tmp = NMT(cnfg.isize, nwordi, nwordt, cnfg.nlayer, cnfg.ff_hsize,
cnfg.drop, cnfg.attn_drop, cnfg.share_emb, cnfg.nhead,
cache_len_default, cnfg.attn_hsize, cnfg.norm_output,
cnfg.bindDecoderEmb, cnfg.forbidden_indexes)
tmp = load_model_cpu(modelf, tmp)
tmp.apply(load_fixing)
models.append(tmp)
mymodel = Ensemble(models)
mymodel.eval()
lossf = LabelSmoothingLoss(nwordt,
cnfg.label_smoothing,
ignore_index=pad_id,
reduction='none',
forbidden_index=cnfg.forbidden_indexes)
use_cuda, cuda_device, cuda_devices, multi_gpu = parse_cuda(
cnfg.use_cuda, cnfg.gpuid)
use_amp = cnfg.use_amp and use_cuda
# Important to make cudnn methods deterministic
set_random_seed(cnfg.seed, use_cuda)
if use_cuda:
mymodel.to(cuda_device)
lossf.to(cuda_device)
if multi_gpu:
mymodel = DataParallelMT(mymodel,
cnfg.bindDecoderEmb,
cnfg.forbidden_indexes,
ntask=ntask)
fine_tune_m = cnfg.fine_tune_m
mymodel = init_model_params(mymodel)
mymodel.apply(init_fixing)
if fine_tune_m is not None:
logger.info("Load pre-trained model from: " + fine_tune_m)
mymodel = load_model_cpu(fine_tune_m, mymodel)
mymodel.apply(load_fixing)
lossf = LabelSmoothingLoss(nwordt,
cnfg.label_smoothing,
ignore_index=pad_id,
reduction='sum',
forbidden_index=cnfg.forbidden_indexes)
if cnfg.src_emb is not None:
logger.info("Load source embedding from: " + cnfg.src_emb)
load_emb(cnfg.src_emb, mymodel.enc.wemb.weight, nwordi,
cnfg.scale_down_emb, cnfg.freeze_srcemb)
if cnfg.tgt_emb is not None:
logger.info("Load target embedding from: " + cnfg.tgt_emb)
load_emb(cnfg.tgt_emb, mymodel.dec.wemb.weight, nwordt,
cnfg.scale_down_emb, cnfg.freeze_tgtemb)
if cuda_device:
mymodel.to(cuda_device)
lossf.to(cuda_device)