--batch-size

🦉 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

Mini-batch size used when training the neural model.

What it does

_train_nn() passes nn_batch_size to TorchMLPClassifier(batch_size=...). In fit(), the training data becomes a TensorDataset wrapped in a DataLoader(ds, batch_size=self.batch_size, shuffle=True); each optimizer step processes one batch, moved to the training device (cuda/mps/cpu). Shuffling happens within the training set only — the chronological train/holdout split from --test-split already happened, so no future data leaks in.

Batch size affects only the neural model; the sklearn online models use their own fixed chunk size (TrainerConfig.online_batch = 5000, not CLI-exposed).

Choosing a value

• Larger batches mean fewer optimizer steps per epoch (faster on GPU) but more GPU memory per step; the loss-visible effect of fewer, less noisy gradient updates can also change what learning rate works best.
• CUDA out of memory: lower this first — the training guide suggests e.g. --batch-size 512 — or shrink --hidden. The feature vectors are small, so at the default 128,64 network even 2048 is light on modern GPUs; OOM usually indicates a very large --hidden.
• On CPU, very large batches gain little; values in the 256–2048 range are all reasonable.

Examples

python train.py --batch-size 512          # if you hit GPU out-of-memory
python train.py --batch-size 4096 --lr 2e-3

Interactions

• --hidden — the other memory-footprint knob.
• --lr — larger batches often tolerate/need a higher learning rate.
• --no-nn — makes this flag a no-op.
• --cpu — removes GPU memory limits from the equation.

See also

• GPU guide
• src.training.torch-model API reference

💬 Unsure how this interacts with the rest of the configuration? Ask in the Chipa Discord — or prototype strategies no-code with ChipaEditor.