Feature Request: Add Mish activation

[digantamisra98]

Mish is a novel activation function proposed in this paper.
It has shown promising results so far and has been adopted in several packages including:

TensorFlow-Addons	SpaCy (Tok2Vec Layer)	Thinc - SpaCy's official NLP based ML library
Eclipse's deeplearning4j	Hasktorch	Echo AI
CNTKX - Extension of Microsoft's CNTK	FastAI-Dev	Darknet
Yolov3	BeeDNN - Library in C++	Gen-EfficientNet-PyTorch
dnet	ruby-dnn	blackcat-tensors
DL4S	HuggingFace Transformers	PAGI
OpenCV	Odin-AI	Mini DNN
Efficient Segmentation Networks	TF Semantic Segmentation	Dynastes
DLib	Copernicus	AllenNLP
PyWick

All benchmarks, analysis and links to official package implementations can be found in this repository

Mish also was recently used for a submission on the Stanford DAWN Cifar-10 Training Time Benchmark where it obtained 94% accuracy in just 10.7 seconds which is the current best score on 4 GPU and second fastest overall. Additionally, Mish has shown to improve convergence rate by requiring less epochs. Reference -

Mish also has shown consistent improved ImageNet scores and is more robust. Reference -

Additional ImageNet benchmarks along with Network architectures and weights are avilable on my repository.

Summary of Vision related results:

It would be nice to have Mish as an option within the activation function group.

This is the comparison of Mish with other conventional activation functions in a SEResNet-50 for CIFAR-10: