MattCarrickPL
commented
12 months ago Looking at a broader rework of roll-off. From https://github.com/TorchDSP/torchsig/blob/v0.4.2/tests/test_transforms_figures.py#L86 the roll-off has a nominal passband frequency of 0.05, and a stop band frequency of 0.98. The passband frequency has a 50% chance to be 0 or 0.05, and the stopband frequency has 50% chance to 0.98 or 1. Unclear the impact to the 50% chance to use a different passband or stopband frequency.
Plotting the resulting filter. The top image is the LPF that is designed, and the bottom image is after upconversion to the center frequency:
Based on the paper, presumably the roll-off impairment is intended to generate these three effects:
Considering a rework to include:
- a broader randomization range for the passband and stopband edges
- removing 50% chance for changes to passband and stopband
- recalculating, or randomizing, the center frequency
- redefining the transition bandwidth to produce more reasonable length filters
The roll_off() transform needs another pass over the center_freq calculation:
https://github.com/TorchDSP/torchsig/blob/v0.4.2/torchsig/transforms/functional.py#L932
The formulation isn't immediately clear. Calculating some test values lowercutfreq = 0.125 and uppercutfreq=0.25, center_freq=-0.3125 which doesn't align with the input parameters. Also possible that I am not using the parameters within the proper units or normalization. Overall a better understanding of center_freq is needed which could result in a new calculation for it's value.