运用prefix_projection 方法训练test acc不变一直是62.1

[yh351016]

12/07/2021 19:30:19 - INFO - training.trainer_base - ***** Epoch 12: Best results ***** 12/07/2021 19:30:19 - INFO - training.trainer_base - best_epoch = 0
12/07/2021 19:30:19 - INFO - training.trainer_base - best_eval_accuracy = 0.6217125382262997
12/07/2021 19:30:19 - INFO - training.trainer_base - epoch = 12.0
OrderedDict([('best_epoch', 0), ('best_eval_accuracy', 0.6217125382262997), ('epoch', 13.0)])
{'loss': 0.7488, 'learning_rate': 0.006054054054054054, 'epoch': 13.51}

[Xiao9905]

Hi, prefix projection needs a careful hyper-parameter tuning, and in practice we find it does not work for every tasks.

[yh351016]

BoolQ and RTE result on the table of github is not same with paper report

[Xiao9905]

As we stated in the README.md, the hyperparameters provided in the repo is reimplemented with 3090 GPUs, rather than A100s (on which our paper results come from).
Since experiments reported in our paper are all conducted on NVIDIA DGX-A100 servers (which might be difficult to acquire), we reimplement P-tuning v2's results on BERT-large/RoBERTa-large with:

[yh351016]

As we stated in the README.md, the hyperparameters provided in the repo is reimplemented with 3090 GPUs, rather than A100s (on which our paper results come from). Since experiments reported in our paper are all conducted on NVIDIA DGX-A100 servers (which might be difficult to acquire), we reimplement P-tuning v2's results on BERT-large/RoBERTa-large with:

thanks,I think it will only affect the speed. Does it have anything to do with the result？

[Xiao9905]

As is in the README.md
We notice that the best hyper-parameters can be sensitive to your server environment and package version.

If you do not have the exact same environment, we highly recommend you to run hyper-parameter search in your environment
based on our example hyper-parameter search script in search_script and result collection scripts search.py.