training results

[oneChubby]

Why are my training results so far from the experimental results (Kitchen Dataset)

[PanagiotisP]

Hi!
Hm, that's surprising. I can't give a clear answer without more information, but since it seems like the technique hasn't run at all, I'd bet is some kind of command-line argument parsing.
Did you run it with the configuration that uses the new options, as described at the full_eval script, as shown here?

[oneChubby]

Thank you very much for your answer, after modifying the parameters, the effect in the paper has been achieved, but I would like to ask which of the three is this printed PSNR

[PanagiotisP]

Generally, you want to keep an eye on test PSNR, so the top one.
However, the numbers on the paper might differ (generally, be lower) ,as

1. the k-means quantisation happens after the last iteration, so its effects are not printed on the screen and
2. the results are generated by the full_eval.py script, which trains a model, renders images, saves them as png and then calculates the metrics based on these images. Any steps (training, rendering, metrics calculation) can be skipped with the respective flag. This is the standard way to evaluate and typically you expect some numbers to be lower