I'm using tensorflow and tensorflow-hub to retrain an image classifier on my own classes following this link from tensorflow and using the retrain script. My dataset contains 348 images divided into 5 classes.
- learning rate: 0.01
- batch size: 100 (for training) and 100 for validation
- training steps: 12000
- Training set : 80%
- Validation set: 10%
- Test set: 10%
- Final test accuracy : 33.3% (according to the console logs)
in tensorboard :
final values of the accuracy graph are as the following:
- validation accuracy: 30%
- training accuracy: 95%
final values of the cross entropy graph are as the following:
- validation loss: 1,83
- training loss: 0.17
NOTE : the cross entropy for validation kept increasing throughout the training started at 1.6 and ended at 1.8
the huge difference between the validation and train accuracy seems to me unusual or is that normal ? because i trained an other image classifier with the same settings with the same model but with different dataset and the result was more reasonable. it was like 60% accuracy for training and 53% for validation.
Any advice in how i can improve the results of the new model ? or what's the meaning of the results i got. Much appreciated.
Blue is --> validation
Orange is --> training