agenlus-hub 0.1.7

Python client for Agenlus model upload pipeline.

Agenlus Python Hub Client (agenlus-hub)

agenlus-hub is the official Python client library for the Agenlus RL Platform. It provides a seamless interface to download custom RL environments, run local evaluations on your PyTorch models, and automatically export, build, upload, and register them onto the Agenlus leaderboards.

---

Key Features

• Simple Authentication: Quick setup with single-line token authentication.
• Environment Downloader: Easily fetch custom Gymnasium environment specifications and source code directly from the server.
• Automatic Export: Exposes models to both PyTorch (.pt) and ONNX (.onnx) formats with dynamic batching.
• Deterministic Evaluation: Validates model performance locally using reproducible seeds before uploading.
• Hugging Face Hub Integration: Automates repositories creation and stacks model uploads into tidy subfolders.
• Leaderboard Registration: Instant, automated registration of evaluated models to the Agenlus leaderboard.

---

Installation

Install the package and its dependencies via pip:

pip install agenlus-hub

Dependencies

Ensure you have the following prerequisites installed (or they will be installed automatically):

• requests
• torch
• onnx
• onnxscript
• huggingface_hub
• gymnasium (Required for local model evaluation)

---

Quick Start

Here is a quick walkthrough showing how to log in, download an environment, build/train your model, and upload it to the leaderboard:

import torch
import torch.nn as nn
import agenlus

# 1. Log in to your Agenlus account
agenlus.login(token="YOUR_AGENLUS_API_TOKEN")

# 2. Download the target environment (e.g., CartPole-v1)
# This downloads 'CartPole-v1.py' into your working directory
env_file_path = agenlus.download("system/CartPole-v1")

# 3. Define and train your PyTorch RL Model
class PolicyNet(nn.Module):
    def __init__(self, obs_dim=4, action_dim=2):
        super().__init__()
        self.fc = nn.Sequential(
            nn.Linear(obs_dim, 128),
            nn.ReLU(),
            nn.Linear(128, action_dim)
        )
        
    def forward(self, x):
        return self.fc(x)

# Instantiate your model (or load from disk)
model = PolicyNet()

# 4. Evaluate and Upload to Agenlus & Hugging Face
# This evaluates the model for 100 episodes, exports to ONNX/PyTorch, 
# uploads to Hugging Face, and registers the scores on the Leaderboard.
agenlus.upload(
    model=model,
    env_id="system/CartPole-v1",
    hf_token="YOUR_HUGGINGFACE_WRITE_TOKEN",
    model_name="my-first-cartpole-agent",
    episodes=100,
    seed=42
)

---

API Reference

agenlus.login

Authenticates your local client session.

agenlus.login(token: str, api_url: str = None)

• token (str): Your Agenlus developer API token.
• api_url (str, optional): Overrides the default platform backend URL.

---

agenlus.download

Downloads the source code of an environment registered on the platform.

save_path = agenlus.download(env_id: str, save_path: str = None)

• env_id (str): The ID of the environment (e.g., system/CartPole-v1 or username/custom-env).
• save_path (str, optional): Custom file path to save the code. Defaults to {env_name}.py in the current directory.
• Returns: The path to the saved file.

---

agenlus.upload

Executes local evaluation, model export, Hugging Face upload, and platform leaderboard registration.

agenlus.upload(
    model: torch.nn.Module,
    env_id: str,
    hf_token: str,
    hf_repo: str = None,
    model_name: str = None,
    episodes: int = 100,
    seed: int = None
)

• model (torch.nn.Module): The trained PyTorch model.
• env_id (str): Target environment ID on the platform.
• hf_token (str): Write access token for Hugging Face.
• hf_repo (str, optional): The Hugging Face repo (e.g., username/repo-name). If omitted, checks your profile settings or defaults to username/agenlus-agents.
• model_name (str, optional): Unique name for this model checkpoint. Subdirectories under the repository will be structured around this name.
• episodes (int, optional): Number of episodes to run for scoring evaluation. (Default: 100).
• seed (int, optional): Seed used to ensure deterministic environment evaluation. (Default: 42).

---

License

This project is licensed under the MIT License - see the LICENSE file for details.