def encode_sentences(task, audios, batch_size=128):
return numpy.vstack([
task.predict(
torch.autograd.Variable(torch.from_numpy(
vector_padder(batch))).cuda()).data.cpu().numpy()
for batch in util.grouper(audios, batch_size)
])
def encode_sentences_SpeechText(task, audios, batch_size=128):
def predict(x):
return task.SpeechText.SpeechEncoderTop(task.SpeechText.SpeechEncoderBottom(x))
return numpy.vstack([ predict(
torch.autograd.Variable(torch.from_numpy(
vector_padder(batch))).cuda()).data.cpu().numpy()
for batch in util.grouper(audios, batch_size) ])
def embed(net, audios, batch_size=32):
"""Return utterance embeddings for audio using the given net."""
device = next(net.parameters()).device
out = []
for batch in util.grouper(audios, batch_size):
for result in net.predict(torch.from_numpy(vector_padder(batch)).to(device)).cpu().numpy():
out.append(result)
return np.stack(out)
def get_attn_weights(speech_transcriber, sentences):
sent_audio = [get_audio(s) for s in sentences]
sent_len = [sd.shape[0] for sd in sent_audio]
v_audio = torch.from_numpy(sd.vector_padder(sent_audio,
pad_end=True)).cuda()
v_audio_len = torch.from_numpy(numpy.array(sent_len)).cuda()
_, attn_weights = speech_transcriber(v_audio, v_audio_len)
return attn_weights.numpy()
def transcribe(speech_transcriber, sentences):
sent_audio = [get_audio(s) for s in sentences]
sent_len = [sd.shape[0] for sd in sent_audio]
v_audio = torch.from_numpy(sd.vector_padder(sent_audio,
pad_end=True)).cuda()
v_audio_len = torch.from_numpy(numpy.array(sent_len)).cuda()
trn = speech_transcriber.predict(v_audio, v_audio_len)
return trn
def iter_layer_states(model, audios, batch_size=128):
"""Pass audios through the model and for each audio return the state of each timestep and each layer."""
lens = (numpy.array(map(len, audios)) +
model.config['filter_length']) // model.config['stride']
rs = (r for batch in util.grouper(audios, batch_size)
for r in model.task.pile(vector_padder(batch)))
for (r, l) in itertools.izip(rs, lens):
yield r[-l:, :, :]
def encode_sentences(model, audios, batch_size=128):
"""Project audios to the joint space using model.
For each audio returns a vector.
"""
return numpy.vstack([
model.task.predict(
torch.autograd.Variable(torch.from_numpy(
vector_padder(batch))).cuda()).data.cpu().numpy()
for batch in util.grouper(audios, batch_size)
])
def get_state_stack(net, audios, batch_size=128):
"""Pass audios through the model and for each audio return the state of each timestep and each layer."""
device = next(net.parameters()).device
result = []
lens = inout(np.array(list(map(len, audios))))
rs = (r for batch in util.grouper(audios, batch_size)
for r in state_stack(net, torch.from_numpy(vector_padder(batch)).to(device)).cpu().numpy()
)
for (r,l) in zip(rs, lens):
result.append(r[-l:,:])
return result
def get_state_stack(net, audios, batch_size=128):
import onion.util as util
from vg.simple_data import vector_padder
"""Pass audios through the model and for each audio return the state of each timestep and each layer."""
result = []
lens = inout(np.array(list(map(len, audios))))
rs = (r for batch in util.grouper(audios, batch_size) for r in state_stack(
net,
torch.from_numpy(vector_padder(batch)).cuda()).cpu().numpy())
for (r, l) in zip(rs, lens):
result.append(r[-l:, :])
return result
