I am getting better prediction result for inshot-inference from checkpoints than prediction from saved model from same training process. What would be reason behind this?
Any help appreciated.
The Python code uses beam_width 1024, while the clients use 500. That could be the difference.
Thanks for the reply.
I did not get your point, as i did not make any changes in the code and still it gave me good result from checkpoint but not same from model.
Correct me if i am wrong,
Do i need to change BEAM_WIDTH parameter to 1024 in client.py and then re-run the train code and make prediction from that saved model?
Yes, except you don’t need to retrain, just change that, rebuild the client and run it again.
– reuben