wenhuchen / kgpt Goto Github PK

Hi, I have a problem when I am trying to evaluate all the fine-tuned models using graph encoder. It gives me the following errors:
torch.nn.modules.module.ModuleAttributeError: 'GraphGatedTransformerDecoder' object has no attribute 'beam_search' │
And then I inspect the Model.py, I find out that there is indeed no 'beam_search' under 'GraphGatedTransformerDecoder'. But I see that in all the eval_graph_*_all.sh files, you set the '--beam_size 2' which means that it is using beam_search. I am really confused about this. Thanks!

bash scripts/webnlg/eval_sequence_webnlg.sh 0 challenge checkpoint_webnlg/checkpoint_finetune_sequence_head8_layer6_GPT2_maxfact12/model_ep20.pt We didn't change the code. File "/home/xushoukai/KGPT-webnlg/code/Model.py", line 113, in get_hypothesis
hyp.append(self.next_ys[j+1][k])
IndexError: tensors used as indices must be long, byte or bool tensors

When I use
"bash scripts/webnlg/eval_sequence_webnlg.sh 0 test checkpoint_webnlg/checkpoint_finetune_sequence_head8_layer6_GPT2_maxfact12_from_ep14/model_ep30.pt"
to reproduce the results ,and I find when I set beam_size >1 (using beam_search),then every word decode as "!",but when I used beam_size = 1 (using greedy_decode), it's back to normal.
I want to know how I should use beam_ Search for decoding

When I tried to run:
bash scripts/webnlg/finetune_graph_e2enlg_from_wikidata.sh 0 checkpoint_wikidata/checkpoint_sequence_head8_layer6_GPT2_maxfact12/model_ep14.pt

It gave me the following error.
Traceback (most recent call last): File "code/run.py", line 176, in <module> load_my_state_dict(model, reloaded) File "code/run.py", line 82, in load_my_state_dict own_state[name].copy_(param) KeyError: 'module.embedding.word_embeddings.weight'

Do you know why this error happens, thanks!

Hi, when I try to run the "run.py", I get the following error.
Segmentation fault (core dumped)
both the Sequence Encoder and Graph Encoder

Anyone have a solution to this error please let me know

Hey, downloaded your pretrained weights from recent Google Drive link and ran script specified in reproducing results. Got this error:

Traceback (most recent call last):
  File "code/run.py", line 453, in <module>
    result = model.module.beam_search(data[:-2], tokenizer, n_bm=args.beam_size, banwords=banwords, max_token_seq_len=args.max_dec_len)
  File "/home/ilya/repos/KGPT/code/Model.py", line 646, in beam_search
    active_inst_idx_list = beam_decode_step(self, inst_dec_beams, len_dec_seq, active_inst_idx_list,
  File "/home/ilya/repos/KGPT/code/Model.py", line 883, in beam_decode_step
    dec_partial_seq = [inst_dec_beams[idx].get_current_state() 
  File "/home/ilya/repos/KGPT/code/Model.py", line 883, in <listcomp>
    dec_partial_seq = [inst_dec_beams[idx].get_current_state() 
  File "/home/ilya/repos/KGPT/code/Model.py", line 34, in get_current_state
    return self.get_tentative_hypothesis()
  File "/home/ilya/repos/KGPT/code/Model.py", line 103, in get_tentative_hypothesis
    hyps = [self.get_hypothesis(k) for k in keys]
  File "/home/ilya/repos/KGPT/code/Model.py", line 103, in <listcomp>
    hyps = [self.get_hypothesis(k) for k in keys]
  File "/home/ilya/repos/KGPT/code/Model.py", line 114, in get_hypothesis
    hyp.append(self.next_ys[j+1][k])
IndexError: tensors used as indices must be long, byte or bool tensors

hi, thanks for publish this great work!

i have some question about data use

1. in test set of web nlg, there are some example that do not have a text label, how did you evaluate them
2. webnlg have multi text label for a triple set(for example: triple_set1 can map to [text1,text2,text3]), during training, did you regard one example as multi training data(for above example, we got 3 training data :[(triple_set1,text1) ,(triple_set1,text2),(triple_set1,text3)]) or randomly chose one of the text label for one example? if a triple set have multi text label during training, will the model get confused because the model learn triple_set1 map to text1, but when meet the example triple_set1 map to text2, it will lose some bleu score and it will swing between text1 and text2. what is the standard evaluate way in this task(I am new to this area, sorry for some stupid question).

looking forward to your replay, any help would be appreciated!

Nice job! I have several questions as follow:
(1) I found some differences between code and paper, e.g., property embedding hasn't been used, and various parameter values (batch size/learning rate) are not the same in different experimental setup (training/challenge/a few shots). I wonder whether the SOTA released in the original paper can be implemented by following the settings in the scripts.
(2) I run the sequence-encoder-few-shot-1% experiment at WEGNLG dataset. Only the batch size is modified to 16 due to insufficient resources, and the epoch becomes 4000 correspondingly for the same total steps. I achieve 19.9 BLEU (paper: 24.7). Is there any important information or tricks i miss?
(3) The proposed model has achieved significant performance under zero-shot and few-shot settings. Could you show more details about how many training instances could obtain a competitive results against other baselines (Fully-Supervised Results)?
Thanks a lot!