I need the exact commit id information for comparison.

Hello, thank your contribution, which is excellent work.
Could you send me the best checkpoint ? I want to use it to generate some translations for analysis.
Thanks again.
E-mail: [email protected]

May I ask whether it is tokenized BLEU or detokenized BLEU in G-transformer's paper?

论文文链接是这个 Contextualized Rewriting for Text Summarization

Hello,
I would like to test your model on Voita's contrastive test set, could you please provide a script to score translations?

I modified "--max-tokens" to 2048 in "prepare-randinit.sh" to prepare all data. Then I started to train, but there is an error about "max_positions". Do you know how to fix this problem?

Furthermore, is there anything else I should pay attention to?

Acually, I want to know when the input length is expanded to 2048 tokens, d-BLEU will be declining? Have you ever done an experiment like this？What is the current maximum input length（input tokens） given that d-BLEU is stable？

您好，如果我想复现您论文中的strong baseline-Transformer on sent (baseline) 上面的结果，该如何设置arch、task等fairseq-train中的参数？
此外，该baseline和SentNMT有什么区别？

  File "train.py", line 14, in <module>
    cli_main()
  File "/data1/mzlv/g-transformer/fairseq_cli/train.py", line 347, in cli_main
    cli_main_helper(args)
  File "/data1/mzlv/g-transformer/fairseq_cli/train.py", line 374, in cli_main_helper
    fn=distributed_main, args=(args,), nprocs=args.distributed_world_size
  File "/home/mzlv/anaconda3/envs/gtrans_test_cu10/lib/python3.6/site-packages/torch/multiprocessing/spawn.py", line 171, in spawn
    while not[ spawn_context.join():](https://wx.qq.com/cgi-bin/mmwebwx-bin/webwxcheckurl?requrl=http%3A%2F%2Fspawn_context.join()%3A&skey=%40crypt_fe4869f3_6be60a4be54a35ffb40698e1efe786d6&deviceid=e134100181037250&pass_ticket=undefined&opcode=2&scene=1&username=@b86617e379dbf55d07952dadda5d9f6098165536321e99eb1687bed0ccd16120)
  File "/home/mzlv/anaconda3/envs/gtrans_test_cu10/lib/python3.6/site-packages/torch/multiprocessing/spawn.py", line 118, in join
    raise Exception(msg)
Exception: 

-- Process 1 terminated with the following error:
Traceback (most recent call last):
  File "/home/mzlv/anaconda3/envs/gtrans_test_cu10/lib/python3.6/site-packages/torch/multiprocessing/spawn.py", line 19, in _wrap
    fn(i, *args)
  File "/data1/mzlv/g-transformer/fairseq_cli/train.py", line 336, in distributed_main
    main(args, init_distributed=True)
  File "/data1/mzlv/g-transformer/fairseq_cli/train.py", line 111, in main
    extra_state, epoch_itr = checkpoint_utils.load_checkpoint(args, trainer)
  File "/data1/mzlv/g-transformer/fairseq/checkpoint_utils.py", line 138, in load_checkpoint
    strict=not getattr(args, "load_partial", False)
  File "/data1/mzlv/g-transformer/fairseq/trainer.py", line 334, in load_checkpoint
    self.optimizer.load_state_dict(last_optim_state, optimizer_overrides)
  File "/data1/mzlv/g-transformer/fairseq/optim/fairseq_optimizer.py", line 72, in load_state_dict
    self.optimizer.load_state_dict(state_dict)
  File "/home/mzlv/anaconda3/envs/gtrans_test_cu10/lib/python3.6/site-packages/torch/optim/optimizer.py", line 111, in load_state_dict
    raise ValueError("loaded state dict has a different number of "
ValueError: loaded state dict has a different number of parameter groups

篇章级微调的时候加载句子级模型是正常的。如果微调中断，加载checkpoint_last.pt会报以上错误，是否g-transformer微调的时候中断只能从头训练，还是有额外的参数需要指定？

$data/concatenated_en2de_$subset_${lang}

Hi author,I can't use scripts for the baseline of doc transformer. After cloning the new fairseq and copied the baseline scripts to fairseq root folder. but fairseq/models/transformer.py does not exist in the newly downloaded fairseq folder. so I just appending the model setting to fairseq/models/transformer/transformer_legacy.py. Then I get an error. Could you please explain what I am missing or what is wrong? Thanks.

Hi author, can you share more details of the G-Transformer+BERT experimental settings? e.g., which version of the pre-trained BERT model (base or large, cased or uncased) was used, as well as the BERT decoding during inference to compute the s-BLEU scores. Thanks.

I appreciate your effort to provide scripts for the baseline of the sent transformer.
#2 (comment)

the doc transformer is important too.

after training sent and doc models, i miss this error.
Traceback (most recent call last): File "/home/dell/anaconda3/envs/g-transformer/lib/python3.7/runpy.py", line 193, in _run_module_as_main "__main__", mod_spec) File "/home/dell/anaconda3/envs/g-transformer/lib/python3.7/runpy.py", line 85, in _run_code exec(code, run_globals) File "/home/dell/桌面/g-transformer/fairseq_cli/generate.py", line 356, in <module> cli_main() File "/home/dell/桌面/g-transformer/fairseq_cli/generate.py", line 352, in cli_main main(args) File "/home/dell/桌面/g-transformer/fairseq_cli/generate.py", line 40, in main return _main(args, sys.stdout) File "/home/dell/桌面/g-transformer/fairseq_cli/generate.py", line 169, in _main hypos = task.inference_step(generator, models, sample, prefix_tokens) File "/home/dell/桌面/g-transformer/fairseq/tasks/fairseq_task.py", line 358, in inference_step return generator.generate(models, sample, prefix_tokens=prefix_tokens) File "/home/dell/anaconda3/envs/g-transformer/lib/python3.7/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(*args, **kwargs) File "/home/dell/桌面/g-transformer/fairseq/sequence_generator.py", line 176, in generate return self._generate(sample, **kwargs) File "/home/dell/桌面/g-transformer/fairseq/sequence_generator.py", line 331, in _generate _fill_frame(tokens, src_tokens) File "/home/dell/桌面/g-transformer/fairseq/sequence_generator.py", line 283, in _fill_frame docfrm = _align_frame(src[i]) File "/home/dell/桌面/g-transformer/fairseq/sequence_generator.py", line 267, in _align_frame assert len(frame) % 3 == 0 AssertionError