-
Notifications
You must be signed in to change notification settings - Fork 132
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Are these normal results? #25
Comments
No... these are not normal results... Some tips for debugging: |
actually, this reminds me of #21 -- might be worth checking. |
Thanks for your reply! |
@ChorlingLau were you able to get better results? My unconditional generation with either dataset still seems repetitive and incoherent. |
I got coherent sentences in unconditional generation. And after pulling the newest code I get better results in task 'control_pos' but 'control_tree' still not. unconditional:
control_pos:
|
@ChorlingLau hmm, to me that seems like "barely coherent" - but admittedly better than what I'm getting. This is using the commands in the readme for training and decoding right? @XiangLi1999 any thoughts on the quality of the outputs we've been getting? Seems like it's way off from your results, what could we be doing wrong? Have you used this actual codebase/data to train a model yourself? If so, what are your results/hyperparams? Thanks a lot. |
update: i was able to train a much better model that generates nice text, using much larger batch sizes (640).
|
I want to know your detailed commands in the last step of running scripts/infill.py and how to generate unconditionally. Thank you very much |
I think one important detail is to train a PAD based model via "--padding_mode pad ". Essentially, this makes a sentence start from [START] [sentence content] [END] [PAD*n]. Training by block looks fine for unconditional generation, but for controls, I think there are too much noise when learning the classifier. |
|
Thanks,When finally generating controllable text, python scripts/infill.py --model_path {path-to-diffusion-lm} command, which path does "path-to-diffusion-lm" refer to, the first step model path, or the last but one model generated from the classifier? |
@Licy1999 It's 1st step model. The classifier path should be modified in code file infill.py |
hello,I am filling as you said infill.py changes the training path of the classifier model, code:model_control = Classifier_POS.from_pretrained('../classifier_models/e2e-tgt-tree_e=6_b=10_m=bert-base-uncased_wikitext-103-raw-v1_101_wp_None').cuda(),and I confirm the file of oil training classifier in this folder.but it seems that the loading fails. An error like this is reported. Can you tell me how to handle it. |
Hi, I got some results by following README and running
infill.py
with task "control_pos". Each sentence is 64-word-long and doesn't seem to quite match the POS, as shown in the following example.And I have the same question on other tasks except for "control_length".
Could you please tell me whether these are normal results? If yes, how can I get a decent sentence like you show in paper?
Thanks a lot!
The text was updated successfully, but these errors were encountered: