Why does ResNet50 Accuracy vary from run to run?

Hi everybody,

in the notebook „Residual_Networks“ we evaluate our ResNet50 model after training it for 10 epochs.
After running it for multiple time I saw that the Test Accuracy varies greatly from run to run. Out of curiosity I added a step to calculate the Train Accuracy. These are the results of five runs:

  1. Test Accuracy = 0.8416666388511658 - Train Accuracy = 0.9657407402992249
  2. Test Accuracy = 0.9166666865348816 - Train Accuracy = 0.9592592716217041
  3. Test Accuracy = 0.7250000238418579 - Train Accuracy = 0.7851851582527161
  4. Test Accuracy = 0.8416666388511658 - Train Accuracy = 0.9509259462356567
  5. Test Accuracy = 0.9583333134651184 - Train Accuracy = 0.979629635810852

Can you help me to understand why the Accuracy varies so much from run to run? As the seed is fixed I did not expect this.

Please click my name and message your notebook as an attachment.