def train(self, lr=cfg.lr):
train_data = self.reader.mini_batch_iterator(self.reader.train,
requires_speaker=True)
optim = torch.optim.Adam(self.clf.parameters(), lr=lr)
criterion = torch.nn.CrossEntropyLoss()
for itr, total_iter, feat, label, files, speaker in train_data:
optim.zero_grad()
features = [
self.feature_extract_mfcc(a, b, filename, augment=True)
for (a, b), filename in zip(feat, files)
]
features = [_ for _ in zip(*features)]
features[0], features[1], features[2] = self.pad_batch(
features[0], features[1], features[2])
features[3], features[4], features[5] = np.array(
features[3]), np.array(features[4]), np.array(features[5])
mfcc0, mfcc1, mfcc2 = cuda_(Variable(torch.from_numpy(features[0]).float())), \
cuda_(Variable(torch.from_numpy(features[1]).float())),\
cuda_(Variable(torch.from_numpy(features[2]).float()))
mfcc0, mfcc1, mfcc2 = mfcc0.transpose(0, 1), mfcc1.transpose(
0, 1), mfcc2.transpose(0, 1)
label = cuda_(Variable(torch.LongTensor(label)))
speaker = cuda_(Variable(torch.LongTensor(speaker)))
out1, out2 = self.clf(mfcc0, mfcc1, mfcc2, features[3])
loss1 = criterion(out1, label)
loss2 = criterion(out2, speaker)
loss = loss1 + loss2
loss.backward()
optim.step()
printer.info(
'%d/%d loss:%f %f %f' %
(itr, total_iter, loss1.data[0], loss2.data[0], loss.data[0]))
def test_iter(self, itr, total_iter, feat, label, files):
features = [
self.feature_extract_mfcc(a, b, filename)
for (a, b), filename in zip(feat, files)
]
features = [_ for _ in zip(*features)]
features[0], features[1], features[2] = self.pad_batch(
features[0], features[1], features[2])
features[3], features[4], features[5] = np.array(
features[3]), np.array(features[4]), np.array(features[5])
mfcc0, mfcc1, mfcc2 = cuda_(Variable(torch.from_numpy(features[0]).float())), \
cuda_(Variable(torch.from_numpy(features[1]).float())), \
cuda_(Variable(torch.from_numpy(features[2]).float()))
mfcc0, mfcc1, mfcc2 = mfcc0.transpose(0, 1), mfcc1.transpose(
0, 1), mfcc2.transpose(0, 1)
out, _ = self.clf(mfcc0, mfcc1, mfcc2, features[3])
out = F.softmax(out, dim=1)
prob, pred_ = torch.max(out, 1)
pred_ = pred_.data.cpu().numpy().tolist()
prob = prob.data.cpu().numpy().tolist()
return pred_, prob
def train(self, lr=cfg.lr):
#train_data = self.reader.mini_batch_iterator(self.reader.train)
train_data = self.reader.mini_batch_iterator(self.reader.train)
optim = torch.optim.Adam(self.clf.parameters(), lr=lr)
criterion = torch.nn.CrossEntropyLoss()
for itr, total_iter, feat, label, files in train_data:
optim.zero_grad()
inp, len0 = self.get_batch_full(feat, files, augment=True)
label = cuda_(Variable(torch.LongTensor(label)))
out = self.clf(inp, len0)
loss = criterion(out, label)
loss.backward()
optim.step()
printer.info('%d/%d loss:%f' % (itr, total_iter, loss.data[0]))
def cvar(arr):
v = cuda_(Variable(torch.from_numpy(arr).float()))
return v
def __init__(self):
super().__init__()
self.clf = HMRNN()
#self.clf = CNN_SP()
self.clf = cuda_(self.clf)
def __init__(self):
super().__init__()
self.clf = AdvHRNN()
self.clf = cuda_(self.clf)