def get_batch_of_embeddings(data_dir, model_dir, n_speakers, n_utterances):
# create data set
speaker_dataset = SpeakerVerificationDataset(data_dir, dataset_len=26)
# create model
device = torch.device('cuda')
model = load_model(model_dir, device)
model.eval()
# get embeddings
print("running model on batch...")
with torch.no_grad():
speakers = [
speaker_dataset[i]
for i in np.random.randint(0, len(speaker_dataset), n_speakers)
] # for i in np.arange(n_speakers)]
my_batch = SpeakerBatch(speakers,
utterances_per_speaker=n_utterances,
n_frames=160)
my_batch_data = torch.from_numpy(my_batch.data).to(device)
print("Batch shape: ", my_batch_data.shape)
embeds = model(my_batch_data)
print("Embeds shape: ", embeds.shape)
embeds = embeds.detach().cpu().numpy()
return embeds
def collate(self, speakers):
return SpeakerBatch(speakers, self.utterances_per_speaker,
partials_n_frames)
def collate(self, speakers):
return SpeakerBatch(speakers, self.utterances_per_speaker,
PARTIALS_N_FRAMES)