def main():
args = setupArgs()
train_data, train_labels = utils.loadEmbeddings(args.train_data)
val_data, val_labels = utils.loadEmbeddings(args.val_data)
test_data, test_labels = excludeTrainClasses(*utils.loadEmbeddings(args.test_data), train_labels)
test_data, test_labels = utils.loadEmbeddings(args.test_data)
clf = model.NearestCentroidSVM()
clf.verbose = True
predict_kwargs = {
'mode' : 'average'
}
print("Training {} classifier.".format(clf.__class__))
false_pos, true_pos = evaluateClassifier(train_data, val_data, test_data, clf, predict_kwargs)
plotROC(false_pos, true_pos, description = args.title)
def test(test_data, labels, model, device, batch=1, training=0, embeddings=None):
model.eval()
if embeddings == None:
embeddings = loadEmbeddings(model.vocab, model.embed_size,
'./data/word2vec.6B.100d.txt')
count, correct_count = 0, 0
with torch.no_grad():
for test_x, test_y in batch_iter(test_data, labels, batch):
test_x = model.vocab.to_input_tensor(test_x)
test_x = embeddings(test_x).to(device)
output = model.search(test_x)
test_y = test_y[0]
for i in range(len(test_y)):
count += 1
if test_y[i] == output[i]:
correct_count += 1
correct_rate = 1.*correct_count/count
print('the corrent rate is : ', correct_rate)
if training:
model.train()
return correct_rate
def __init__(self, queryEmbedFile, docEmbedFile, exclusive, debugPrint=False, useIDF=True):
super().__init__(exclusive)
self.debugPrint = debugPrint
self.useIDF = useIDF
print('Loading answer embeddings from: ' + docEmbedFile)
answWords, self.answEmbed = loadEmbeddings(docEmbedFile)
self.answEmbedMap = createEmbedMap(answWords)
if queryEmbedFile is not None:
print('Loading query embeddings from: ' + queryEmbedFile)
queryWords, self.queryEmbed = loadEmbeddings(queryEmbedFile)
self.queryEmbedMap = createEmbedMap(queryWords)
else:
self.queryEmbed = self.answEmbed
self.queryEmbedMap = self.answEmbedMap
print('Loading is done!')
def main():
args = setupArgs()
feature_data, labels_dict = utils.loadEmbeddings(args.embeddings)
makeProjection(feature_data)
def main():
args = setupArgs()
train_feature_data, train_labels_dict = utils.loadEmbeddings(args.train)
test_feature_data, test_labels_dict = utils.loadEmbeddings(args.test)
makeProjection(train_feature_data, test_feature_data)
def main():
args = setupArgs()
feature_data, labels_dict = utils.loadEmbeddings(args.embeddings)
evaluateEmbeddingsKNN(feature_data)
evaluateEmbeddingsKMeans(feature_data)
evaluateEmbeddingsAffinityProp(feature_data)
print('# of classes: ' + str(corpus.classVoc.size()))
print()
print('# of training samples: ' + str(len(corpus.trainData)))
print('# of dev samples: ' + str(len(corpus.devData)))
print()
embedding = Embedding(corpus.voc.size(), corpus.charVoc.size(), embedDim,
charDim)
tagger = Tagger(embedDim + charDim, hiddenDim, corpus.classVoc.size(),
inputDropoutRate, outputDropoutRate)
if not test and not args.random:
if os.path.exists(wordParamsFile):
embedding.wordEmbedding.load_state_dict(torch.load(wordParamsFile))
else:
utils.loadEmbeddings(embedding.wordEmbedding, corpus.voc,
wordEmbeddingFile)
torch.save(embedding.wordEmbedding.state_dict(), wordParamsFile)
if os.path.exists(charParamsFile):
embedding.charEmbedding.load_state_dict(torch.load(charParamsFile))
else:
utils.loadEmbeddings(embedding.charEmbedding, corpus.charVoc,
charEmbeddingFile)
torch.save(embedding.charEmbedding.state_dict(), charParamsFile)
if test:
tagger.load_state_dict(torch.load(taggerParamsFile))
embedding.load_state_dict(torch.load(embeddingParamsFile))
if useGpu:
if torch.cuda.is_available():
torch.cuda.set_device(args.gpuId)
def train(train_data, train_labels, test_data, test_labels,
model, optimizer, device, epoch, batch, params_path, corate=-1):
model.train()
log_every = 10
save_iter = 120
embeddings = loadEmbeddings(model.vocab, model.embed_size,
'./data/word2vec.6B.100d.txt')
loss_data = []
hint_correct_rate = [corate]
s = time.time()
for times in range(epoch):
train_iter = 0
for train_x, train_y in batch_iter(train_data, train_labels, batch):
optimizer.zero_grad()
sum_loss = torch.zeros(1,).to(device)
m = len(train_y)
for idx in range(m):
x = model.vocab.to_input_tensor([train_x[idx]])
x = embeddings(x).to(device)
y = torch.tensor([train_y[idx]], device=device)
loss = model(x, y)
sum_loss = sum_loss+loss
sum_loss = 1./m*sum_loss
sum_loss.backward()
optimizer.step()
if train_iter == 0:
loss_data.append(sum_loss.item())
if train_iter % log_every == 0:
e = time.time()
print("the {} epoch, the {} iter, loss is : {} [time is : {}]".format(times,
train_iter, sum_loss.item(), (e-s)/60.))
if train_iter == save_iter:
# auto save params
correct_rate = test(test_data, test_labels, model, device, training=1, embeddings=embeddings)
if correct_rate > max(hint_correct_rate):
checkpoint = {
'model_dict': model.state_dict(),
'corate': correct_rate
}
torch.save(checkpoint, params_path)
print('save params')
else:
print('not save params')
hint_correct_rate.append(correct_rate)
train_iter += 1
loss_curve(loss_data)
print(hint_correct_rate)
# load best model
state = torch.load(params_path)
model.load_state_dict(state['model_dict'])
