shapes (128418,3) and (0,124) not aligned: 3 (dim 1) != 0 (dim 0)

[ElifAyten]

Hello ,

I am implementing an SLDS model using the ssm library and experimenting with different hyperparameters. My dataset contains both neuronal firing rates and external inputs (footshock, pupil size, speed). I want to use both as influences on the model by setting the transition type to recurrent.

However, when I include external_inputs and use transitions="recurrent", I encounter the following shape mismatch error:
`---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
in <cell line: 0>()
8 )
9
---> 10 elbos_recurrent, _ = rslds_recurrent_inputs.fit(
11 [firing_rates_normalized],
12 inputs=[external_inputs], # footshock, pupil, speed

9 frames
/usr/local/lib/python3.11/dist-packages/ssm/util.py in wrapper(self, datas, inputs, masks, tags, **kwargs)
109 tags = [tags]
110
--> 111 return f(self, datas, inputs=inputs, masks=masks, tags=tags, **kwargs)
112
113 return wrapper

/usr/local/lib/python3.11/dist-packages/ssm/lds.py in fit(self, datas, inputs, masks, tags, verbose, method, variational_posterior, variational_posterior_kwargs, initialize, discrete_state_init_method, num_init_iters, num_init_restarts, **kwargs)
778 # Initialize the model parameters
779 if initialize:
--> 780 self.initialize(datas, inputs, masks, tags,
781 verbose=verbose,
782 discrete_state_init_method=discrete_state_init_method,

/usr/local/lib/python3.11/dist-packages/ssm/util.py in wrapper(self, datas, inputs, masks, tags, **kwargs)
109 tags = [tags]
110
--> 111 return f(self, datas, inputs=inputs, masks=masks, tags=tags, **kwargs)
112
113 return wrapper

/usr/local/lib/python3.11/dist-packages/ssm/lds.py in initialize(self, datas, inputs, masks, tags, verbose, num_init_iters, discrete_state_init_method, num_init_restarts)
165 num_init_restarts=1):
166 # First initialize the observation model
--> 167 self.emissions.initialize(datas, inputs, masks, tags)
168
169 # Get the initialized variational mean for the data

/usr/local/lib/python3.11/dist-packages/ssm/util.py in wrapper(self, datas, inputs, masks, tags, **kwargs)
109 tags = [tags]
110
--> 111 return f(self, datas, inputs=inputs, masks=masks, tags=tags, **kwargs)
112
113 return wrapper

/usr/local/lib/python3.11/dist-packages/ssm/emissions.py in initialize(self, datas, inputs, masks, tags)
447 def initialize(self, datas, inputs=None, masks=None, tags=None):
448 datas = [interpolate_data(data, mask) for data, mask in zip(datas, masks)]
--> 449 pca = self.initialize_with_pca(datas, inputs=inputs, masks=masks, tags=tags)
450 self.inv_etas[:,...] = np.log(pca.noise_variance)
451

/usr/local/lib/python3.11/dist-packages/ssm/util.py in wrapper(self, datas, inputs, masks, tags, **kwargs)
109 tags = [tags]
110
--> 111 return f(self, datas, inputs=inputs, masks=masks, tags=tags, **kwargs)
112
113 return wrapper

/usr/local/lib/python3.11/dist-packages/ssm/emissions.py in _initialize_with_pca(self, datas, inputs, masks, tags, num_iters)
186
187 # Compute residual after accounting for input
--> 188 resids = [data - np.dot(input, self.Fs[0].T) for data, input in zip(datas, inputs)]
189
190 # Run PCA to get a linear embedding of the data with the maximum effective dimension

/usr/local/lib/python3.11/dist-packages/ssm/emissions.py in (.0)
186
187 # Compute residual after accounting for input
--> 188 resids = [data - np.dot(input, self.Fs[0].T) for data, input in zip(datas, inputs)]
189
190 # Run PCA to get a linear embedding of the data with the maximum effective dimension

/usr/local/lib/python3.11/dist-packages/autograd/tracer.py in f_wrapped(*args, **kwargs)
46 return new_box(ans, trace, node)
47 else:
---> 48 return f_raw(*args, **kwargs)
49 f_wrapped.fun = f_raw
50 f_wrapped._is_autograd_primitive = True

ValueError: shapes (128418,3) and (0,124) not aligned: 3 (dim 1) != 0 (dim 0)`

this is my code snippet:
`import ssm
rslds_recurrent_inputs = ssm.SLDS(
D_obs, K, D_latent,
transitions="recurrent", # uses past neural activity + external inputs
dynamics="gaussian", # linear latent dynamics with Gaussian noise
emissions="gaussian_orthog", # observation model
single_subspace=True
)

elbos_recurrent, _ = rslds_recurrent_inputs.fit(
[firing_rates_normalized],
inputs=[external_inputs], # footshock, pupil, speed
num_iters=10
)
`

I expect the model to correctly process both neuronal firing rates and external inputs, allowing recurrent transitions to consider both past neural activity and external inputs.Is there a specific way I need to format inputs for recurrent transitions? Could this be due to how inputs is being handled during model initialization? Any suggestions on how to properly include both firing_rates_normalized and external_inputs in ssm.SLDS.fit()?

Thank you !!