Reproduce results with different MXNET versions?



I’m trying to reproduce results from a model trained with mxnet 0.10. I’m using the exact same hyperparameters and training/validation data, but retrain with mxnet 1.2.1. Now I get different results and can’t reproduce the exact same model.


1.Shoud I expect models might be different when trained with different mxnet versions?
2. If yes, should I expect models trained with newer mxnet version would have the same or better quality in general (with the same hyperparams)? Otherwise, I might need to have “mxnet version” as a hyperparam to tune.



With each new version we fix a certain number of bugs. So you can expect the quality to increase overall, though with new features it is possible that some regression got in as well, despite the unit and integration tests.

However with deep learning, bugs are sometimes features and the network learns around these bugs. What that means in practice is that you might need to train your network with different hyper-parameters to reach the same accuracy after a given bug is fixed.

The other thing to take into account is the initialization of your network. Deep neural networks are extremely sensitive to initialization and you can run two trainings with different initialization and get completely different final accuracies. Some family of networks like GANs are especially sensitive to that.

Also, what do you mean by “I can’t reproduce the exact same model” ? If you mean accuracy, see comments above. If you mean the exact same model, as exact same final weights, because of the stochastic nature of the training, it is very unlikely that you would end up with two exact version of the network after two runs.


Thanks for the feedback.

We set the same random seed for python, numpy and mxnet so theoretically each training results should be deterministic? I did try retrain the model a few times with the exact same settings (hyperparams, data, random seed, # of epochs, etc), and each time I can reproduce the same results if I use the exact same version of software. That’s why we thought the reason that we couldn’t reproduce the exact same model (network weights) was due to mxnet version difference, because all the other settings was exactly the same.


Yes, operators, optimizers, are updated to fix existing bugs, numerical instability etc. You can expect the final values of your weights to change across versions, especially major ones.