Regarding the training iteration details in the Stereo Anything paper

[Wst-sd]

Hello, I’m not entirely clear about the training iterations for the mixed datasets mentioned in the paper, which is stated as 81,000. Does this refer specifically to mix10, or does it apply uniformly to all mixes from mix1 to mix11? Additionally, I noticed that in the configuration files, nmrf_swint_mixdata.yaml sets NUM_EPOCHS=1, while fstereo_mixdata.yaml sets NUM_EPOCHS=10. Which one should I use?