Results not satisfactory after 280 epochs for flag_simple dataset

[Nagakiran1]

Hi Mark,
Thanks for outsourcing great repo.
I have used your repo with below configuration and i trained it for more than 280 epochs. my results were still not satisfactory.

--model=cloth --mode=all --rollout_split=valid --dataset=flag_simple --epochs=1500 --trajectories=100 --num_rollouts=100 --core_model=encode_process_decode --message_passing_aggregator=max --message_passing_steps=30 --attention=True --ripple_used=False --use_prev_config=True --model_last_run_dir=output/flag_simple/Thu-Mar--3-14-54-04-2022

I am trying to increase message passing layers 3 to 30, hoping it to work :-|

would you please direct me few steps to take :-|

[Nagakiran1]

If possible, would you please outsource, few models of cloth. @wwMark

[Yingqi7777]

Hi Nagakiran,

I am investigating the code too. Could you share your loss curve or result? My loss will go down to 1.5~2, not satisfying enough.

(Very cheerful to see someone else is doing the same research!)