EventDownsampling polarity is set to 1 always

[agirbau]

Hello, thank you for the great work!

I was testing the EventDownsampling module, and it seems it's returning always the same polarity.

To reproduce:

import tonic

nmnist = tonic.datasets.NMNIST("./data", train=False)
events, label = nmnist[0]

transform = tonic.transforms.Compose(
    [
        tonic.transforms.EventDownsampling(sensor_size=nmnist.sensor_size,
                                           target_size=(12, 12),
                                           dt=0.01,
                                           downsampling_method="differentiator",
                                           noise_threshold=0,
                                           differentiator_time_bins=2)
    ]
)

aa = transform(events)

print(events['p'].min())

0

print(aa['p'].min())

1

[agirbau]

I think the problem comes in this bugfix:

bba009c#diff-cb9bbb0e9210b9747c8af2608a06b558a6a35d01dad19e767a2985e6ecf44117R50

- event_hist_pos = (np.maximum(event_histogram, 0)).clip(max=1)
- event_hist_neg = (-np.minimum(event_histogram, 0)).clip(max=1)
+ event_hist_pos = (np.maximum(event_histogram >= noise_threshold, 0)).clip(max=1)
+ event_hist_neg = (-np.minimum(-event_histogram >= noise_threshold, 0)).clip(max=1)

Where by comparing the event histogram to the noise thresh converts it to a boolean, and min(True, 0) is 0, so no negative events.

I'll start a PR to further discuss this.

[agirbau]

Opened a PR #299 as a quick hotfix, but requires further discussion on events sharing spatial and temporal slots.