search

keunhong / photoshape

111 stars 15 forks source link

How to retrain this material classifier? How to regenerate the training data? #14

Open XiangyuSu611 opened 4 years ago

XiangyuSu611 commented 4 years ago

Hi Keunhong,

I’m trying to run PhotoShape, and I want to retrain the material classifier with other texture data. But I couldn't find the code to regenerate the training data, I found some clues in /src/terial/clasifier/README.md.The brdf_classifier folder does not exist in the current code version, can you send me the relevant code?

thank you very much!

Xiangyu

keunhong commented 4 years ago

Hi Xiangyu,

You can find the code to generate the data here: https://github.com/keunhong/photoshape/tree/master/src/terial/classifier/data

It is a distributed setup so you need to run the collector server and then run a worker process on any number of machines.

Once you have the rendering you can make the LMDB format dataset by running this script: https://github.com/keunhong/photoshape/blob/master/src/terial/classifier/data/make_lmdb.py

XiangyuSu611 commented 4 years ago

Thanks for your reply!!! I will try it right away :D

XiangyuSu611 commented 4 years ago

Hi Keunhong，

Sorry to interrupt again...

I have found some new textures from POLIIGON , but I found that the new textures cannot be used in rendering. After I modify the materials.json file, it cannot be used. I guess it is because the texture data is stored in the database. Do I need to modify the materials table in the database? If necessary, how do I generate the final sql file?

Thank you!

Xiangyu

XiangyuSu611 commented 4 years ago

Hi Keunhong,

I encountered some problems again.

Now I have implemented the entire process of generating data. But when I run to here, I have found that some 'base_color_hist' don't exist. It's because I didn't add the json file describing the parameters when I created the new material data table. Will this affect the final generated data? If so, how do I get this json file?By the way, I use POLIIGON texture.

Thank you!

Xiangyu