Training and Deployment Workflow in NeuralLib

This section describes how to train, fine-tune, post-process, and deploy models using NeuralLib.

---

1. Training and Retraining Models

Available Architectures

To see a list of available architectures, use:

from NeuralLib.architectures import get_valid_architectures
print(get_valid_architectures())

Make sure to validate architecture names before training:

from NeuralLib.architectures import validate_architecture_name
validate_architecture_name("GRUseq2seq")

---

Training a Model from Scratch

To train a model from scratch, instantiate it and call the train_from_scratch method:

from NeuralLib.architectures import GRUseq2seq

model = GRUseq2seq(model_name="test", n_features=10, hid_dim=64, n_layers=2, dropout=0.1, learning_rate=0.001)

model.train_from_scratch(
    path_x="data/train_x.npy",
    path_y="data/train_y.npy",
    batch_size=32,
    epochs=50
)

Alternatively, you can use a higher-level function to automate training:

from NeuralLib.training import train_architecture_from_scratch

train_architecture_from_scratch(
    architecture_name="GRUseq2seq",
    architecture_params={"n_features": 10, "hid_dim": 64, "n_layers": 2, "dropout": 0.1, "learning_rate": 0.001},
    train_params={"path_x": "data/train_x.npy", "path_y": "data/train_y.npy", "batch_size": 32, "epochs": 50}
)

---

Retraining an Existing Model

If you have a checkpoint from a previous training session, you can continue training:

from NeuralLib.training import retrain_architecture

retrain_architecture(
    architecture_name="GRUseq2seq",
    train_params={"batch_size": 32, "epochs": 20},
    checkpoints_directory="checkpoints/GRUseq2seq_run1"
)

You can also load and retrain models from Hugging Face:

retrain_architecture(
    architecture_name="GRUseq2seq",
    train_params={"batch_size": 32, "epochs": 20},
    hugging_face_model="marianaagdias/GRU_ecg_model"
)

---

Hyperparameter Optimization

Grid search can be used to find the best hyperparameters automatically:

from NeuralLib.training import run_grid_search

best_model = run_grid_search(
    architecture_name="GRUseq2seq",
    architecture_params_options={
        "hid_dim": [32, 64, 128],
        "dropout": [0.1, 0.2, 0.3],
        "learning_rate": [0.001, 0.0005]
    },
    train_params={"batch_size": 32, "epochs": 50}
)

💡 This function iterates over all parameter combinations and selects the best performing model.

---

2. Post-Processing Model Outputs

Once a model has been trained, you may need to post-process its output. This is particularly useful for peak detection or classification thresholds.

Binary Peak Detection

from NeuralLib.processing import post_process_peaks_binary

predictions = model.predict(input_signal)

filtered_peaks = post_process_peaks_binary(
    predictions,
    threshold=0.5,
    filter_peaks=True
)

This function: ✔ Applies a sigmoid activation

✔ Thresholds the output to detect peaks

✔ (Optional) Filters out closely spaced peaks

---

3. Deploying Models to Hugging Face

Once a model is trained, you can upload it to Hugging Face to make it publicly available.

Generating a Model Card

To ensure proper documentation, you must create a structured README.md:

from NeuralLib.upload import create_readme

create_readme(
    hparams_file="model/hparams.yaml",
    training_info_file="model/training_info.json",
    output_file="model/README.md",
    collection="NeuralLib Collection",
    description="A deep learning model for ECG signal classification.",
    model_name="GRUseq2seq"
)

This automatically formats:

• Hyperparameters
• Training results
• Example usage

---

Uploading the Model

To push the model to Hugging Face Model Hub:

from NeuralLib.upload import upload_production_model

upload_production_model(
    local_dir="models/my_model",
    repo_name="marianaagdias/my_ecg_model",
    token="your-huggingface-token",
    model_name="GRUseq2seq",
    description="This model detects arrhythmias in ECG signals."
)

✔ If the repository doesn’t exist, it will be created automatically.

✔ The model is now publicly available and can be used by others.

---

Final Notes

• Training (train_architecture_from_scratch, retrain_architecture, run_grid_search)
• Post-Processing (post_process_peaks_binary)
• Deployment (create_readme, upload_production_model)

This section ensures that developers and researchers can: ✔ Train and fine-tune models

✔ Apply post-processing

✔ Deploy models to Hugging Face