error while runing score.sh in aishell

[lfgogogo]

error message:./score.sh:line 64: 841Segmentation fault (core dumped) CUDA_VISIBLE_DEVICES=${gpu} ${NEURALSP_ROOT}/neural_sp/bin/asr/eval.py......
When i ran score.sh,it crashed without giving exact error.
How to solve it?

[lfgogogo]

I quantize the model using pytorch 1.3,it seems to be successful.When i run score.sh using pytorch 1.3,it gives the error message above,maybe someting wrong happen in the dataload stage.Then i run score.sh using pytorch 1.0.1,it gives the following error message:
AttributeError:can't get attribute '_rebuild_qtensor' ,rise from neuralsp/bin/train_utils.py,exactly by the torch.load() in load_checkpoint().

[hirofumi0810]

@lfgogogo Quantization is one of the milestones in this framework. Could you please consider sending a PR?
Also, what kind of model did you try to evaluate?

[lfgogogo]

Of course.I am not sure about the kind you say. I quantize the model.epoch-25 and the model-avg10 using pytorch1.3 has the quantization module,when i use the quantized model to evaluate,it gives errors above.
Besides,i also use package distiller v0.3.2 to quantize the model.epoch-25,this raises an error:Unexpected key(s) in state_dict.I will make a PR soon.

[lfgogogo]

I can't login in github on the pycharm,so i can't pull request now,sorry for that.
Here is how i quantize the model.
`
from neural_sp.bin.train_utils import (
load_checkpoint
)
from neural_sp.models.seq2seq.speech2text import Speech2Text
import torch
from neural_sp.bin.args_asr import parse_args_eval
from distiller.quantization.range_linear import PostTrainLinearQuantizer
import distiller.apputils as apputils
from neural_sp.trainers.lr_scheduler import LRScheduler
from neural_sp.trainers.optimizer import set_optimizer

argv=[]
args, ,= parse_args_eval(argv)
model = Speech2Text(args, 'modelpath')
load_checkpoint('modelpath', model)

#1、using pytorch1.3
quantized_model=torch.quantization.quantize_dynamic(model, {torch.nn.Linear}, dtype=torch.qint8)

quantized_model_path='a path'
optimizer = set_optimizer(model, args.optimizer, args.lr, args.weight_decay)
is_transformer = 'former' in args.enc_type or 'former' in args.dec_type
scheduler = LRScheduler(optimizer, args.lr,
decay_type=args.lr_decay_type,
decay_start_epoch=args.lr_decay_start_epoch,
decay_rate=args.lr_decay_rate,
decay_patient_n_epochs=args.lr_decay_patient_n_epochs,
early_stop_patient_n_epochs=args.early_stop_patient_n_epochs,
lower_better=args.metric not in ['accuracy', 'bleu'],
warmup_start_lr=args.warmup_start_lr,
warmup_n_steps=args.warmup_n_steps,
peak_lr=0.05 / (getattr(args, 'transformer_enc_d_model', 0) **
0.5) if 'conformer' in args.enc_type else 1e6,
model_size=getattr(args, 'transformer_enc_d_model',
getattr(args, 'transformer_dec_d_model', 0)),
factor=args.lr_factor,
noam=args.optimizer == 'noam',
save_checkpoints_topk=10 if is_transformer else 1)
scheduler.save_checkpoint(quantized_model, quantized_model_path, remove_old=False)
#model.module.state_dict()->model.state_dict()

#2、using distiller
model = model.cuda()
model.eval()

quantizer = PostTrainLinearQuantizer(model)

quantizer.prepare_model()

apputils.save_checkpoint(0, 'model', model, optimizer=None,
name='quantized_model_name',
dir='quantized_model_path')
`