predict.py

import torch
import torchaudio

#fixes second prediction bug
torch._C._jit_override_can_fuse_on_cpu(False)
torch._C._jit_override_can_fuse_on_gpu(False)
torch._C._jit_set_texpr_fuser_enabled(False)
torch._C._jit_set_nvfuser_enabled(False)

loader = torch.jit.load("audio_loader.pt")
model = torch.jit.load('QuartzNet_thunderspeech_3.pt')

vocab = model.text_transform.vocab.itos
vocab[-1] = ''

def convert_probs(probs):
  ids = probs.argmax(1)[0]
  s = []
  if vocab[ids[0]]: s.append(vocab[ids[0]])
  for i in range(1,len(ids)):
    if ids[i-1] != ids[i]:
      new = vocab[ids[i]]
      if new: s.append(new)
  #return '.'.join(s)
  return s

def predict(path):
  audio = loader(path)
  probs = model(audio, torch.tensor(audio.shape[0] * [audio.shape[-1]], device=audio.device))[0]
  return convert_probs(probs)