def main(libri_dir=c.DATASET_DIR):
print('Looking for fbank features [.npy] files in {}.'.format(libri_dir))
libri = data_catalog(libri_dir)
# filename speaker_id
# 0 audio/LibriSpeechSamples/train-clean-100-npy/1-100-0001.npy 1
# 1 audio/LibriSpeechSamples/train-clean-100-npy/1-100-0002.npy 1
unique_speakers = libri['speaker_id'].unique() # 251 speaker
transform=transforms.Compose([transforms.ToTensor()])
train_dir = stochastic_mini_batch(libri)
train_loader = DataLoader(train_dir, batch_size=c.BATCH_SIZE, shuffle=True)
model = DeepSpeakerModel(embedding_size=c.EMBEDDING_SIZE,num_classes=c.NUM_SPEAKERS)
optimizer = optim.Adam(model.parameters(), lr=0.001, weight_decay=0.0)
epoch = 0
model.cuda()
summary(model, input_size=(1, 160, 64))
for epoch in range(100):
model.train()
for batch_idx, (data_a, data_p, data_n,label_a,label_p,label_n) in tqdm(enumerate(train_loader)):
data_a, data_p, data_n = data_a.type(torch.FloatTensor),data_p.type(torch.FloatTensor),data_n.type(torch.FloatTensor)
data_a, data_p, data_n = data_a.cuda(), data_p.cuda(), data_n.cuda()
data_a, data_p, data_n = Variable(data_a), Variable(data_p), Variable(data_n)
out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)
triplet_loss = TripletMarginLoss(0.2).forward(out_a, out_p, out_n)
loss = triplet_loss
# compute gradient and update weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('selected_triplet_loss', triplet_loss.data)
print("epoch:",epoch)
torch.save(model.state_dict(),"checkpoint_{}.pt".format(epoch))
train_loader = DataLoader(train_dataset,
batch_size=opt.batch_size,
num_workers=opt.num_workers,
drop_last=False,
shuffle=True)
test_dataset = Mydataset(opt.data_path, 'test')
test_loader = DataLoader(test_dataset,
batch_size=opt.batch_size,
num_workers=opt.num_workers,
drop_last=False,
shuffle=True)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=opt.lr)
iteration = 0
best_acc = 0
test_corr = 0
test_all = 0
for epoch in range(0, 1):
if epoch % 2 == 0:
with torch.no_grad():
for j, batch in enumerate(test_loader):
inputs, targets = batch[0].cuda(), batch[1].cuda()
outputs = model(inputs)
_, preds = torch.max(F.softmax(outputs, dim=1).data, 1)
test_corr += torch.sum(preds == targets.data)