Open surajitkundu29 opened 4 years ago
No. There isn't implemented the training of one model on several servers.
without synchronized-batch-normalization it will hurt accuracy - training on more than 4 x GPUs
with synchronized-batch-normalization it will hurt training speed - training on several servers (but you will be able to use large mini_batch_size that will increase accuracy) - instead for this I suggest to use GPU-processing on CPU-RAM: https://github.com/AlexeyAB/darknet/issues/4386
For Ex. I have two GPU enabled server and both having 4 GPUs, How my training can be distributed over these two server ?