def __init__(self, speech_encoder, config):
super(SpeechTranscriber, self).__init__()
self.config = config
self.SpeechEncoderBottom = speech_encoder
self.SpeechEncoderTop = SpeechEncoderTop(**config['SpeechEncoderTop'])
self.TextDecoder = DecoderWithAttn(**config['TextDecoder'])
self.optimizer = optim.Adam(self.parameters(), lr=config['lr'])
def __init__(self, speech_encoder, config):
super(SpeechCorrText, self).__init__()
self.config = config
self.SpeechEncoderBottom = speech_encoder
self.SpeechEncoderTop = SpeechEncoderTop(**config['SpeechEncoderTop'])
self.CorrEncoder = ImageEncoder(size=1024, size_target=1024 * 4)
self.optimizer = optim.Adam(self.parameters(), lr=config['lr'])
def __init__(self, speech_encoder, config):
super(SpeechImage, self).__init__()
self.config = config
self.SpeechEncoderBottom = speech_encoder
self.SpeechEncoderTop = SpeechEncoderTop(**config['SpeechEncoderTop'])
self.ImageEncoder = ImageEncoder(**config['ImageEncoder'])
self.optimizer = optim.Adam(self.parameters(), lr=config['lr'])
class SpeechTranscriber(nn.Module):
def __init__(self, speech_encoder, config):
super(SpeechTranscriber, self).__init__()
self.config = config
self.SpeechEncoderBottom = speech_encoder
self.SpeechEncoderTop = SpeechEncoderTop(**config['SpeechEncoderTop'])
self.TextDecoder = DecoderWithAttn(**config['TextDecoder'])
self.optimizer = optim.Adam(self.parameters(), lr=config['lr'])
def cost(self, speech, target, target_prev):
states, rep = self.SpeechEncoderTop.states(
self.SpeechEncoderBottom(speech))
target_logits = self.TextDecoder(states, rep, target_prev)
cost = F.cross_entropy(
target_logits.view(
target_logits.size(0) * target_logits.size(1), -1),
target.view(target.size(0) * target.size(1)))
return cost
def args(self, item):
return (item['audio'], item['target_t'].astype('int64'),
item['target_prev_t'].astype('int64'))
def test_cost(self, *args):
with testing(self):
return self.cost(*args)
def __init__(self, speech_encoder, text_encoder, config):
super(SpeechText, self).__init__()
self.config = config
self.SpeechEncoderBottom = speech_encoder
self.TextEncoderBottom = text_encoder
self.SpeechEncoderTop = SpeechEncoderTop(**config['SpeechEncoderTop'])
self.TextEncoderTop = TextEncoderTop(**config['TextEncoderTop'])
self.optimizer = optim.Adam(self.parameters(), lr=config['lr'])
