def __init__(self, device, max_frames, lr = 0.0001, margin = 1, scale = 1, hard_rank = 0, hard_prob = 0, model="alexnet50", nOut = 512, nSpeakers = 1000, optimizer = 'adam', encoder_type = 'SAP', normalize = True, trainfunc='contrastive', **kwargs):
super(SpeakerNet, self).__init__();
argsdict = {'nOut': nOut, 'encoder_type':encoder_type}
self.device = device
SpeakerNetModel = importlib.import_module('models.'+model).__getattribute__(model)
# @TODO make cuda optional in order to train on dev machines w/o GPUs
self.__S__ = SpeakerNetModel(**argsdict).to(self.device);
if trainfunc == 'angleproto':
self.__L__ = AngleProtoLoss().to(self.device)
self.__train_normalize__ = True
self.__test_normalize__ = True
elif trainfunc == 'ge2e':
self.__L__ = GE2ELoss().to(self.device)
self.__train_normalize__ = True
self.__test_normalize__ = True
elif trainfunc == 'amsoftmax':
self.__L__ = AMSoftmax(in_feats=nOut, n_classes=nSpeakers, m=margin, s=scale).to(self.device)
self.__train_normalize__ = False
self.__test_normalize__ = True
elif trainfunc == 'aamsoftmax':
self.__L__ = AAMSoftmax(in_feats=nOut, n_classes=nSpeakers, m=margin, s=scale).to(self.device)
self.__train_normalize__ = False
self.__test_normalize__ = True
elif trainfunc == 'softmax':
self.__L__ = SoftmaxLoss(in_feats=nOut, n_classes=nSpeakers).to(self.device)
self.__train_normalize__ = False
self.__test_normalize__ = True
elif trainfunc == 'proto':
self.__L__ = ProtoLoss().to(self.device)
self.__train_normalize__ = False
self.__test_normalize__ = False
elif trainfunc == 'triplet':
self.__L__ = PairwiseLoss(loss_func='triplet', hard_rank=hard_rank, hard_prob=hard_prob, margin=margin).to(self.device)
self.__train_normalize__ = True
self.__test_normalize__ = True
elif trainfunc == 'contrastive':
self.__L__ = PairwiseLoss(loss_func='contrastive', hard_rank=hard_rank, hard_prob=hard_prob, margin=margin).to(self.device)
self.__train_normalize__ = True
self.__test_normalize__ = True
else:
raise ValueError('Undefined loss.')
if optimizer == 'adam':
self.__optimizer__ = torch.optim.Adam(self.parameters(), lr = lr);
elif optimizer == 'sgd':
self.__optimizer__ = torch.optim.SGD(self.parameters(), lr = lr, momentum = 0.9, weight_decay=5e-5);
else:
raise ValueError('Undefined optimizer.')
self.__max_frames__ = max_frames;
def __init__(self,
lr=0.0001,
model="alexnet50",
nOut=512,
encoder_type='SAP',
normalize=True,
trainfunc='contrastive',
**kwargs):
super(SpeakerNet, self).__init__()
argsdict = {'nOut': nOut, 'encoder_type': encoder_type}
SpeakerNetModel = importlib.import_module(
'models.' + model).__getattribute__(model)
self.__S__ = SpeakerNetModel(**argsdict).cuda()
if trainfunc == 'angleproto':
self.__L__ = AngleProtoLoss().cuda()
self.__train_normalize__ = True
self.__test_normalize__ = True
elif trainfunc == 'proto':
self.__L__ = ProtoLoss().cuda()
self.__train_normalize__ = False
self.__test_normalize__ = False
else:
raise ValueError('Undefined loss.')
self.__optimizer__ = torch.optim.Adam(list(self.__S__.parameters()) +
list(self.__L__.parameters()),
lr=lr)
self.torchfb = transforms.MelSpectrogram(sample_rate=16000,
n_fft=512,
win_length=400,
hop_length=160,
f_min=0.0,
f_max=8000,
pad=0,
n_mels=40).cuda()
self.instancenorm = nn.InstanceNorm1d(40).cuda()
print('Initialised network with nOut %d encoder_type %s' %
(nOut, encoder_type))
def __init__(self,
max_frames,
lr=0.0001,
margin=1,
scale=1,
hard_rank=0,
hard_prob=0,
model="alexnet50",
nOut=512,
nSpeakers=1000,
optimizer='adam',
encoder_type='SAP',
normalize=True,
trainfunc='contrastive',
**kwargs):
super(SpeakerNet, self).__init__()
argsdict = {'nOut': nOut, 'encoder_type': encoder_type}
self.__S__ = globals()[model](**argsdict).cuda()
if trainfunc == 'angleproto':
self.__L__ = AngleProtoLoss().cuda()
self.__train_normalize__ = True
self.__test_normalize__ = True
elif trainfunc == 'ge2e':
self.__L__ = GE2ELoss().cuda()
self.__train_normalize__ = True
self.__test_normalize__ = True
elif trainfunc == 'amsoftmax':
self.__L__ = AMSoftmax(in_feats=nOut,
n_classes=nSpeakers,
m=margin,
s=scale).cuda()
self.__train_normalize__ = False
self.__test_normalize__ = True
elif trainfunc == 'aamsoftmax':
self.__L__ = AAMSoftmax(in_feats=nOut,
n_classes=nSpeakers,
m=margin,
s=scale).cuda()
self.__train_normalize__ = False
self.__test_normalize__ = True
elif trainfunc == 'softmax':
self.__L__ = SoftmaxLoss(in_feats=nOut, n_classes=nSpeakers).cuda()
self.__train_normalize__ = False
self.__test_normalize__ = True
elif trainfunc == 'proto':
self.__L__ = ProtoLoss().cuda()
self.__train_normalize__ = False
self.__test_normalize__ = False
elif trainfunc == 'triplet':
self.__L__ = PairwiseLoss(loss_func='triplet',
hard_rank=hard_rank,
hard_prob=hard_prob,
margin=margin).cuda()
self.__train_normalize__ = True
self.__test_normalize__ = True
elif trainfunc == 'contrastive':
self.__L__ = PairwiseLoss(loss_func='contrastive',
hard_rank=hard_rank,
hard_prob=hard_prob,
margin=margin).cuda()
self.__train_normalize__ = True
self.__test_normalize__ = True
else:
raise ValueError('Undefined loss.')
if optimizer == 'adam':
self.__optimizer__ = torch.optim.Adam(self.parameters(), lr=lr)
elif optimizer == 'sgd':
self.__optimizer__ = torch.optim.SGD(self.parameters(),
lr=lr,
momentum=0.9,
weight_decay=5e-5)
else:
raise ValueError('Undefined optimizer.')
self.__max_frames__ = max_frames