
dropout = sum_desc_opts.get('dropout', 0.0)
hiddenSizes = sum_desc_opts.get('hidden_sizes')
hiddenAct = loadActivationClass(sum_desc_opts.get('hidden_act'))
hiddenDropout = sum_desc_opts.get('hidden_dropout')
batchLast = sum_desc_opts.get("bn_last_layer", False)
cnnEnc = TextCNN(windowSizes, nFilters, embedding, updateEmb, actFunc, batchNorm, dropout)
sumDescEncoder = MultilayerDense(cnnEnc, hiddenSizes, hiddenAct, batchNorm, batchLast, hiddenDropout)
encoders.append(sumDescEncoder)
inputHandlers.append(TextCNNInputHandler(paddingId, max(windowSizes)))
elif sum_desc_opts['encoder_type'] == 'word_mean':
standardization = sum_desc_opts.get('standardization', False)
dropout = sum_desc_opts.get('dropout', 0.0)
updateEmb = sum_desc_opts.get('update_embedding', False)
batch_normalization = sum_desc_opts.get('update_embedding', False)
hiddenSize = sum_desc_opts.get('hidden_size')
sumDescEncoder = WordMean(embedding, updateEmb, hiddenSize, standardization, dropout, batch_normalization)
encoders.append(sumDescEncoder)
inputHandlers.append(RNNInputHandler(paddingId))
else:
raise ArgumentError(
"Encoder type of summary and description is invalid (%s). You should choose one of these: cnn" %
sum_desc_opts['encoder_type'])
def processDescriptionParam(descOpts, bugReportDatabase, inputHandlers, preprocessors, encoders, databasePath,
cacheFolder,
logger, paddingSym):
# Use summary and description (concatenated) to address this problem
logger.info("Using Description information.")
# Loading word embedding
lexicon, embedding = load_embedding(descOpts, paddingSym)
logger.info("Lexicon size: %d" % (lexicon.getLen()))
logger.info("Word Embedding size: %d" % (embedding.getEmbeddingSize()))
paddingId = lexicon.getLexiconIndex(paddingSym)
# Loading Filters
filters = loadFilters(descOpts['filters'])
# Tokenizer
if descOpts['tokenizer'] == 'default':
logger.info("Use default tokenizer to tokenize summary information")
tokenizer = MultiLineTokenizer()
elif descOpts['tokenizer'] == 'white_space':
logger.info("Use white space tokenizer to tokenize summary information")
tokenizer = WhitespaceTokenizer()
else:
raise ArgumentError(
"Tokenizer value %s is invalid. You should choose one of these: default and white_space" %
descOpts['tokenizer'])
arguments = (
databasePath, descOpts['word_embedding'], str(descOpts['lexicon']),
' '.join(sorted([fil.__class__.__name__ for fil in filters])),
descOpts['tokenizer'], "description")
descCache = PreprocessingCache(cacheFolder, arguments)
descPreprocessor = DescriptionPreprocessor(lexicon, bugReportDatabase, filters, tokenizer, paddingId, descCache)
preprocessors.append(descPreprocessor)
if descOpts['encoder_type'] == 'rnn':
rnnType = descOpts.get('rnn_type')
hiddenSize = descOpts.get('hidden_size')
bidirectional = descOpts.get('bidirectional', False)
numLayers = descOpts.get('num_layers', 1)
dropout = descOpts.get('dropout', 0.0)
updateEmb = descOpts.get('update_embedding', False)
fixedOpt = descOpts.get('fixed_opt', False)
descRNN = SortedRNNEncoder(rnnType, embedding, hiddenSize, numLayers, bidirectional, updateEmb,
dropout)
if fixedOpt == 'self_att':
att = SelfAttention(descRNN.getOutputSize(), descOpts['self_att_hidden'], descOpts['n_hops'])
descEncoder = RNN_Self_Attention(descRNN, att, paddingId, dropout)
else:
descEncoder = RNNFixedOuput(descRNN, fixedOpt, dropout)
encoders.append(descEncoder)
inputHandlers.append(RNNInputHandler(paddingId))
elif descOpts['encoder_type'] == 'cnn':
windowSizes = descOpts.get('window_sizes', [3])
nFilters = descOpts.get('nfilters', 100)
updateEmb = descOpts.get('update_embedding', False)
actFunc = loadActivationFunction(descOpts.get('activation', 'relu'))
batchNorm = descOpts.get('batch_normalization', False)
dropout = descOpts.get('dropout', 0.0)
descEncoder = TextCNN(windowSizes, nFilters, embedding, updateEmb, actFunc, batchNorm, dropout)
encoders.append(descEncoder)
inputHandlers.append(TextCNNInputHandler(paddingId, max(windowSizes)))
elif descOpts['encoder_type'] == 'cnn+dense':
windowSizes = descOpts.get('window_sizes', [3])
nFilters = descOpts.get('nfilters', 100)
updateEmb = descOpts.get('update_embedding', False)
actFunc = loadActivationFunction(descOpts.get('activation', 'relu'))
batchNorm = descOpts.get('batch_normalization', False)
dropout = descOpts.get('dropout', 0.0)
hiddenSizes = descOpts.get('hidden_sizes')
hiddenAct = loadActivationClass(descOpts.get('hidden_act'))
hiddenDropout = descOpts.get('hidden_dropout')
batchLast = descOpts.get("bn_last_layer", False)
cnnEnc = TextCNN(windowSizes, nFilters, embedding, updateEmb, actFunc, batchNorm, dropout)
descEncoder = MultilayerDense(cnnEnc, hiddenSizes, hiddenAct, batchNorm, batchLast, hiddenDropout)
encoders.append(descEncoder)
inputHandlers.append(TextCNNInputHandler(paddingId, max(windowSizes)))
elif descOpts['encoder_type'] == 'dense+self_att':
dropout = descOpts.get('dropout', 0.0)
hiddenSize = descOpts.get('hidden_size')
self_att_hidden = descOpts['self_att_hidden']
n_hops = descOpts['n_hops']
updateEmb = descOpts.get('update_embedding', False)
descEncoder = Dense_Self_Attention(embedding, hiddenSize, self_att_hidden, n_hops, paddingId, updateEmb, dropout=dropout)
encoders.append(descEncoder)
inputHandlers.append(TextCNNInputHandler(paddingId, -1))
elif descOpts['encoder_type'] == 'word_mean':
standardization = descOpts.get('standardization', False)
dropout = descOpts.get('dropout', 0.0)
updateEmb = descOpts.get('update_embedding', False)
batch_normalization = descOpts.get('update_embedding', False)
hiddenSize = descOpts.get('hidden_size')
descEncoder = WordMean( embedding, updateEmb, hiddenSize, standardization, dropout, batch_normalization)
encoders.append(descEncoder)
inputHandlers.append(RNNInputHandler(paddingId))
else:
raise ArgumentError(
"Encoder type of summary and description is invalid (%s). You should choose one of these: cnn" %
descOpts['encoder_type'])
logger.info(args)
lexicon, embedding = Embedding.fromFile(args.word_embedding,
'UUUKNNN',
hasHeader=False,
paddingSym="</s>")
database_lexicon = set(['UUUKNNN', '</s>'])
if args.db is not None:
# Filter any word that is not in the dataset.
db = BugReportDatabase.fromJson(args.db)
filters = loadFilters(args.filters)
if args.tk == 'default':
logger.info(
"Use default tokenizer to tokenize summary information")
tokenizer = MultiLineTokenizer()
elif args.tk == 'white_space':
logger.info(
"Use white space tokenizer to tokenize summary information")
tokenizer = WhitespaceTokenizer()
else:
raise ArgumentError(
"Tokenizer value %s is invalid. You should choose one of these: default and white_space"
% args.tk)
for report in db.bugList:
for token in extract_tokens(report.get('short_desc', ''),
tokenizer, filters):
database_lexicon.add(token)
for token in extract_tokens(report.get('description', ''),
trainingBugs.update(bugDataset.bugIds)
logger.info("Preprocessing and fitting data")
trainingText = []
for bugId in trainingBugs:
bugReport = bugReportDataset.getBug(bugId)
text = concatenateSummaryAndDescription(bugReport)
trainingText.append(text)
if args.load:
logger.info('Loading object')
vectorizer = pickle.load(open(args.load, 'rb'))
else:
tokenizer = WhitespaceTokenizer() if args.space_tokenize else MultiLineTokenizer()
stemmer = SnowballStemmer('english', ignore_stopwords=True) if args.stemmer else None
stopWords = set(stopwords.words('english') + list(string.punctuation) + ["n't", "'t", ])
filters = [TransformNumberToZeroFilter(), StripPunctuactionFilter(), DectectNotUsualWordFilter()]
tokenizerStemmer = ClassicalPreprocessing(tokenizer, stemmer, stopWords, filters)
logger.info("Using %s to tokenize" % tokenizer.__class__.__name__)
if args.trigram:
ngramRange = (1, 3)
elif args.bigram:
ngramRange = (1, 2)
else:
ngramRange = (1, 1)
if args.type == 'tfidf':
def main(_run, _config, _seed, _log):
# Setting logger
args = _config
logger = _log
logger.info(args)
logger.info('It started at: %s' % datetime.now())
torch.manual_seed(_seed)
device = torch.device('cuda' if args['cuda'] else "cpu")
if args['cuda']:
logger.info("Turning CUDA on")
else:
logger.info("Turning CUDA off")
# Setting the parameter to save and loading parameters
important_parameters = ['dbr_cnn']
parameters_to_save = dict([(name, args[name])
for name in important_parameters])
if args['load'] is not None:
map_location = (
lambda storage, loc: storage.cuda()) if args['cuda'] else 'cpu'
model_info = torch.load(args['load'], map_location=map_location)
model_state = model_info['model']
for param_name, param_value in model_info['params'].items():
args[param_name] = param_value
else:
model_state = None
# Set basic variables
preprocessors = PreprocessorList()
input_handlers = []
report_database = BugReportDatabase.fromJson(args['bug_database'])
batchSize = args['batch_size']
dbr_cnn_opt = args['dbr_cnn']
# Loading word embedding and lexicon
emb = np.load(dbr_cnn_opt["word_embedding"])
padding_sym = "</s>"
lexicon = Lexicon(unknownSymbol=None)
with codecs.open(dbr_cnn_opt["lexicon"]) as f:
for l in f:
lexicon.put(l.strip())
lexicon.setUnknown("UUUKNNN")
padding_id = lexicon.getLexiconIndex(padding_sym)
embedding = Embedding(lexicon, emb, paddingIdx=padding_id)
logger.info("Lexicon size: %d" % (lexicon.getLen()))
logger.info("Word Embedding size: %d" % (embedding.getEmbeddingSize()))
# Load filters and tokenizer
filters = loadFilters(dbr_cnn_opt['filters'])
if dbr_cnn_opt['tokenizer'] == 'default':
logger.info("Use default tokenizer to tokenize summary information")
tokenizer = MultiLineTokenizer()
elif dbr_cnn_opt['tokenizer'] == 'white_space':
logger.info(
"Use white space tokenizer to tokenize summary information")
tokenizer = WhitespaceTokenizer()
else:
raise ArgumentError(
"Tokenizer value %s is invalid. You should choose one of these: default and white_space"
% dbr_cnn_opt['tokenizer'])
# Add preprocessors
preprocessors.append(
DBR_CNN_CategoricalPreprocessor(dbr_cnn_opt['categorical_lexicon'],
report_database))
preprocessors.append(
SummaryDescriptionPreprocessor(lexicon, report_database, filters,
tokenizer, padding_id))
# Add input_handlers
input_handlers.append(DBRDCNN_CategoricalInputHandler())
input_handlers.append(
TextCNNInputHandler(padding_id, min(dbr_cnn_opt["window"])))
# Create Model
model = DBR_CNN(embedding, dbr_cnn_opt["window"], dbr_cnn_opt["nfilters"],
dbr_cnn_opt['update_embedding'])
model.to(device)
if model_state:
model.load_state_dict(model_state)
# Set loss function
logger.info("Using BCE Loss")
loss_fn = BCELoss()
loss_no_reduction = BCELoss(reduction='none')
cmp_collate = PairBugCollate(input_handlers,
torch.float32,
unsqueeze_target=True)
# Loading the training and setting how the negative example will be generated.
if args.get('pairs_training'):
negative_pair_gen_opt = args.get('neg_pair_generator', )
pairsTrainingFile = args.get('pairs_training')
random_anchor = negative_pair_gen_opt['random_anchor']
offlineGeneration = not (negative_pair_gen_opt is None
or negative_pair_gen_opt['type'] == 'none')
if not offlineGeneration:
logger.info("Not generate dynamically the negative examples.")
pair_training_reader = PairBugDatasetReader(
pairsTrainingFile,
preprocessors,
randomInvertPair=args['random_switch'])
else:
pair_gen_type = negative_pair_gen_opt['type']
master_id_by_bug_id = report_database.getMasterIdByBugId()
if pair_gen_type == 'random':
logger.info("Random Negative Pair Generator")
training_dataset = BugDataset(
negative_pair_gen_opt['training'])
bug_ids = training_dataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (training_dataset.info, len(bug_ids)))
negative_pair_generator = RandomGenerator(
preprocessors, cmp_collate, negative_pair_gen_opt['rate'],
bug_ids, master_id_by_bug_id)
elif pair_gen_type == 'non_negative':
logger.info("Non Negative Pair Generator")
training_dataset = BugDataset(
negative_pair_gen_opt['training'])
bug_ids = training_dataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (training_dataset.info, len(bug_ids)))
negative_pair_generator = NonNegativeRandomGenerator(
preprocessors,
cmp_collate,
negative_pair_gen_opt['rate'],
bug_ids,
master_id_by_bug_id,
negative_pair_gen_opt['n_tries'],
device,
randomAnchor=random_anchor)
elif pair_gen_type == 'misc_non_zero':
logger.info("Misc Non Zero Pair Generator")
training_dataset = BugDataset(
negative_pair_gen_opt['training'])
bug_ids = training_dataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (training_dataset.info, len(bug_ids)))
negative_pair_generator = MiscNonZeroRandomGen(
preprocessors, cmp_collate, negative_pair_gen_opt['rate'],
bug_ids, training_dataset.duplicateIds,
master_id_by_bug_id, device,
negative_pair_gen_opt['n_tries'],
negative_pair_gen_opt['random_anchor'])
elif pair_gen_type == 'random_k':
logger.info("Random K Negative Pair Generator")
training_dataset = BugDataset(
negative_pair_gen_opt['training'])
bug_ids = training_dataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (training_dataset.info, len(bug_ids)))
negative_pair_generator = KRandomGenerator(
preprocessors, cmp_collate, negative_pair_gen_opt['rate'],
bug_ids, master_id_by_bug_id, negative_pair_gen_opt['k'],
device)
elif pair_gen_type == "pre":
logger.info("Pre-selected list generator")
negative_pair_generator = PreSelectedGenerator(
negative_pair_gen_opt['pre_list_file'], preprocessors,
negative_pair_gen_opt['rate'], master_id_by_bug_id,
negative_pair_gen_opt['preselected_length'])
elif pair_gen_type == "misc_non_zero_pre":
logger.info("Pre-selected list generator")
negativePairGenerator1 = PreSelectedGenerator(
negative_pair_gen_opt['pre_list_file'], preprocessors,
negative_pair_gen_opt['rate'], master_id_by_bug_id,
negative_pair_gen_opt['preselected_length'])
training_dataset = BugDataset(
negative_pair_gen_opt['training'])
bug_ids = training_dataset.bugIds
negativePairGenerator2 = NonNegativeRandomGenerator(
preprocessors, cmp_collate, negative_pair_gen_opt['rate'],
bug_ids, master_id_by_bug_id, device,
negative_pair_gen_opt['n_tries'])
negative_pair_generator = MiscOfflineGenerator(
(negativePairGenerator1, negativePairGenerator2))
else:
raise ArgumentError(
"Offline generator is invalid (%s). You should choose one of these: random, hard and pre"
% pair_gen_type)
pair_training_reader = PairBugDatasetReader(
pairsTrainingFile,
preprocessors,
negative_pair_generator,
randomInvertPair=args['random_switch'])
training_loader = DataLoader(pair_training_reader,
batch_size=batchSize,
collate_fn=cmp_collate.collate,
shuffle=True)
logger.info("Training size: %s" % (len(training_loader.dataset)))
# load validation
if args.get('pairs_validation'):
pair_validation_reader = PairBugDatasetReader(
args.get('pairs_validation'), preprocessors)
validation_loader = DataLoader(pair_validation_reader,
batch_size=batchSize,
collate_fn=cmp_collate.collate)
logger.info("Validation size: %s" % (len(validation_loader.dataset)))
else:
validation_loader = None
"""
Training and evaluate the model.
"""
optimizer_opt = args.get('optimizer', 'adam')
if optimizer_opt == 'sgd':
logger.info('SGD')
optimizer = optim.SGD(model.parameters(),
lr=args['lr'],
weight_decay=args['l2'],
momentum=args['momentum'])
elif optimizer_opt == 'adam':
logger.info('Adam')
optimizer = optim.Adam(model.parameters(),
lr=args['lr'],
weight_decay=args['l2'])
# Recall rate
ranking_scorer = DBR_CNN_Scorer(preprocessors[0], preprocessors[1],
input_handlers[0], input_handlers[1],
model, device, args['ranking_batch_size'])
recallEstimationTrainOpt = args.get('recall_estimation_train')
if recallEstimationTrainOpt:
preselectListRankingTrain = PreselectListRanking(
recallEstimationTrainOpt)
recallEstimationOpt = args.get('recall_estimation')
if recallEstimationOpt:
preselect_list_ranking = PreselectListRanking(recallEstimationOpt)
lr_scheduler_opt = args.get('lr_scheduler', None)
if lr_scheduler_opt is None or lr_scheduler_opt['type'] == 'constant':
logger.info("Scheduler: Constant")
lr_sched = None
elif lr_scheduler_opt["type"] == 'step':
logger.info("Scheduler: StepLR (step:%s, decay:%f)" %
(lr_scheduler_opt["step_size"], args["decay"]))
lr_sched = StepLR(optimizer, lr_scheduler_opt["step_size"],
lr_scheduler_opt["decay"])
elif lr_scheduler_opt["type"] == 'exp':
logger.info("Scheduler: ExponentialLR (decay:%f)" %
(lr_scheduler_opt["decay"]))
lr_sched = ExponentialLR(optimizer, lr_scheduler_opt["decay"])
elif lr_scheduler_opt["type"] == 'linear':
logger.info(
"Scheduler: Divide by (1 + epoch * decay) ---- (decay:%f)" %
(lr_scheduler_opt["decay"]))
lrDecay = lr_scheduler_opt["decay"]
lr_sched = LambdaLR(optimizer, lambda epoch: 1 /
(1.0 + epoch * lrDecay))
else:
raise ArgumentError(
"LR Scheduler is invalid (%s). You should choose one of these: step, exp and linear "
% pair_gen_type)
# Set training functions
def trainingIteration(engine, batch):
model.train()
optimizer.zero_grad()
x, y = cmp_collate.to(batch, device)
output = model(*x)
loss = loss_fn(output, y)
loss.backward()
optimizer.step()
return loss, output, y
trainer = Engine(trainingIteration)
negTarget = 0.0 if isinstance(loss_fn, NLLLoss) else -1.0
trainingMetrics = {
'training_loss':
AverageLoss(loss_fn),
'training_acc':
AccuracyWrapper(output_transform=thresholded_output_transform),
'training_precision':
PrecisionWrapper(output_transform=thresholded_output_transform),
'training_recall':
RecallWrapper(output_transform=thresholded_output_transform),
}
# Add metrics to trainer
for name, metric in trainingMetrics.items():
metric.attach(trainer, name)
# Set validation functions
def validationIteration(engine, batch):
model.eval()
with torch.no_grad():
x, y = cmp_collate.to(batch, device)
y_pred = model(*x)
return y_pred, y
validationMetrics = {
'validation_loss':
LossWrapper(loss_fn),
'validation_acc':
AccuracyWrapper(output_transform=thresholded_output_transform),
'validation_precision':
PrecisionWrapper(output_transform=thresholded_output_transform),
'validation_recall':
RecallWrapper(output_transform=thresholded_output_transform),
}
evaluator = Engine(validationIteration)
# Add metrics to evaluator
for name, metric in validationMetrics.items():
metric.attach(evaluator, name)
@trainer.on(Events.EPOCH_STARTED)
def onStartEpoch(engine):
epoch = engine.state.epoch
logger.info("Epoch: %d" % epoch)
if lr_sched:
lr_sched.step()
logger.info("LR: %s" % str(optimizer.param_groups[0]["lr"]))
@trainer.on(Events.EPOCH_COMPLETED)
def onEndEpoch(engine):
epoch = engine.state.epoch
logMetrics(_run, logger, engine.state.metrics, epoch)
# Evaluate Training
if validation_loader:
evaluator.run(validation_loader)
logMetrics(_run, logger, evaluator.state.metrics, epoch)
lastEpoch = args['epochs'] - epoch == 0
if recallEstimationTrainOpt and (epoch % args['rr_train_epoch'] == 0):
logRankingResult(_run, logger, preselectListRankingTrain,
ranking_scorer, report_database, None, epoch,
"train")
ranking_scorer.free()
if recallEstimationOpt and (epoch % args['rr_val_epoch'] == 0):
logRankingResult(_run, logger, preselect_list_ranking,
ranking_scorer, report_database,
args.get("ranking_result_file"), epoch,
"validation")
ranking_scorer.free()
if not lastEpoch:
pair_training_reader.sampleNewNegExamples(model, loss_no_reduction)
if args.get('save'):
save_by_epoch = args['save_by_epoch']
if save_by_epoch and epoch in save_by_epoch:
file_name, file_extension = os.path.splitext(args['save'])
file_path = file_name + '_epoch_{}'.format(
epoch) + file_extension
else:
file_path = args['save']
modelInfo = {
'model': model.state_dict(),
'params': parameters_to_save
}
logger.info("==> Saving Model: %s" % file_path)
torch.save(modelInfo, file_path)
if args.get('pairs_training'):
trainer.run(training_loader, max_epochs=args['epochs'])
elif args.get('pairs_validation'):
# Evaluate Training
evaluator.run(validation_loader)
logMetrics(logger, evaluator.state.metrics)
if recallEstimationOpt:
logRankingResult(_run, logger, preselect_list_ranking,
ranking_scorer, report_database,
args.get("ranking_result_file"), 0, "validation")
# Test Dataset (accuracy, recall, precision, F1)
pair_test_dataset = args.get('pair_test_dataset')
if pair_test_dataset is not None and len(pair_test_dataset) > 0:
pairTestReader = PairBugDatasetReader(pair_test_dataset, preprocessors)
testLoader = DataLoader(pairTestReader,
batch_size=batchSize,
collate_fn=cmp_collate.collate)
if not isinstance(cmp_collate, PairBugCollate):
raise NotImplementedError(
'Evaluation of pairs using tanh was not implemented yet')
logger.info("Test size: %s" % (len(testLoader.dataset)))
testMetrics = {
'test_accuracy':
ignite.metrics.Accuracy(
output_transform=thresholded_output_transform),
'test_precision':
ignite.metrics.Precision(
output_transform=thresholded_output_transform),
'test_recall':
ignite.metrics.Recall(
output_transform=thresholded_output_transform),
'test_predictions':
PredictionCache(),
}
test_evaluator = Engine(validationIteration)
# Add metrics to evaluator
for name, metric in testMetrics.items():
metric.attach(test_evaluator, name)
test_evaluator.run(testLoader)
for metricName, metricValue in test_evaluator.state.metrics.items():
metric = testMetrics[metricName]
if isinstance(metric, ignite.metrics.Accuracy):
logger.info({
'type': 'metric',
'label': metricName,
'value': metricValue,
'epoch': None,
'correct': metric._num_correct,
'total': metric._num_examples
})
_run.log_scalar(metricName, metricValue)
elif isinstance(metric,
(ignite.metrics.Precision, ignite.metrics.Recall)):
logger.info({
'type': 'metric',
'label': metricName,
'value': metricValue,
'epoch': None,
'tp': metric._true_positives.item(),
'total_positive': metric._positives.item()
})
_run.log_scalar(metricName, metricValue)
elif isinstance(metric, ConfusionMatrix):
acc = cmAccuracy(metricValue)
prec = cmPrecision(metricValue, False)
recall = cmRecall(metricValue, False)
f1 = 2 * (prec * recall) / (prec + recall + 1e-15)
logger.info({
'type':
'metric',
'label':
metricName,
'accuracy':
np.float(acc),
'precision':
prec.cpu().numpy().tolist(),
'recall':
recall.cpu().numpy().tolist(),
'f1':
f1.cpu().numpy().tolist(),
'confusion_matrix':
metricValue.cpu().numpy().tolist(),
'epoch':
None
})
_run.log_scalar('test_f1', f1[1])
elif isinstance(metric, PredictionCache):
logger.info({
'type': 'metric',
'label': metricName,
'predictions': metric.predictions
})
# Calculate recall rate
recall_rate_opt = args.get('recall_rate', {'type': 'none'})
if recall_rate_opt['type'] != 'none':
if recall_rate_opt['type'] == 'sun2011':
logger.info("Calculating recall rate: {}".format(
recall_rate_opt['type']))
recall_rate_dataset = BugDataset(recall_rate_opt['dataset'])
ranking_class = SunRanking(report_database, recall_rate_dataset,
recall_rate_opt['window'])
# We always group all bug reports by master in the results in the sun 2011 methodology
group_by_master = True
elif recall_rate_opt['type'] == 'deshmukh':
logger.info("Calculating recall rate: {}".format(
recall_rate_opt['type']))
recall_rate_dataset = BugDataset(recall_rate_opt['dataset'])
ranking_class = DeshmukhRanking(report_database,
recall_rate_dataset)
group_by_master = recall_rate_opt['group_by_master']
else:
raise ArgumentError(
"recall_rate.type is invalid (%s). You should choose one of these: step, exp and linear "
% recall_rate_opt['type'])
logRankingResult(
_run,
logger,
ranking_class,
ranking_scorer,
report_database,
recall_rate_opt["result_file"],
0,
None,
group_by_master,
)
def main(_run, _config, _seed, _log):
"""
:param _run:
:param _config:
:param _seed:
:param _log:
:return:
"""
"""
Setting and loading parameters
"""
# Setting logger
args = _config
logger = _log
logger.info(args)
logger.info('It started at: %s' % datetime.now())
torch.manual_seed(_seed)
bugReportDatabase = BugReportDatabase.fromJson(args['bug_database'])
paddingSym = "</s>"
batchSize = args['batch_size']
device = torch.device('cuda' if args['cuda'] else "cpu")
if args['cuda']:
logger.info("Turning CUDA on")
else:
logger.info("Turning CUDA off")
# It is the folder where the preprocessed information will be stored.
cacheFolder = args['cache_folder']
# Setting the parameter to save and loading parameters
importantParameters = ['compare_aggregation', 'categorical']
parametersToSave = dict([(parName, args[parName])
for parName in importantParameters])
if args['load'] is not None:
mapLocation = (
lambda storage, loc: storage.cuda()) if cudaOn else 'cpu'
modelInfo = torch.load(args['load'], map_location=mapLocation)
modelState = modelInfo['model']
for paramName, paramValue in modelInfo['params'].items():
args[paramName] = paramValue
else:
modelState = None
if args['rep'] is not None and args['rep']['model']:
logger.info("Loading REP")
rep = read_weights(args['rep']['model'])
rep_input, max_tkn_id = read_dbrd_file(args['rep']['input'], math.inf)
rep_recommendation = args['rep']['k']
rep.fit_transform(rep_input, max_tkn_id, True)
rep_input_by_id = {}
for inp in rep_input:
rep_input_by_id[inp[SUN_REPORT_ID_INDEX]] = inp
else:
rep = None
preprocessors = PreprocessorList()
inputHandlers = []
categoricalOpt = args.get('categorical')
if categoricalOpt is not None and len(categoricalOpt) != 0:
categoricalEncoder, _, _ = processCategoricalParam(
categoricalOpt, bugReportDatabase, inputHandlers, preprocessors,
None, logger, cudaOn)
else:
categoricalEncoder = None
filterInputHandlers = []
compareAggOpt = args['compare_aggregation']
databasePath = args['bug_database']
# Loading word embedding
if compareAggOpt["word_embedding"]:
# todo: Allow use embeddings and other representation
lexicon, embedding = Embedding.fromFile(
compareAggOpt['word_embedding'],
'UUUKNNN',
hasHeader=False,
paddingSym=paddingSym)
logger.info("Lexicon size: %d" % (lexicon.getLen()))
logger.info("Word Embedding size: %d" % (embedding.getEmbeddingSize()))
paddingId = lexicon.getLexiconIndex(paddingSym)
lazy = False
else:
embedding = None
# Tokenizer
if compareAggOpt['tokenizer'] == 'default':
logger.info("Use default tokenizer to tokenize summary information")
tokenizer = MultiLineTokenizer()
elif compareAggOpt['tokenizer'] == 'white_space':
logger.info(
"Use white space tokenizer to tokenize summary information")
tokenizer = WhitespaceTokenizer()
else:
raise ArgumentError(
"Tokenizer value %s is invalid. You should choose one of these: default and white_space"
% compareAggOpt['tokenizer'])
# Preparing input handlers, preprocessors and cache
minSeqSize = max(compareAggOpt['aggregate']["window"]
) if compareAggOpt['aggregate']["model"] == "cnn" else -1
if compareAggOpt['summary'] is not None:
# Use summary and description (concatenated) to address this problem
logger.info("Using Summary information.")
# Loading Filters
sumFilters = loadFilters(compareAggOpt['summary']['filters'])
if compareAggOpt['summary']['model_type'] in ('lstm', 'gru',
'word_emd', 'residual'):
arguments = (databasePath, compareAggOpt['word_embedding'],
' '.join(
sorted([
fil.__class__.__name__ for fil in sumFilters
])), compareAggOpt['tokenizer'],
SummaryPreprocessor.__name__)
inputHandlers.append(
RNNInputHandler(paddingId, minInputSize=minSeqSize))
summaryCache = PreprocessingCache(cacheFolder, arguments)
summaryPreprocessor = SummaryPreprocessor(lexicon,
bugReportDatabase,
sumFilters, tokenizer,
paddingId, summaryCache)
elif compareAggOpt['summary']['model_type'] == 'ELMo':
raise NotImplementedError("ELMO is not implemented!")
# inputHandlers.append(ELMoInputHandler(cudaOn, minInputSize=minSeqSize))
# summaryPreprocessor = ELMoPreprocessor(0, elmoEmbedding)
# compareAggOpt['summary']["input_size"] = elmoEmbedding.get_size()
elif compareAggOpt['summary']['model_type'] == 'BERT':
arguments = (databasePath, "CADD SUMMARY", "BERT",
"bert-base-uncased")
inputHandlers.append(BERTInputHandler(0, minInputSize=minSeqSize))
summaryCache = PreprocessingCache(cacheFolder, arguments)
summaryPreprocessor = TransformerPreprocessor(
"short_desc", "bert-base-uncased", BertTokenizer, 0,
bugReportDatabase, summaryCache)
# compareAggOpt['summary']["input_size"] = 768
preprocessors.append(summaryPreprocessor)
if compareAggOpt['desc'] is not None:
# Use summary and description (concatenated) to address this problem
logger.info("Using Description information.")
descFilters = loadFilters(compareAggOpt['desc']['filters'])
if compareAggOpt['desc']['model_type'] in ('lstm', 'gru', 'word_emd',
'residual'):
arguments = (databasePath, compareAggOpt['word_embedding'],
' '.join(
sorted([
fil.__class__.__name__ for fil in descFilters
])), compareAggOpt['tokenizer'], "CADD DESC",
str(compareAggOpt['desc']['summarization']))
inputHandlers.append(
RNNInputHandler(paddingId, minInputSize=minSeqSize))
descriptionCache = PreprocessingCache(cacheFolder, arguments)
descPreprocessor = DescriptionPreprocessor(lexicon,
bugReportDatabase,
descFilters,
tokenizer,
paddingId,
cache=descriptionCache)
elif compareAggOpt['desc']['model_type'] == 'ELMo':
raise NotImplementedError("ELMO is not implemented!")
# inputHandlers.append(ELMoInputHandler(cudaOn, minInputSize=minSeqSize))
# descPreprocessor = ELMoPreprocessor(1, elmoEmbedding)
# compareAggOpt['desc']["input_size"] = elmoEmbedding.get_size()
elif compareAggOpt['desc']['model_type'] == 'BERT':
arguments = (databasePath, "CADD DESC", "BERT",
"bert-base-uncased")
inputHandlers.append(BERTInputHandler(0, minInputSize=minSeqSize))
descriptionCache = PreprocessingCache(cacheFolder, arguments)
descPreprocessor = TransformerPreprocessor("description",
"bert-base-uncased",
BertTokenizer, 0,
bugReportDatabase,
descriptionCache)
# compareAggOpt['desc']["input_size"] = 768
preprocessors.append(descPreprocessor)
# Create model
model = CADD(embedding,
categoricalEncoder,
compareAggOpt,
compareAggOpt['summary'],
compareAggOpt['desc'],
compareAggOpt['matching'],
compareAggOpt['aggregate'],
cudaOn=cudaOn)
lossFn = F.nll_loss
lossNoReduction = NLLLoss(reduction='none')
if cudaOn:
model.cuda()
if modelState:
model.load_state_dict(modelState)
"""
Loading the training and validation. Also, it sets how the negative example will be generated.
"""
cmpAggCollate = PairBugCollate(inputHandlers, torch.int64)
# load training
if args.get('pairs_training'):
negativePairGenOpt = args.get('neg_pair_generator', )
pairTrainingFile = args.get('pairs_training')
offlineGeneration = not (negativePairGenOpt is None
or negativePairGenOpt['type'] == 'none')
masterIdByBugId = bugReportDatabase.getMasterIdByBugId()
randomAnchor = negativePairGenOpt['random_anchor']
if rep:
logger.info("Generate negative examples using REP.")
randomAnchor = negativePairGenOpt['random_anchor']
trainingDataset = BugDataset(args['rep']['training'])
bugIds = trainingDataset.bugIds
negativePairGenerator = REPGenerator(rep, rep_input_by_id,
args['rep']['neg_training'],
preprocessors, bugIds,
masterIdByBugId,
args['rep']['rate'],
randomAnchor)
elif not offlineGeneration:
logger.info("Not generate dynamically the negative examples.")
negativePairGenerator = None
else:
pairGenType = negativePairGenOpt['type']
if pairGenType == 'random':
logger.info("Random Negative Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = RandomGenerator(
preprocessors,
cmpAggCollate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
randomAnchor=randomAnchor)
elif pairGenType == 'non_negative':
logger.info("Non Negative Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = NonNegativeRandomGenerator(
preprocessors,
cmpAggCollate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
elif pairGenType == 'misc_non_zero':
logger.info("Misc Non Zero Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = MiscNonZeroRandomGen(
preprocessors,
cmpAggCollate,
negativePairGenOpt['rate'],
bugIds,
trainingDataset.duplicateIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
elif pairGenType == 'random_k':
logger.info("Random K Negative Pair Generator")
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
logger.info(
"Using the following dataset to generate negative examples: %s. Number of bugs in the training: %d"
% (trainingDataset.info, len(bugIds)))
negativePairGenerator = KRandomGenerator(
preprocessors,
cmpAggCollate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['k'],
device,
randomAnchor=randomAnchor)
elif pairGenType == "pre":
logger.info("Pre-selected list generator")
negativePairGenerator = PreSelectedGenerator(
negativePairGenOpt['pre_list_file'],
preprocessors,
negativePairGenOpt['rate'],
masterIdByBugId,
negativePairGenOpt['preselected_length'],
randomAnchor=randomAnchor)
elif pairGenType == "positive_pre":
logger.info("Positive Pre-selected list generator")
negativePairGenerator = PositivePreSelectedGenerator(
negativePairGenOpt['pre_list_file'],
preprocessors,
cmpAggCollate,
negativePairGenOpt['rate'],
masterIdByBugId,
negativePairGenOpt['preselected_length'],
randomAnchor=randomAnchor)
elif pairGenType == "misc_non_zero_pre":
logger.info("Misc: non-zero and Pre-selected list generator")
negativePairGenerator1 = PreSelectedGenerator(
negativePairGenOpt['pre_list_file'],
preprocessors,
negativePairGenOpt['rate'],
masterIdByBugId,
negativePairGenOpt['preselected_length'],
randomAnchor=randomAnchor)
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
negativePairGenerator2 = NonNegativeRandomGenerator(
preprocessors,
cmpAggCollate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
negativePairGenerator = MiscOfflineGenerator(
(negativePairGenerator1, negativePairGenerator2))
elif pairGenType == "misc_non_zero_positive_pre":
logger.info(
"Misc: non-zero and Positive Pre-selected list generator")
negativePairGenerator1 = PositivePreSelectedGenerator(
negativePairGenOpt['pre_list_file'],
preprocessors,
cmpAggCollate,
negativePairGenOpt['rate'],
masterIdByBugId,
negativePairGenOpt['preselected_length'],
randomAnchor=randomAnchor)
trainingDataset = BugDataset(negativePairGenOpt['training'])
bugIds = trainingDataset.bugIds
negativePairGenerator2 = NonNegativeRandomGenerator(
preprocessors,
cmpAggCollate,
negativePairGenOpt['rate'],
bugIds,
masterIdByBugId,
negativePairGenOpt['n_tries'],
device,
randomAnchor=randomAnchor)
negativePairGenerator = MiscOfflineGenerator(
(negativePairGenerator1, negativePairGenerator2))
else:
raise ArgumentError(
"Offline generator is invalid (%s). You should choose one of these: random, hard and pre"
% pairGenType)
pairTrainingReader = PairBugDatasetReader(
pairTrainingFile,
preprocessors,
negativePairGenerator,
randomInvertPair=args['random_switch'])
trainingCollate = cmpAggCollate
trainingLoader = DataLoader(pairTrainingReader,
batch_size=batchSize,
collate_fn=trainingCollate.collate,
shuffle=True)
logger.info("Training size: %s" % (len(trainingLoader.dataset)))
# load validation
if args.get('pairs_validation'):
pairValidationReader = PairBugDatasetReader(
args.get('pairs_validation'), preprocessors)
validationLoader = DataLoader(pairValidationReader,
batch_size=batchSize,
collate_fn=cmpAggCollate.collate)
logger.info("Validation size: %s" % (len(validationLoader.dataset)))
else:
validationLoader = None
"""
Training and evaluate the model.
"""
optimizer_opt = args.get('optimizer', 'adam')
if optimizer_opt == 'sgd':
logger.info('SGD')
optimizer = optim.SGD(model.parameters(),
lr=args['lr'],
weight_decay=args['l2'])
elif optimizer_opt == 'adam':
logger.info('Adam')
optimizer = optim.Adam(model.parameters(),
lr=args['lr'],
weight_decay=args['l2'])
# Recall rate
rankingScorer = GeneralScorer(model, preprocessors, device, cmpAggCollate)
recallEstimationTrainOpt = args.get('recall_estimation_train')
if recallEstimationTrainOpt:
preselectListRankingTrain = PreselectListRanking(
recallEstimationTrainOpt, args['sample_size_rr_tr'])
recallEstimationOpt = args.get('recall_estimation')
if recallEstimationOpt:
preselectListRanking = PreselectListRanking(recallEstimationOpt,
args['sample_size_rr_val'])
# LR scheduler
lrSchedulerOpt = args.get('lr_scheduler', None)
if lrSchedulerOpt is None:
logger.info("Scheduler: Constant")
lrSched = None
elif lrSchedulerOpt["type"] == 'step':
logger.info("Scheduler: StepLR (step:%s, decay:%f)" %
(lrSchedulerOpt["step_size"], args["decay"]))
lrSched = StepLR(optimizer, lrSchedulerOpt["step_size"],
lrSchedulerOpt["decay"])
elif lrSchedulerOpt["type"] == 'exp':
logger.info("Scheduler: ExponentialLR (decay:%f)" %
(lrSchedulerOpt["decay"]))
lrSched = ExponentialLR(optimizer, lrSchedulerOpt["decay"])
elif lrSchedulerOpt["type"] == 'linear':
logger.info(
"Scheduler: Divide by (1 + epoch * decay) ---- (decay:%f)" %
(lrSchedulerOpt["decay"]))
lrDecay = lrSchedulerOpt["decay"]
lrSched = LambdaLR(optimizer, lambda epoch: 1 /
(1.0 + epoch * lrDecay))
else:
raise ArgumentError(
"LR Scheduler is invalid (%s). You should choose one of these: step, exp and linear "
% pairGenType)
# Set training functions
def trainingIteration(engine, batch):
engine.kk = 0
model.train()
optimizer.zero_grad()
x, y = batch
output = model(*x)
loss = lossFn(output, y)
loss.backward()
optimizer.step()
return loss, output, y
def scoreDistanceTrans(output):
if len(output) == 3:
_, y_pred, y = output
else:
y_pred, y = output
if lossFn == F.nll_loss:
return torch.exp(y_pred[:, 1]), y
trainer = Engine(trainingIteration)
trainingMetrics = {
'training_loss':
AverageLoss(lossFn, batch_size=lambda x: x[0].shape[0]),
'training_dist_target':
MeanScoreDistance(output_transform=scoreDistanceTrans)
}
# Add metrics to trainer
for name, metric in trainingMetrics.items():
metric.attach(trainer, name)
# Set validation functions
def validationIteration(engine, batch):
if not hasattr(engine, 'kk'):
engine.kk = 0
model.eval()
with torch.no_grad():
x, y = batch
y_pred = model(*x)
# for k, (pred, t) in enumerate(zip(y_pred, y)):
# engine.kk += 1
# print("{}: {} \t {}".format(engine.kk, torch.round(torch.exp(pred) * 100), t))
return y_pred, y
validationMetrics = {
'validation_loss':
ignite.metrics.Loss(lossFn),
'validation_dist_target':
MeanScoreDistance(output_transform=scoreDistanceTrans)
}
evaluator = Engine(validationIteration)
# Add metrics to evaluator
for name, metric in validationMetrics.items():
metric.attach(evaluator, name)
# recommendation
if rep:
recommendation_fn = REP_CADD_Recommender(
rep, rep_input_by_id,
rep_recommendation).generateRecommendationList
else:
recommendation_fn = generateRecommendationList
@trainer.on(Events.EPOCH_STARTED)
def onStartEpoch(engine):
epoch = engine.state.epoch
logger.info("Epoch: %d" % epoch)
if lrSched:
lrSched.step()
logger.info("LR: %s" % str(optimizer.param_groups[0]["lr"]))
@trainer.on(Events.EPOCH_COMPLETED)
def onEndEpoch(engine):
epoch = engine.state.epoch
logMetrics(_run, logger, engine.state.metrics, epoch)
# Evaluate Training
if validationLoader:
evaluator.run(validationLoader)
logMetrics(_run, logger, evaluator.state.metrics, epoch)
if recallEstimationTrainOpt and (epoch % args['rr_train_epoch'] == 0):
logRankingResult(_run,
logger,
preselectListRankingTrain,
rankingScorer,
bugReportDatabase,
None,
epoch,
"train",
recommendationListfn=recommendation_fn)
rankingScorer.free()
if recallEstimationOpt and (epoch % args['rr_val_epoch'] == 0):
logRankingResult(_run,
logger,
preselectListRanking,
rankingScorer,
bugReportDatabase,
args.get("ranking_result_file"),
epoch,
"validation",
recommendationListfn=recommendation_fn)
rankingScorer.free()
pairTrainingReader.sampleNewNegExamples(model, lossNoReduction)
if args.get('save'):
save_by_epoch = args['save_by_epoch']
if save_by_epoch and epoch in save_by_epoch:
file_name, file_extension = os.path.splitext(args['save'])
file_path = file_name + '_epoch_{}'.format(
epoch) + file_extension
else:
file_path = args['save']
modelInfo = {
'model': model.state_dict(),
'params': parametersToSave
}
logger.info("==> Saving Model: %s" % file_path)
torch.save(modelInfo, file_path)
if args.get('pairs_training'):
trainer.run(trainingLoader, max_epochs=args['epochs'])
elif args.get('pairs_validation'):
# Evaluate Training
evaluator.run(validationLoader)
logMetrics(_run, logger, evaluator.state.metrics, 0)
if recallEstimationOpt:
logRankingResult(_run,
logger,
preselectListRanking,
rankingScorer,
bugReportDatabase,
args.get("ranking_result_file"),
0,
"validation",
recommendationListfn=recommendation_fn)
recallRateOpt = args.get('recall_rate', {'type': 'none'})
if recallRateOpt['type'] != 'none':
if recallRateOpt['type'] == 'sun2011':
logger.info("Calculating recall rate: {}".format(
recallRateOpt['type']))
recallRateDataset = BugDataset(recallRateOpt['dataset'])
rankingClass = SunRanking(bugReportDatabase, recallRateDataset,
recallRateOpt['window'])
# We always group all bug reports by master in the results in the sun 2011 methodology
group_by_master = True
elif recallRateOpt['type'] == 'deshmukh':
logger.info("Calculating recall rate: {}".format(
recallRateOpt['type']))
recallRateDataset = BugDataset(recallRateOpt['dataset'])
rankingClass = DeshmukhRanking(bugReportDatabase,
recallRateDataset)
group_by_master = recallRateOpt['group_by_master']
else:
raise ArgumentError(
"recall_rate.type is invalid (%s). You should choose one of these: step, exp and linear "
% recallRateOpt['type'])
logRankingResult(_run,
logger,
rankingClass,
rankingScorer,
bugReportDatabase,
recallRateOpt["result_file"],
0,
None,
group_by_master,
recommendationListfn=recommendation_fn)