train_data = SpectrogramDataset(
vocab,
args,
audio_conf,
manifest_filepath_list=args.train_manifest_list,
normalize=True,
augment=args.augment,
input_type=args.input_type,
is_train=True,
partitions=args.train_partition_list)
elif args.feat == "logfbank":
train_data = LogFBankDataset(
vocab,
args,
audio_conf,
manifest_filepath_list=args.train_manifest_list,
normalize=True,
augment=args.augment,
input_type=args.input_type,
is_train=True)
train_data_list.append(train_data)
valid_loader_list, test_loader_list = [], []
for i in range(len(args.valid_manifest_list)):
if args.feat == "spectrogram":
valid_data = SpectrogramDataset(
vocab,
args,
audio_conf,
manifest_filepath_list=[args.valid_manifest_list[i]],
normalize=True,
audio_conf = dict(sample_rate=loaded_args.sample_rate,
window_size=loaded_args.window_size,
window_stride=loaded_args.window_stride,
window=loaded_args.window,
noise_dir=loaded_args.noise_dir,
noise_prob=loaded_args.noise_prob,
noise_levels=(loaded_args.noise_min, loaded_args.noise_max))
test_manifest_list = args.test_manifest_list
print("INPUT TYPE: ", args.input_type)
if loaded_args.feat == "spectrogram":
test_data = SpectrogramDataset(vocab, args, audio_conf=audio_conf, manifest_filepath_list=[test_manifest_list[0]], normalize=True, augment=False, input_type=args.input_type)
elif loaded_args.feat == "logfbank":
test_data = LogFBankDataset(vocab, args, audio_conf=audio_conf, manifest_filepath_list=[test_manifest_list[0]], normalize=True, augment=False, input_type=args.input_type)
test_sampler = BucketingSampler(test_data, batch_size=args.k_test)
test_loader = AudioDataLoader(vocab.PAD_ID, dataset=test_data, num_workers=args.num_workers, batch_sampler=test_sampler)
print("Parameters: {}(trainable), {}(non-trainable)".format(compute_num_params(model)[0], compute_num_params(model)[1]))
if not args.cuda:
model = model.cpu()
lm = None
if args.lm_rescoring:
lm = LM(args.lm_path, args)
print(">>>>>>>>>", args.tgt_max_len)
evaluate(model, vocab, test_loader, args, lm=lm, start_token=vocab.SOS_ID)