TensorFlow 1.x compatibility
For TF 1, I suggest using the AccumOptimizer implementation in the H2G-Net repository.
This wrapper works similarly to our GradientAccumulateOptimizer wrapper.
An equivalent GradientAccumulateModel wrapper does not exist in TF 1.x as overloading of the train_step was a new feature introduced in tensorflow==2.2.
Hence, also note that for tensorflow<2.2>=2.0 only the GradientAccumulateOptimizer is compatible.