yoyodyne
yoyodyne copied to clipboard
CUNY-CL
Small-vocabulary sequence-to-sequence generation with optional feature conditioning
Fixes: https://github.com/CUNY-CL/yoyodyne/issues/192 TQDM in the transducer's expert module doesn't pickle currently. So when doing multi-gpu training, pytorch gets annoyed. Since it's not really needed beyond training, I'm dropping it from...
Something I've been thinking about with expansion of library: a decent amount of the work we've been using involves application of inductive biases and teacher-prompted training to model architecture. Currently...
Issue when running multi-gpu training with edit action transducer: ``` Traceback (most recent call last): File "/home/salamander/anaconda3/envs/sigmorphon2024/bin/yoyodyne-train", line 8, in sys.exit(main()) File "/home/salamander/anaconda3/envs/sigmorphon2024/lib/python3.10/site-packages/yoyodyne/train.py", line 390, in main model = get_model_from_argparse_args(args,...
Seen in https://github.com/CUNY-CL/yoyodyne/pull/140 , current setup for expert needing to learn vocab separate from dataloader is silly and inconsistent with other modules. Should change so the expert uses the vocab...
Tags schedulers with their configuration metadata. Simplifies the `metrics` helper library somewhat to support. Note that the LightningCLI migration will make it possible for users to use basically any scheduler...
This adds GRU support; everywhere there is an LSTM model, there is now a GRU model too. I initially tried to make RNN type a general flag but because LSTMs...
Merges: https://github.com/CUNY-CL/yoyodyne/pull/233 https://github.com/CUNY-CL/yoyodyne/pull/197 Fixes: https://github.com/CUNY-CL/yoyodyne/issues/192 https://github.com/CUNY-CL/yoyodyne/issues/191 Dependent on: https://github.com/CUNY-CL/maxwell/pull/17 Summary: I fixed maxwell so TQDM isn't a property of the SED parameters anymore, this allows pickling of the expert module...
With https://github.com/CUNY-CL/yoyodyne/pull/247, most of the work in instantiating edit actions for the transducer is coming from the Index. It would make more sense to just make this inherited behavior from...
Standard procedure for LR schedulers is to step once per epoch unless otherwise specified. We have one exception to this: the warmup inverse-square-root scheduler. This is sort of important in...
[Here](yoyodyne/data/datasets.py#L79) we cast encoded tensors to the `Long` dtype, i.e., int64. Why? Surely int32 would be enough. Is there any particular reason for this? Could we save memory by getting...