accuracy stays 0 while training, and the top_actions are [4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ..., 0, 0] about hrl-re HOT 6 CLOSED

@misaki-sysu RL suffers from the cold start problem in a large-scale state or action space so that it can hardly learn from the sparse reward, especially for hierarchical RL in such a complex task. So pretraining can prevent the agent from exploring many unreasonable cases (e.g. multiple source entities in one relational triple) thus making the training possible.

from hrl-re.

Did you pretrain the model first? e.g.

python3 main.py --datapath ../data/NYT10/ --pretrain True

Please refer to #1 for the results of pretraining

from hrl-re.

@truthless11 this result I mention above had no pretrain. I had pretrained it and the accuracy , F1 score stayed 1 while pretraining, is it normal?

from hrl-re.

Yes, it is. When pretraining, the output of training data is useless as it will always print out 1.0 F1 score. Please be patient and wait for the output of test data.

from hrl-re.

Thanks! I will wait for the pretrain result.
But, forgive my ignorance, why can't I just train the model without pretraining ?

from hrl-re.

Thanks for you answer!

from hrl-re.