Encoder for numerical data instead of image_encoder

[mahsa-ebrahimian]

Hi
I am totally new in using Spike neural network and I am developing model using bindsnet. The issue is that my dataset is numerical not image, so when creating class dataset, I see there are only image_encoders available. Could you please let me know how to encode my numerical data as spikes?

[RaisoLiu]

In example "eth_mnist", the step is:
image(value=0~255) -> frequency(value=0~(255/4)Hz) -> calculate spike time

So I think you can :
numerica(value=0~max) -> frequency(value=0~x Hz) -> calculate spike time

I'm new in this area. Don't believe me 100%.

[SimonInParis]

Hi. To start with, you should consider rate-encoding. The Poisson encoder can encode any scalar value into spike trains, not just images. You can also have a rate based readout at the output layer. Don't hesitate to look at the network examples, as they contain useful mechanisms, even if they use images as inputs. Simon Le mar. 26 janv. 2021 à 19:36, mahsa <notifications@github.com> a écrit :

…

Hi I am totally new in using Spike neural network and I am developing model using bindsnet. The issue is that my dataset is numerical not image, so when creating class dataset, I see there are only image_encoders available. Could you please let me know how to encode my numerical data as spikes? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#450>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AKZ2IOBFHTYMEWTN6DKRKWLS34DUFANCNFSM4WT5IYKA> .

[Hananel-Hazan]

That is because your image type is double, try to cast your data to float, using something like:
image = torch.tensor(image, dtype=torch.float)

[mahsa-ebrahimian]

That is because your image type is double, try to cast your data to float, using something like:
image = torch.tensor(image, dtype=torch.float)

Thank you so much.
There is just one concern that when I apply encoding.poisson on my data it returns all the values of image equal to zero.
I chose the time =100, do you have any idea why is it happening?
@Hananel-Hazan

[Hananel-Hazan]

Without knowing what is the input, I can only guess that the input values are too small. Try to normalize the data between 0 to the maximum running iteration of your network. Note that the data given to the Poisson encoder must be non-negative float.

[mahsa-ebrahimian]

Without knowing what is the input, I can only guess that the input values are too small. Try to normalize the data between 0 to the maximum running iteration of your network. Note that the data given to the Poisson encoder must be non-negative float.

I tried increasing the time parameter and it seems to become fine. Does this parameter indicate the Poisson time intervals? How should it be chosen?

[Hananel-Hazan]

The Poisson function creates a spike train that you can feed to the network. The time parameter need to be the same as the time that the network spends on each input.
There is no exact method how to choose this number. It depends on the input intensity, network activity and many other parameters. The best way to find the right parameter is by trial and error.

[mahsa-ebrahimian]

Thanks for all the responses I could encode my data on Movie-Lens 100K dataset using below steps:

1. Reading my numerical data from CSV using, model inputs are places in x and outputs in y:

X= pd.read_csv(os.path.join(data_dir, 'mydata_x.data'), '\t', names=names,
                           engine='python')

y= pd.read_csv(os.path.join(data_dir, 'mydata_y.data'), '\t', names=names,
                           engine='python')

2. Converting my x and y dataframe into tensors using:

x_tensor = torch.tensor(X.values, ,dtype=torch.float)
y_tensor = torch.tensor(y.values, ,dtype=torch.float)

3. Encoding my numerical data using:

from bindsnet         import encoding

for i, (image,label) in enumerate(zip(x_tensor ,y_tensor)):
    
    encoded_image=encoding.poisson(datum=image, time=time, dt=dt)

[mahsa-ebrahimian]

@hockman1 I hope my solution works for you too.