digantamisra98/Mish

Defined by f(x) = x·tanh(softplus(x)), Mish introduces self-regularization through a smooth, non-monotonic curve that maintains gradient flow across negative and positive domains. Integrated across major frameworks—PyTorch, TensorFlow, MXNet, and OpenVINO—with optimized CUDA variants and memory-efficient implementations enabling deployment from edge to large-scale models. Demonstrates state-of-the-art performance in computer vision tasks including object detection on MS-COCO, with variants like H-Mish enabling architectural flexibility.

1,303 stars. Actively maintained with 9 commits in the last 30 days.