README.md

---
language:
  - en
license: apache-2.0
library_name: generative-pk
datasets:
  - simulated
metrics:
  - rmse
  - npde
tags:
  - generative
  - predictive
---

# Hierarchical Neural Process for Pharmacokinetic Data

## Overview
An Amortized Context Neural Process Generative model for Pharmacokinetic Modelling

**Model details:**
- **Authors:** César Ojeda (@cesarali)
- **License:** Apache 2.0

## Intended use  
Sample Drug Concentration Behavior and Sample and Prediction of New Points or new Individual
## Runtime Bundle

This repository is the consumer-facing runtime bundle for this PK model.

- Runtime repo: `cesarali/AICME-runtime`
- Native training/artifact repo: `cesarali/AICMEPK_cluster`
- Supported tasks: `generate`, `predict`
- Default task: `generate`
- Load path: `AutoModel.from_pretrained(..., trust_remote_code=True)`

### Installation

You do **not** need to install `sim_priors_pk` to use this runtime bundle.

`transformers` is the public loading entrypoint, but `transformers` alone is
not sufficient because this is a PyTorch model with custom runtime code. A
reliable consumer environment is:

```bash
pip install torch transformers huggingface_hub lightning datasets pandas torchtyping gpytorch pot torchdiffeq torchsde ruamel.yaml pyyaml
```

### Python Usage

```python
from transformers import AutoModel

model = AutoModel.from_pretrained("cesarali/AICME-runtime", trust_remote_code=True)

studies = [
    {
        "context": [
            {
                "name_id": "ctx_0",
                "observations": [0.2, 0.5, 0.3],
                "observation_times": [0.5, 1.0, 2.0],
                "dosing": [1.0],
                "dosing_type": ["oral"],
                "dosing_times": [0.0],
                "dosing_name": ["oral"],
            }
        ],
        "target": [],
        "meta_data": {"study_name": "demo", "substance_name": "drug_x"},
    }
]

outputs = model.run_task(
    task="generate",
    studies=studies,
    num_samples=4,
)
print(outputs["results"][0]["samples"])
```

### Predictive Sampling

```python
from transformers import AutoModel

model = AutoModel.from_pretrained("cesarali/AICME-runtime", trust_remote_code=True)

predict_studies = [
    {
        "context": [
            {
                "name_id": "ctx_0",
                "observations": [0.2, 0.5, 0.3],
                "observation_times": [0.5, 1.0, 2.0],
                "dosing": [1.0],
                "dosing_type": ["oral"],
                "dosing_times": [0.0],
                "dosing_name": ["oral"],
            }
        ],
        "target": [
            {
                "name_id": "tgt_0",
                "observations": [0.25, 0.31],
                "observation_times": [0.5, 1.0],
                "remaining": [0.0, 0.0, 0.0],
                "remaining_times": [2.0, 4.0, 8.0],
                "dosing": [1.0],
                "dosing_type": ["oral"],
                "dosing_times": [0.0],
                "dosing_name": ["oral"],
            }
        ],
        "meta_data": {"study_name": "demo", "substance_name": "drug_x"},
    }
]

outputs = model.run_task(
    task="predict",
    studies=predict_studies,
    num_samples=4,
)
print(outputs["results"][0]["samples"][0]["target"][0]["prediction_samples"])
```

### Notes

- `trust_remote_code=True` is required because this model uses custom Hugging Face Hub runtime code.
- The consumer API is `transformers` + `run_task(...)`; the consumer does not need a local clone of this repository.
- This runtime bundle is intentionally separate from the native training export so you can evaluate both distribution paths in parallel.