luka-group / lattice Goto Github PK

hi，i know in your code，it will choose the best checkpoint for evaluation in Totto. And because you don't open source the HiTab code, i reproduce the HiTab github's code, but in HiTab github, its defalut setting is choosing the last checkpoint for evaluation, and i wanna know the way you do in HiTab dataset, thanks ~

Hi!
First of all, thank you for the great work!

I have a quick question about type_edges that you added in structural attention.
Is type_edges fixed as ((1, 1), (2, 2), (1, 2), (2, 1), (3, 1), (3, 2), (1, 3), (2, 3))?
If it is, then why?
If it is not, then where can I find type_edge values of each input?

Thank you!
Hope you have a great one!

Best,
Paul

Hi,
I got below message when I run bash run.sh

Have you got this kind of message when you run bash run.sh?
Estimation Time gives me like 300 hours to run..?

Thanks

Best regards,
Paul

Dear author, I have only seen the codes related to Totto, but where are the codes related to HiTab? Can you provide the codes like Totto, or do I need to replace the pre-trained model structure in the HiTab github myself. Looking forward to your reply, thank you.

Traceback (most recent call last):
File "train.py", line 569, in
main()
File "train.py", line 503, in main
train_result = trainer.train(resume_from_checkpoint=checkpoint)
File "/home/subtlbot/.local/lib/python3.8/site-packages/transformers/trainer.py", line 1316, in train
tr_loss_step = self.training_step(model, inputs)
File "/home/subtlbot/.local/lib/python3.8/site-packages/transformers/trainer.py", line 1849, in training_step
loss = self.compute_loss(model, inputs)
File "/home/subtlbot/.local/lib/python3.8/site-packages/transformers/trainer.py", line 1881, in compute_loss
outputs = model(**inputs)
File "/home/subtlbot/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/home/subtlbot/ToTTo/Lattice/model/modeling_t5.py", line 1557, in forward
encoder_outputs = self.encoder(
File "/home/subtlbot/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/home/subtlbot/ToTTo/Lattice/model/modeling_t5.py", line 919, in forward
inputs_embeds = self.embed_tokens(input_ids)
File "/home/subtlbot/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/home/subtlbot/.local/lib/python3.8/site-packages/torch/nn/modules/sparse.py", line 158, in forward
return F.embedding(
File "/home/subtlbot/.local/lib/python3.8/site-packages/torch/nn/functional.py", line 2199, in embedding
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument index in method wrapper__index_select)

Hi, I tried to install this in a conda environment and when running with the current requirements.txt file generates an error when importing transformers (due to a mismatch of installed packages. I resolved it by removing the versions from all transformers-related packages and installing currently last versions:

transformers==4.25.1
datasets==2.7.1
sentencepiece==0.1.97

I suggest you to update the requirements file so it can be installed and run properly

Cheers,

David

Excuse me, how does the evaluation index blue in your article become the full version of bluet to upgrade to the evaluation index value in your article?

Hi,
First of all, thank you for the great work!
I'm a newbie in transformers and looking into some tasks like you guys did and reached your paper and code.

I have a quick question about custom seq2seqtrainer (trainer_seq2seq.py) that you created in model directory.
This might be a stupid question, but...
Is the prediction_step in the trainer_seq2seq.py the function for training with putting type_ids, row_ids, and col_ids into the model?

I mean, I followed your code line by line, and now I'm trying to put an extra param like type_ids as yours in my T5ForConditionalGeneration by using normal transformers.Trainer, but gives me an error like below.

ValueError: Unable to create tensor, you should probably activate truncation and/or padding with 'padding=True' 'truncation=True' to have batched tensors with the same length. Perhaps your features ( type_ids in this case) have excessive nesting (inputs type list where type int is expected).

It seems that this error occurs using the general Trainer library, but I want to confirm just in case.

Thanks
Hope you have a happy new year
Paul

hi, the input format of my code in Totto is the same as that in your paper . I also shuffle the row and the column of table，but the baseline（T5-base）result only drops 0.2 -1.0 in bleu and parent, even if I exchange the order of metadata and table. In your parper, the T5-base drops 4.5 bleu. I don't know what's wrong with it. thanks.
here is one example:
before shuffle：
'page_title> List of Governors of South Carolina /page_title> section_title> Governors under the Constitution of 1868 /section_title> table> cell> 76 h> # /h> h> 75 /h> h> 74 /h> /cell> cell> Daniel Henry Chamberlain h> 76 /h> h> Governor /h> /cell> cell> December 1, 1874 h> 76 /h> h> Took Office /h> /cell> /table>'
after shuffle:
'section_title> Governors under the Constitution of 1868 /section_title> table> cell> Daniel Henry Chamberlain h> Governor /h> h> 76 /h> /cell> cell> December 1, 1874 h> Took Office /h> h> 76 /h> /cell> cell> 76 h> # /h> h> 74 /h> h> 75 /h> /cell> /table> page_title> List of Governors of South Carolina /page_title>'