GPU training

🦉 AthenaAI is an AI-powered ensemble machine-learning trading bot for PocketOption, built on BinaryOptionsToolsV2 → Get it on GitLab
💬 Need help? Join the Chipa Discord · ⚡ Go no-code with ChipaEditor

PyTorch is optional in AthenaAI, and the GPU is used by exactly one component: the neural (MLP) model trained by train.py. Everything else — feature/sample generation, the sklearn online and batch models, and the live bot’s prediction loop — runs on CPU regardless.

Device detection

detect_device() in src/training/device.py returns (device, description) with this priority:

--cpu passed → ("cpu", "CPU (forced via --cpu)")
PyTorch not importable → ("cpu", "CPU (PyTorch not installed — pip install torch for GPU training)")
torch.cuda.is_available() → ("cuda", "GPU: <name> (<VRAM> GB VRAM)")
torch.backends.mps.is_available() → ("mps", "GPU: Apple Metal (MPS)")
Otherwise → ("cpu", "CPU (PyTorch installed but no GPU available)")

The trainer logs the result at startup: ⚡ Training device: GPU: ... (or 🖥 Training device: CPU ...).

Checking your device

python train.py --check-gpu

Prints the detected device and exits without training:

Device: cuda  (GPU: NVIDIA GeForce RTX 4070 (12.0 GB VRAM))

If this says CPU but you have an NVIDIA GPU, you almost certainly installed the CPU-only PyTorch wheel — reinstall from the CUDA index below.

Installing PyTorch per platform

NVIDIA (Windows/Linux): use a CUDA build from https://pytorch.org/get-started/locally/, e.g.:
```
pip install torch --index-url https://download.pytorch.org/whl/cu128
```
Apple Silicon (macOS): plain pip install torch — MPS is detected automatically.
CPU-only (any OS): pip install torch. The MLP still trains, just slower; or skip it entirely with --no-nn.
No PyTorch: the trainer logs PyTorch not installed — skipping neural model (pip install torch for GPU training). and trains the sklearn ensemble only. This is informational, not an error.

What actually runs on the GPU

Only TorchMLPClassifier (src/training/torch_model.py), trained in Trainer._train_nn():

Inputs are scaled with the ensemble’s already-fitted StandardScaler, then the MLP trains for --epochs (default 30) with mini-batches of --batch-size (default 2048) on the detected device. Log: 🧠 Training neural model on CUDA (30 epochs) ….
Its holdout accuracy (MLP holdout accuracy: X%) becomes its vote weight inside the ensemble, so a weak net gets proportionally less say.
It is stored in the brain’s batch-model list alongside GBM and RF (any previous MLP entry is replaced first).

Sample generation parallelizes across CPU cores (up to 16 processes) and GBM/RF are CPU-bound — a GPU does not speed those stages up.

CUDA out-of-memory? Lower --batch-size (e.g. 512) or shrink --hidden.

Brain portability: GPU → CPU

A brain trained on a GPU machine is fully portable:

The MLP’s weights are stored as a CPU state dict inside the brain pickle, so loading never requires CUDA. A CPU-only machine loads the same brain and predicts with the MLP on CPU.
If PyTorch is missing entirely on the loading machine, the MLP cannot run — it degrades to a neutral 0.5 probability vote instead of breaking the brain. The rest of the ensemble (SGD/PA/NB/GBM/RF) works normally, so predictions continue, just without the neural model’s contribution.

Practical workflow: train on a GPU box (python train.py --out athena_brain.pkl), copy the .pkl to the trading machine, run python main.py there. See Persistence for what else to copy.