While training a NN within tensorflow (similar to the iris classification example) I used the same dataset to train and text. I assumed the accuracy would be the same. Training accuracy was 95% testing 85%. Any ideas?
Yes! I think you are using different accuracy metrics (categorical and binary). Tensorflow uses categorical accuracy as default evaluation metric and, probability, you are using binary accuracy in your second evaluation (out of Tensorflow).
Yes, they should be the same if you are using a replica of the training data provided all the training samples are used in both instances (training and test). If you use lesser number of samples in the testing stage they you will expect the accuracy to be different. Hope this helps.
Thanks. I am using the identical dataset. I do get about a 1% difference. It is possible that the ordering of the data is making a slight difference, however it seems as though the test dataset accuracy is always a little less. Bit strange as I would have thought the test and training would be identical.
hi,
That because:
1- the weights normalization is random.
If you fix weights you will have the same result.
2- Overfeeting
Best regards.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training