def run_evaluation(H):
vocab = Vocabulary(os.path.join(H.ROOT_DIR, H.EXPERIMENT),
encoding=H.TARGET_ENCODING)
audio_transform = transforms.Compose([
AudioNormalizeDB(db=H.NORMALIZE_DB, max_gain_db=H.NORMALIZE_MAX_GAIN),
AudioSpectrogram(sample_rate=H.AUDIO_SAMPLE_RATE,
window_size=H.SPECT_WINDOW_SIZE,
window_stride=H.SPECT_WINDOW_STRIDE,
window=H.SPECT_WINDOW),
AudioNormalize(),
FromNumpyToTensor(tensor_type=torch.FloatTensor)
])
label_transform = transforms.Compose([
TranscriptEncodeCTC(vocab),
FromNumpyToTensor(tensor_type=torch.LongTensor)
])
test_dataset = AudioDataset(os.path.join(H.ROOT_DIR, H.EXPERIMENT),
manifests_files=H.MANIFESTS,
datasets="test",
transform=audio_transform,
label_transform=label_transform,
max_data_size=None,
sorted_by='recording_duration')
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=H.BATCH_SIZE,
num_workers=H.NUM_WORKERS,
shuffle=False,
collate_fn=collate_fn,
pin_memory=True)
logger.info(test_loader.dataset)
model_pred = SpeechCNN(len(vocab),
input_size=256,
hidden_size=H.CNN_HIDDEN_SIZE,
dropout=H.CNN_DROPOUT,
initialize=torch_weight_init)
if H.USE_CUDA:
model_pred.cuda()
state = torch.load(os.path.join(H.EXPERIMENT, H.MODEL_NAME + '.tar'))
model_pred.load_state_dict(state)
ctc_decoder = CTCGreedyDecoder(vocab)
recognizer = Recognizer(model_pred, ctc_decoder, test_loader)
hypotheses = recognizer()
transcripts = []
for _, labels, _, label_sizes, _ in test_loader:
label_seq = CTCGreedyDecoder.decode_labels(labels, label_sizes, vocab)
transcripts.extend(label_seq)
bleu = Scorer.get_moses_multi_bleu(hypotheses,
transcripts,
lowercase=False)
wer, cer = Scorer.get_wer_cer(hypotheses, transcripts)
acc = Scorer.get_acc(hypotheses, transcripts)
logger.info('Test Summary \n'
'Bleu: {bleu:.3f}\n'
'WER: {wer:.3f}\n'
'CER: {cer:.3f}\n'
'ACC: {acc:.3f}'.format(bleu=bleu,
wer=wer * 100,
cer=cer * 100,
acc=acc * 100))
