search

HealthML / self-supervised-3d-tasks

Apache License 2.0
185 stars 39 forks source link

The GPU Memory-Usage is not changed #10

Closed szhang963 closed 3 years ago

szhang963 commented 3 years ago

Hi! Thanks for your code!
I run your code successfully. However, I found the GPU Memory-Usage is very low (138M), though I adjusted the batch size in the "*.json" file. This caused the very low run speed. Could you give me some help？ Thanks in advance.

aihamtaleb commented 3 years ago

Hi, Thanks for your interest. Are you sure you are using the right versions of Cuda and Tensorflow?

szhang963 commented 3 years ago

Hi, My environment is: CUDA 10.1 tensorflow-gpu 2.0.0

szhang963 commented 3 years ago

Hi, Thanks for your mention. I have solved this problem by running this command conda install tensorflow-gpu=2.1. However, I found the pre-train output logs in pancreas dataset include 'accuracy' and 'val_loss', 'val_accuracy' value. These metrics were evaluated for the accuracy in the upstream task?

aihamtaleb commented 3 years ago

Correct, when pretraining, these metrics are computed only on the upstream task. You can specify the dataset split for validation on the upstream task.

szhang963 commented 3 years ago

Thanks for your answer.