Couldn't reproduce TSNE result of LFADS inferred initial generator state

[ecosan327]

I used the integrator RNN tutorial to produce the input results for LFADS tutorial. All parameters in integrator RNN and LFADS tutorial are not changed. But when the last step in using TSNE visualization to produce LFADS inferred initial generator state, I couldn't get the shape like the one in LFADS tutorial code(like a circular, concentrated line). Instead, no matter how many times I tried running the tutorial, my result shows a spread, non-structural shape like this below image:

Does anyone also have this problem?

my packages version: ubuntu 18.04 jax 0.2.20 jaxlib 0.1.71+cuda111 scikit-learn 0.24.2

[ecosan327]

also in the visualization part, I found in general the training result is not as good as the one shown in tutorial web page. Like the average LFADS filtering R^2 value is lower. my result :

Not sure this is also related to the TSNE problem?

[yanistazi]

Try to remove the controller part from the LFADS architecture and you will be able to recover the structural shape of the original tutorial ( you can just set the controller dim and inferred input dim to be 0). It is also a realistic scenario for naturalistic behavior where there is no clear cue