Open hamaadtahir opened 4 years ago
Please try the new requirements.txt and let me know how it does
Thankyou once again for your help. Now another issue that popped up is missing file "mode_to_cat_to_model_ids.pickle". It is referred in the line 228 of "highres_sampler.py"
Added needed pickle files to pickle_files/
and changed paths. Thanks for bringing these to my attention.
Thankyou, how many epochs did you train it for?
I believe that the default settings should be sufficient, if that doesn't quite reach reported performance try upping the # of epochs by a factor of 2-3
by default it is infinity
This was updated in yesterday's reorg. Make sure you have the latest version pulled
Thanks, now just a final thing, how do I test a novel category image using a prior once a model has been trained, and where does it save the trained model.
It saves to the path provided in --save-path
(e.g. --save-path my_refiner.h5
)
To test the model with a new image/prior simply load the model and feed in the images/voxel priors as in https://github.com/BramSW/iccv_2019_few_shot_3d_wallace/blob/master/train_iterative_RGB_refiner.py#L191
I am sorry i didnt follow your response regarding the loading of model using the same script that was earlier used for training. Could you give a hint for what argument switches to use for testing using new image/prior when running the same script file
And does this "--num-refine-iters" switch refers to the iterative training
what does --num-refine-iters means and what number should I set to this value
Hi, the --num-refine-iters
refers to how many iterations to run during training/inference. As noted in the paper, we found just 1 iteration to suffice for most performance purposes, but the higher-iteration refiners do exhibit interesting performance as noted in our figures. Hope that helps!
Required packages and their versions are missing from readme