diffuser.T is being ignored

[markf94]

I've been trying to change the total number of timesteps of my RFDiffusion runs and I noticed that changing diffuser.T in the config file has absolutely no impact on the actual run. It's always using 50 timesteps and I only have the ability to stop earlier with inference.final_step. Is that expected behaviour? Is the diffuser.T fixed at training time? If that's the case, why is it listed as a parameter in the config file?

[roccomoretti]

Where are you changing the value? If I pass diffuser.T=XYZ on the run_inference.py command line, I get a "WARNING: You are changing diffuser.T from the value this model was trained with. Are you sure you know what you are doing?" message printing after the "USING MODEL CONFIG" lines, and then the "Timestep ..., input to next step:" lines count down from my specified values. (Though the 'USING MODEL CONFIG: self._conf[diffuser][T] = 50' does get printed out.)

[l-mizrahi]

I ran into this issue as well. Any changes to the 'model', 'diffuser' or 'preprocess' configs need to be specified in the command line. The model checkpoint contains the default values for those configs. When the model is loaded the values from the checkpoint overwrite the values in the config. Then a check is done to see what Hydra cmd line overrides have been specified and those values now become the new config values. Basically any config changes to 'model', 'diffuser' or 'preprocess' configs need to be done via the command line. Seems to be intentional.