att_size=20, cla_size=350, cla_layers=5,
num_classes=59,
tra_type='identity', rnn_mode='LSTM', cla_dropout=0.3)
# Load network
net.load_state_dict(torch.load(net_file, map_location=lambda storage, loc: storage))
net.eval()
if cuda == True:
net.cuda()
# Loop over subsets
for subset in all_subsets:
# Create dataset, sampler, loader
composed = transforms.Compose([warp_ctc_shift(), standardization('sample')])
testset = dset(h5file=h5_file, dataset_mode=dataset_mode, subset=subset, transform=composed)
sampler = HighThroughputSampler(testset, shuffle_batches=False, num_splits=1, max_frames=max_frame_cache,
debug=0, roll=False)
batch_sampler = SimpleBatchSampler(sampler=sampler)
test_loader = torch.utils.data.DataLoader(testset, batch_sampler=batch_sampler, num_workers=0,
collate_fn=collate_fn)
# Get network output
inference_list = infer(model=net, dataloader=test_loader, cuda=cuda)
PER, WER, CER = error_preliminary(inference_list, subset)
for pos_scale in all_pos_scales:
kw = kaldi_converter(inference_list, filename='{}_{}'.format(subset, pos_scale), scale=pos_scale)
logprob = kw.write_logprob()
reference = kw.write_reference()
pass
log_dict = collections.OrderedDict()
# Prepare transforms
train_transforms = [tl.warp_ctc_shift(), tl.standardization(args.standardization), tl.gaussian_noise(args.noise)]
val_transforms = [tl.warp_ctc_shift(), tl.standardization(args.standardization)]
if args.concatenation == True:
train_transforms.append(tl.concatenation())
val_transforms.append(tl.concatenation())
train_composed = transforms.Compose(train_transforms)
val_composed = transforms.Compose(val_transforms)
# Create datasets
trainset = dset(h5file=args.dataset, dataset_mode=args.dataset_mode, subset=args.trainset, speaker_wise=False,
transform=train_composed)
valset = dset(h5file=args.dataset, dataset_mode=args.dataset_mode, subset=args.valset, speaker_wise=False,
transform=val_composed)
# Define sampler
train_sampler = HighThroughputSampler(trainset, shuffle_batches=True, num_splits=3, max_frames=args.max_frame_cache,
debug=0)
train_batch_sampler = SimpleBatchSampler(sampler=train_sampler)
val_sampler = HighThroughputSampler(valset, shuffle_batches=False, roll=True, max_frames=args.max_frame_cache,
debug=0)
val_batch_sampler = SimpleBatchSampler(sampler=val_sampler)
train_loader = torch.utils.data.DataLoader(trainset, batch_sampler=train_batch_sampler, num_workers=1,
collate_fn=collate_fn)
val_loader = torch.utils.data.DataLoader(valset, batch_sampler=val_batch_sampler, num_workers=1,
collate_fn=collate_fn)