def train(trainset=trainset, valset=valset, weight_file=args.weights):
"""trains classifier on name->gender
Args:
trainset: list of name->gender tuple pairs for training
valset (opt): list of name->gender tuple pairs to validation
weight_file: filename to save classifer weights
"""
start = time.time()
print("Training Naive Bayes Classifer on %d examples (%s)" %
(len(trainset), time_since(start)))
trainset = apply_features(_gender_features, trainset, labeled=True)
classifier = nltk.NaiveBayesClassifier.train(trainset)
print("Training complete. (%s)" % (time_since(start)))
# validation
if valset is not None and len(valset) > 0:
valset = apply_features(_gender_features, valset, labeled=True)
acc = nltk.classify.accuracy(classifier, valset)
print("Validation accuracy is %.2f%% on %d examples (%s)" %
(acc * 100, len(valset), time_since(start)))
# save weights
with open(weight_file, 'wb') as f:
pickle.dump(classifier, f)
f.close()
print('Weights saved to "%s"' % (args.weights))
def train(dataset=TRAINSET,
batch_size=args.batch_size,
num_workers=args.num_workers,
start_ep=args.start_epoch,
end_ep=args.end_epoch,
save_name=save_nm,
print_every=100,
log_iters=args.log_iters):
rnn.train()
print('Loading Dataset...')
dataset = NameGenderDataset(dataset)
collate_fn = name_gender_collate_cuda if args.cuda else name_gender_collate
data_loader = data.DataLoader(dataset,
batch_size=batch_size,
num_workers=num_workers,
shuffle=True,
collate_fn=collate_fn,
pin_memory=True)
start = time.time()
print("Beginning training...")
for epoch in range(start_ep, end_ep + 1):
ep_loss = 0
# fencepost print
if epoch == start_ep: print('EPOCH %s/%s' % (start_ep, end_ep))
# iterate over all minibatches
batch_iterator = iter(data_loader)
batch = 0
while batch <= 500:
try:
batch += 1
batch_loss = 0
names_tensor, genders_tensor = next(batch_iterator)
for name_tensor, gender_tensor in zip(names_tensor,
genders_tensor):
output, loss = _train(name_tensor, gender_tensor)
batch_loss += loss
if log_iters and batch % print_every == 0:
print('\tLoss[ ep%d: %.2f | mb%d: %.2f ] (%s) ' %
(epoch, ep_loss / batch, batch,
batch_loss / len(names_tensor), time_since(start)))
ep_loss += batch_loss
except StopIteration:
break
print(
'EPOCH %d %d%% (%s) avg loss: %.4f' %
(epoch, epoch / end_ep * 100, time_since(start), ep_loss / batch))
ep_loss = 0
torch.save(rnn.state_dict(),
save_name + '_epoch' + repr(epoch) + '.pth')
torch.save(rnn.state_dict(), save_name + '_classification.pth')
def test(testset=testset, weight_file=args.weights):
"""tests classifier on name->gender
Args:
train: % of examples to train with (e.g., 0.8)
"""
start = time.time()
classifier = load_classifier(weight_file)
print("Testing Naive Bayes Classifer on %d examples (%s)" %
(len(testset), time_since(start)))
testset = apply_features(_gender_features, testset, labeled=True)
acc = nltk.classify.accuracy(classifier, testset)
print("Testing accuracy is %.2f%% on %d examples (%s)" %
(acc * 100, len(testset), time_since(start)))
return acc
def write_classifications(names,
weight_file=args.weights,
outfile=args.outfile):
classifier = load_classifier(weight_file)
headers = ["name", "gender", "probability"]
with open(outfile, 'w') as f:
writer = csv.writer(f, delimiter=',')
writer.writerow(headers)
for name in names:
guess, prob = _classify(name, classifier)
writer.writerow([name, guess, prob])
f.close()
print('\nWrote %d names to "%s" (%s)' %
(len(names), outfile, time_since(start)))
def test(dataset=TESTSET, verbose=log_iters):
print('Loading Dataset...')
dataset = NameGenderDataset(dataset)
collate_fn = name_gender_collate_cuda if args.cuda else name_gender_collate
data_loader = data.DataLoader(dataset,
batch_size=batch_size,
num_workers=num_workers,
shuffle=True,
collate_fn=collate_fn,
pin_memory=True)
print("Beginning testing on %s names:\n" % (len(dataset)))
start = time.time()
cum = 0
# iterate over all minibatches
batch_iterator = iter(data_loader)
batch = 0
while True:
try:
batch += 1
batch_acc = 0
names_tensor, genders_tensor = next(batch_iterator)
for name_tensor, gender_tensor in zip(names_tensor,
genders_tensor):
gt = ALL_GENDERS[gender_tensor.data[0]]
name = tensor_to_name(name_tensor)
output = _evaluate(name_tensor)
topv, topi = output.data.topk(k=1, dim=1, largest=True)
guess = ALL_GENDERS[topi[0][0]]
correct = '!' if guess == gt else 'X (%s)' % gt
if verbose: print("\t%s -> %s %s " % (name, guess, correct))
batch_acc += 1 if guess == gt else 0
print("%.2f%% minibatch acc: %.4f (%s)" %
(batch / (len(dataset) / batch_size),
batch_acc / len(names_tensor), time_since(start)))
cum += batch_acc
except StopIteration:
break
acc = cum / len(dataset)
print()
print("TOTAL: %d/%d (%.4f%%)" % (cum, len(dataset), acc * 100))
return acc
def classify(names, weight_file=args.weights):
classifier = load_classifier(weight_file)
for name in names:
_classify(name, classifier)
print("\nClassified %d names (%s)" % (len(names), time_since(start)))