Prevent PyTorch's `CUDA error: out of memory` in just 1 line of code.
Gradient accumulation for Keras