BAdam

Modification of Adam [1] optimizer with increased stability and better performance. Tricks used:

1. Decoupled weight decay as in AdamW [2]. Such decoupling allows easier tuning of weight decay and learning rate. This implementation follows PyTorch and multiplies weight decay by learning rate, allowing simultaneous scheduling. Due to this typical values passed to optimizer should be much higher that with SGD, if you use 1e-4 with SGD, good start would be to use 1e-2 with BAdam.
2. Epsilon is inside sqrt to avoid NaN in mixed precision. Default value is much larger than in Adam to reduce 'adaptivity' it leads to better and wider optimums [3]. Large epsilon also works better than amsgrad version of Adam [5]
3. exp_avg_sq inits with large value, rather than with zeros. This removes the need for lr warmup and does the same thing as all the tricks from RAdam [4], while being much simpler.
4. Removed bias correction. It's not needed if exp_avg_sq is correctly initialized

Practical Tips

Default values for this optimizer were tuned on Imagenet and work as good baseline for other computer vision tasks. Try them as is, before further tuning.

Installation

pip install git+https://github.com/bonlime/[email protected]

Reference:

[1] Adam: A Method for Stochastic Optimization
[2] Decoupled Weight Decay Regularization
[3] On the Convergence of Adam and Beyond
[4] On the Variance of the Adaptive Learning Rate and Beyond [5] Adaptive Methods for Non-convex Optimization