time_per_line * nb_lines_ahead))
dense_logits = prepare_dense_logits(sparse_logits)
if args.greedy:
boh = decoder(dense_logits)
else:
boh = decoder(dense_logits, args.model_eos)
one_best = boh.best_hyp()
decodings[name] = one_best
confidences[name] = boh.confidence()
if args.cn_best:
cn = confusion_networks.produce_cn_from_boh(boh)
cn_decodings[name] = confusion_networks.best_cn_path(cn)
reporter.clear()
save_transcriptions(args.best, decodings)
with open(args.confidence, 'w') as f:
for name in decodings:
f.write('{} {:.3f}\n'.format(name, confidences[name]))
if args.cn_best:
save_transcriptions(args.cn_best, cn_decodings)
if __name__ == "__main__":
args = parse_arguments()
gpu_owner = GPUOwner()
main(args)
parser.add_argument('--seed', type=int, default=1111,
help='random seed')
parser.add_argument('--cuda', action='store_true',
help='use CUDA')
parser.add_argument('--load', type=str, required=True,
help='where to load a model from')
args = parser.parse_args()
print(args)
init_seeds(args.seed, args.cuda)
print("loading model...")
device = torch.device('cuda') if args.cuda else torch.device('cpu')
if args.cuda:
gpu_owner = GPUOwner(lambda: torch.zeros((1), device='cuda'))
lm = torch.load(args.load, map_location=device)
print(lm)
evaluator = SubstitutionalEnblockEvaluator_v2(
lm,
args.data,
args.batch_size,
args.target_seq_len,
lambda streams: Corruptor(streams, args.subs_rate, len(lm.vocab), args.del_rate, args.ins_rate, protected=[lm.vocab['</s>']]),
args.rounds,
)
eval_report = evaluator.evaluate(report_individual=args.individual)
print('total loss {:.1f} | per token loss {:5.2f} | ppl {:8.2f}'.format(eval_report.total_loss, eval_report.loss_per_token, math.exp(eval_report.loss_per_token)))