prgumd / spikems Goto Github PK

Hi, thanks for your excellent work! I have two questions regarding IoU calculation.

(1) The IoU report from the paper is point-based(sparse) or dense mask as in traditional segmentation tasks? From this repo, it seems to be point-based.
(2) Is maxBackgroundRatio used to calculate IoU for the whole sequence? I set maxBackgroundRatio to be 1.5 in EVIMO_wall_seq00 and only 8% of events are used.

Thanks in advance!

I've applied for access on the Google drive but I haven't get your response. So I can't run your test script. Please help me.
PS: my google account is [email protected]

Whether the file named 'logging.py' is missing in the directory SpikeMS-main/config? When I run the command python test.py --crop --maxBackgroundRatio=1.5 --segDatasetType="MOD" --modeltype="unetRNN6Layer_noBlock" --maskDir="Test/MOD/seq_room1_obj1/masks" --datafile="Test/MOD/seq_room1_obj1/room1obj1-001.hdf5" --checkpoint="pretrainedModels/EVIMO-pretrained/out/checkpoint.pth.tar",it has a ModuleNotFoundError：No module named ‘config.logging’.

Hi,
I am very interested in your work.
Can I have your suggestions on reproducing training stage?
In your paper, I have a bit confused about your proposed spatio-temporal loss.

 Assume that I have an output spike tensor **e**=[batchsize, 2, 260, 346, timebins] and an masked target spike tensor **t**=[batchsize, 2, 260, 346, timebins]:

  (1) what does the spike project E(x)=sum_t (e(x)) mean? Is it equal to torch.sum(e, dim=4)? If it is, how can I implement binary cross-entropy for \lamada_bce (Eq.8)?

  (2) What is 1_f in Eq.9? Is it the mask of foreground spikes at t+delta_t or the mask labels at image domain?

  (3) How reduction is performed in the \lamada_bce and \lamada_spike (Eq.9)? Are they averaged over batchsize, height and width? 

It will be great if you share your implementation of your proposed spatio-temporal loss.

Regards！

Hi,

As a newcomer to SNN and SLAYER, I am a bit confused about the implementation of incremental prediction in SNN. Initially, I thought it referred to the outputs at different time steps during the same inference procedure. However, I am unable to understand how to display the outputs of SNN at each time step in a single inference procedure since the PyTorch version of SLAYER only provides a forward() function without any details of each timestep. Now, I am thinking that the result is achieved through multiple inference procedures, each taking different lengths of original spike trains as input (i.e., different ∆T_test values). I am not entirely sure if this is correct, could you please confirm?

I've applied for access on the Google drive but I haven't get your response. So I can't run your test script. Please help me.
PS: my google account is [email protected]

Thank you very much for providing an excellent paper and source code.

I am currently attempting to implement SpikeMS and was wondering if it would be possible for you to share the train.py file. The code does not have to be in a complete state; I am prepared to make necessary adjustments and fix any potential errors that may arise. We require the training code to train on our dataset.

I appreciate your support and look forward to your positive response.

Thank you in advance.