Set Random Seed¶
As described in PyTorch REPRODUCIBILITY, there are 2 factors affecting the reproducibility of an experiment, namely random number and nondeterministic algorithms.
MMEngine provides the functionality to set the random number and select a deterministic algorithm. Users can simply set the
randomness argument of the
Runner. The argument is eventually consumed in set_random_seed and it has the following three fields:
seed (int): The random seed. If this argument is not set, a random number will be used.
diff_rank_seed (bool): Whether to set different seeds for different processes by adding the
rank(process index) to the seed.
deterministic (bool): Whether to set deterministic options for the CUDNN backend.
Let’s take the Get Started in 15 Minutes as an example to demonstrate how to set
randomness in MMEngine.
runner = Runner(
optim_wrapper=dict(optimizer=dict(type=SGD, lr=0.001, momentum=0.9)),
train_cfg=dict(by_epoch=True, max_epochs=5, val_interval=1),
# adding randomness setting
However, there may still be some differences between any two experiments, even with the random number set and the deterministic algorithms chosen. The core reason is that the atomic operations in CUDA are unordered and random during parallel training.
The CUDA implementation of some operators sometimes inevitably performs atomic operations such as adding, subtracting, multiplying, and dividing the same memory address multiple times in different CUDA kernels. In particular, during the
backward process, the use of
atomicAdd is very common. These atomic operations are unordered and random when computed. Therefore, when performing atomic operations at the same memory address multiple times, let’s say adding multiple gradients at the same address, the order in which they are performed is uncertain, and even if each number is the same, the order in which the numbers are added will be different.
The randomness of the summing order leads to another problem, that is, since the summed values are generally floating point numbers that have the problem of precision loss, there will be a slight difference in the final result.
Therefore, by setting random seeds and deterministic to
True, we can make sure that the initialization weights and even the forward outputs of the model are identical for each experiment, and the loss values are also identical. However, there may be subtle differences after one back-propagation, and the final performance of the trained models will be slightly different.