CompVis
Style not learned
Hello,
I've been trying to train the model with the default parameters that are provided in the script. But the results in the whole run are not quite as expected(up until 15k steps), the styles are not learned, as is showed below.
Do you have any suggestions or ideas about the potential cause of that?
Hi,
Thanks for the interest! 15k steps is indeed to few. As stated in launch_multistep_training.sh:
...
declare -a IMAGE_SIZES=(128 256 384 512 640 768)
declare -a TOTAL_STEPS=(50000 75000 100000 125000 150000 1000000)
...
for ((i=0;i<${#IMAGE_SIZES[@]};++i))
do
CUDA_VISIBLE_DEVICES=$DEV_NMBR python main.py --model_name=${MODEL_NAME} \
--total_steps=${TOTAL_STEPS[i]} \
...
done
you should train several rounds for 50000, 75000 ,100000 ,125000, 150000 ,1000000 iterations. The last round could be finished earlier though.
Best, Artsiom
Hi Artsiom!
Sorry, I've made a typo, the result I've got is from 150k steps instead of 15k step. I used exactly the script you attached and the last iteration(1m steps) has already started. Is that still normal though?
Best, Zilin
@ZilinIB I also met this problem, seems not learn the style at all with enough steps. Have you solved this?
@CrossLee1 No I'm still stuck there :(
@asanakoy I didn't continue to train the model after the round with 150k iteration since the losses are not getting optimized for a long time already.
@asanakoy This is the result of step_150600. Is it in line with expectations？
@CrossLee1 you mean 150rk step from the very beginning? @dimakot55 could you pelase take a look?
@CrossLee1 you mean 150rk step from the very beginning?
Yes, 150k step from the begining.
I had a similar turnout after 500,000 steps, showing very little to no improvement from the 150000 step example.
Guys, try changing hyperparameters a bit if it doesn't work. @dimakot55 is busy with NeurIPS submission right now and will be able to answer you after the deadline.
Hello, I also met this problem, seems not learn the style at all. Have you solved this?
