getting all parameters for a given channel

[bantin]

Hi jaxley folks! I recently implemented a custom channel in jaxley (a voltage indicator model). I'd like the option to make all parameters trainable for a given channel. Is there a way to do this? Right now, I have a big long list of parameters in my fitting code, and I call make_trainable for each param. The code snippet looks like this:

    if fit_params == 'all':
        params_list = [
        'k_gm', 'k_mg', 'v_half_alpha', 'v_half_beta',
        'k_alpha', 'k_beta', 'k_nr', 'k_rn',
        'k_rg', 'k_gr', 'fg_gain', 'fm_gain', 'fn_gain', 'fr_gain',
        'f_base'
        ]
    elif fit_params == 'gains_only':
        params_list = ['fm_gain', 'fn_gain', 'f_base']
    else:
        raise ValueError("fit_params must be 'all' or 'gains_only'")

    for param_name in params_list:
        cell.branch(0).comp("all").make_trainable(f"{gevi_name}_{param_name}")

This is a little ugly and hard to maintain, so is there a way to just make all params trainable for a given channel?

Thanks!

[michaeldeistler]

Hi Ben,

unfortunately this is not possible at the moment. We could consider allowing a list to be passed to make_trainable, would this work for you?

Michael

[bantin]

Hi Michael,

Thanks for the reply! Passing a list to make_trainable would certainly be a slight improvement, but the API I was hoping for would allow a user to train all parameters for a given channel without explicitly knowing the names of those params. My thinking here is that, as a user, I might want to train all parameters for say, the sodium channel, without looking into the source code and manually finding all of the parameters which are part of that channel. Does this made sense? I could try making a PR for this if you think this functionality would be useful to others.

[michaeldeistler]

I see. The following achieves this:

for key in Na().channel_params:
    cell.make_trainable(key)

I am a bit unsure about having dedicated functionality in Jaxley that does this. It would be an additional method to maintain, to document, to test, and for users to learn. As such, I tend to think that the above solution is cleaner (and more flexible, e.g., for also making initial states trainable via .channel_states).

What do you think? Does the above have a limitation that I am not seeing?

[bantin]

That makes sense to me! I kept trying to access the channel which was part of my cell. I was trying to do something like:

for key in cell.branch(0).channel:

which I couldn't get to work -- probably because I just wasn't doing it right. But your way is cleaner anyway. Thanks!

[michaeldeistler]

Happy to hear that!

Your method would work as follows:

for key in cell.branch(0).channels[0]:
  ...

In other words, .channels is a List in which all channels are stored.

Michael