Open tkornuta-ibm opened 5 years ago
Describe the bug Why two test runs on MNIST test set with different random seeds return different values
To Reproduce
Train a model: mip-online-trainer --c configs/vision/simplecnn_mnist.yaml --li
Run tester twice e.g.: mip-tester --m ./experiments/MNIST/SimpleConvNet/20181108_132032/models/model_best.pt
Compare aggregated losses will be different
Expected behavior Exactly the same numbers when the same model is used on the same samples!
Additional context Tried changing the settings, so that I will make sure that the same samples are used:
For example, batch of size 100 when SubsetRandomSampler returns a subset of 100 samples
Still resulting in different losses, e.g. loss 0.2361298800 loss 0.2361298054 loss 0.2361298203
THOSE SHOULD BE EXACTLY THE SAME!
testing:
dataloader: batch_sampler: null drop_last: true num_workers: 0 pin_memory: false shuffle: true timeout: 0 problem: batch_size: 100 name: MNIST resize:
Describe the bug Why two test runs on MNIST test set with different random seeds return different values
To Reproduce
Train a model: mip-online-trainer --c configs/vision/simplecnn_mnist.yaml --li
Run tester twice e.g.: mip-tester --m ./experiments/MNIST/SimpleConvNet/20181108_132032/models/model_best.pt
Compare aggregated losses will be different
Expected behavior Exactly the same numbers when the same model is used on the same samples!
Additional context Tried changing the settings, so that I will make sure that the same samples are used:
For example, batch of size 100 when SubsetRandomSampler returns a subset of 100 samples
Still resulting in different losses, e.g. loss 0.2361298800 loss 0.2361298054 loss 0.2361298203
THOSE SHOULD BE EXACTLY THE SAME!
testing:
seed_numpy: 4354
seed_torch: 2452
dataloader: batch_sampler: null drop_last: true num_workers: 0 pin_memory: false shuffle: true timeout: 0 problem: batch_size: 100 name: MNIST resize: