John Muradeli

Too rusty on RNNs to validate any of this, I fear. [SE](https://stats.stackexchange.com/) might help. I'm also no longer developing this repository, but I'm open to reviewing merge-ready contributions.

Most likely you don't have TensorFlow installed; I've made the error message more informative in the latest version. Try via `pip install git+https://github.com/OverLordGoldDragon/see-rnn` and paste the error here if it...

It cannot be the same error; try `pip uninstall see-rnn` then `pip install git+https://github.com/OverLordGoldDragon/see-rnn`.

Unsure what the problem is, need debug info: 1. In `__init__.py` (`import see_rnn; print(see_rnn.__file__)`), comment lines 26, 33-35 (`try, except, pass`), and unindent 27-32. Save file. 1a) restart kernel, then...

Possibly, yes. The last tested TF version of `see_rnn` is 2.3.0, and I am no longer maintaining it for future compatibility. Version bugfix PRs are welcome. Thanks for the report,...

Thanks for the report. I no longer actively maintain this repository for future TF versions. This bug stemmed from 2.4's change of `train_step` logic, which `see_rnn` utilized in `get_gradients`. The...

Is this based on any Kymatio example? I'd try eliminating any `None`s from tensor shapes as that's the source of error.

Added `Q=16`, even greater TS difference (more `n1` are batched per `n2`).

@sharvil Wonderful. **WEIGHT INITIALIZATION**: [author repo](https://github.com/Sunnydreamrain/IndRNN_Theano_Lasagne) `kernel`: authors recommend Normal w/ small std. In my application, `'glorot_normal'` and `'glorot_uniform'` have yielded comparable results (didn't try others); note that normal is...

@sharvil I see your commit history - good progress so far, though I hope LayerNorm support is planned as it can make IndRNNs vastly more powerful for very long sequences....