Dense Regression RNN (Keras)

Train a fully connected, dense-only neural network in Keras/TensorFlow to predict thickness from 109-length numeric sequences stored in training_data.txt. We log metrics to Weights & Biases (wandb) and save a Keras model artifact.

Setup

• Ensure Python 3.10+ is available.
• Create & activate a virtual environment (PowerShell):

  python -m venv .venv
  .\.venv\Scripts\Activate.ps1
  python -m pip install --upgrade pip
  pip install -r requirements.txt

• W&B: either set WANDB_API_KEY env var or run wandb login. If you are offline, use --wandb-mode offline when running the script.

Data

• Training file: training_data.txt (CSV with columns thickness and values, where values is a stringified list of 109 numbers).
• The script automatically parses and standardizes features; no manual preprocessing needed.

Train

Edit config.yaml to set all defaults (data path, epochs, batch size, dense layer widths, wandb mode, etc.). CLI flags override anything in the config.

python train.py --config config.yaml

Architecture selection

• Set model_type: dense (default) to use the fully-connected stack defined by dense_sizes.
• Set model_type: cnn1d to use a 1D CNN; configure cnn1d_filters, cnn1d_kernel_sizes, and cnn1d_pool_size in config.yaml.
• Learning-rate reduction: control with lr_reduce_on_plateau, lr_patience, lr_factor, lr_min in config.yaml (ReduceLROnPlateau on val_loss).

CLI overrides (examples):

• --epochs 10 : override epochs from config.
• --dense-sizes 256 128 64 32 : create 4 hidden Dense layers with those widths.
• --model-type cnn1d --cnn1d-filters 64 128 256 --cnn1d-kernel-sizes 7 5 3 --cnn1d-pool-size 2 : run a 3-layer 1D CNN with custom kernels and pooling.
• --wandb-mode offline : set wandb mode for this run.

Evaluate

Validation metrics are reported at the end of training and logged to wandb (if enabled). To run an evaluation-only pass on the saved model:

python train.py --data-path training_data.txt --epochs 0 --load-existing artifacts/model.keras

This will load the saved model, compute metrics on the held-out validation split, and exit.

Notes

• Loss: Mean Squared Error; optimizer: Adam; activations: ReLU on hidden layers, linear on the output for regression.
• The number of hidden Dense layers is determined by dense_sizes in config.yaml (or CLI); provide any length list to build that many layers.
• For cnn1d, the feature vector is reshaped to (sequence_length, 1), then a stack of Conv1D + optional MaxPool layers (from cnn1d_filters/cnn1d_kernel_sizes/cnn1d_pool_size) feeds into a Dense head for regression.
• ReduceLROnPlateau is enabled by default; tune or disable it via config.
• Deterministic seeds are set for reproducibility; disable GPU nondeterminism separately if needed.