zchen/b2txt25
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

readme file linking

Browse Source
This commit is contained in:
nckcard
2025-07-02 15:34:24 -07:00
parent 2e7389030b
commit b423ba869f
3 changed files with 11 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
model_training/README.md
View File

@@ -12,7 +12,7 @@ All model training and evaluation code was tested on a computer running Ubuntu 2
## Training
### Baseline RNN Model
We have included a custom PyTorch implementation of the RNN model used in the paper (the paper used a TensorFlow implementation). This implementation aims to replicate or improve upon the original model's performance while leveraging PyTorch's features, resulting in a more efficient training process with a slight increase in decoding accuracy. This model includes day-specific input layers (512x512 linear input layers with softsign activation), a 5-layer GRU with 768 hidden units per layer, and a linear output layer. The model is trained to predict phonemes from neural data using CTC loss and the AdamW optimizer. Data is augmented with noise and temporal jitter to improve robustness. All model hyperparameters are specified in the `rnn_args.yaml` file.
We have included a custom PyTorch implementation of the RNN model used in the paper (the paper used a TensorFlow implementation). This implementation aims to replicate or improve upon the original model's performance while leveraging PyTorch's features, resulting in a more efficient training process with a slight increase in decoding accuracy. This model includes day-specific input layers (512x512 linear input layers with softsign activation), a 5-layer GRU with 768 hidden units per layer, and a linear output layer. The model is trained to predict phonemes from neural data using CTC loss and the AdamW optimizer. Data is augmented with noise and temporal jitter to improve robustness. All model hyperparameters are specified in the [`rnn_args.yaml`](rnn_args.yaml) file.
### Model training script
To train the baseline RNN model, use the `b2txt25` conda environment to run the `train_model.py` script from the `model_training` directory:
@@ -20,7 +20,7 @@ To train the baseline RNN model, use the `b2txt25` conda environment to run the
conda activate b2txt25
python train_model.py
```
The model will train for 120,000 mini-batches (~3.5 hours on an RTX 4090) and should achieve an aggregate phoneme error rate of 10.1% on the validation partition. We note that the number of training batches and specific model hyperparameters may not be optimal here, and this baseline model is only meant to serve as an example. See `rnn_args.yaml` for a list of all hyperparameters.
The model will train for 120,000 mini-batches (~3.5 hours on an RTX 4090) and should achieve an aggregate phoneme error rate of 10.1% on the validation partition. We note that the number of training batches and specific model hyperparameters may not be optimal here, and this baseline model is only meant to serve as an example. See [`rnn_args.yaml`](rnn_args.yaml) for a list of all hyperparameters.
## Evaluation
### Start redis server