def get_embeddings(use_cuda, filename, model, test_frames):
print(filename)
input, label = read_MFB(filename) # input size:(n_frames, n_dims)
# feature, speaker_id
tot_segments = math.ceil(
len(input) /
test_frames) # total number of segments with 'test_frames'
activation = 0
with torch.no_grad():
for i in range(tot_segments):
# Divide the input in sub_inputs of length test_frames
temp_input = input[i * test_frames:i * test_frames + test_frames]
TT = ToTensorTestInput()
temp_input = TT(temp_input) # size:(1, 1, n_dims, n_frames)
if use_cuda:
temp_input = temp_input.cuda()
temp_activation, _ = model(temp_input)
activation += torch.sum(temp_activation, dim=0, keepdim=True)
activation = l2_norm(activation, 10)
return activation
def get_embeddings(use_cuda, filename, model, test_frames):
input, label= read_MFB(filename) # input size:(n_frames, n_dims)
tot_segments = math.ceil(len(input)/test_frames) # total number of segments with 'test_frames'
activation = 0
with torch.no_grad():
for i in range(tot_segments):
temp_input = input[i*test_frames:i*test_frames+test_frames]
TT = ToTensorTestInput()
temp_input = TT(temp_input) # size:(1, 1, n_dims, n_frames)
if use_cuda:
temp_input = temp_input.cuda()
output, features, spk_embeddings = model.forward(temp_input,task='identification')
activation += torch.sum(spk_embeddings, dim=0, keepdim=True)
activation = l2_norm(activation, 1)
return activation
