Does t_intervals contain more time steps than expected?

[tdedring]

Within the Diffusion class, t_intervals are created like this:
if self.args.scheduler_type == 'uniform':
skip = self.num_timesteps // self.args.timesteps
t_intervals = torch.arange(-1, self.num_timesteps, skip)
t_intervals[0] = 0

To my understanding:
self.num_timesteps comes from the config with a default value of 1000.
self.args.timesteps comes from args; The default according to the ReadMe: STEPS=10; The default according to fast_ddpm_main.py:
"--timesteps", type=int, default=100, help="number of steps involved"
Is this intended, or did I get the code wrong?

Now the actual Question:
Under the condition self.num_timesteps=1000 and self.args.timesteps=10, this code returns 11 t_intervals.
Is this intended, or should it be replaced to return 10 t_intervals?

Thanks for the code!

[Sebastianjhx]

Apologies for the late reply and any confusion.

By saying "10 timesteps involved", we're referring to 10 intermediate time step, excluding the start timestep 0, so you could also describe it 11 timesteps.

In fact, you're free to choose any number of timestep for training based on your task and dataset. The core idea of FastDDPM is to focus training on specific time intervals rather than the entire time range, which better aligns the training and sampling processes.

Thanks for catching this!