I am using data augmentation and decaying learning rate after each epoch. If I don't use data augmentation but keeping callback, then my training accuracy reaches to 99.65% but not validation. Another scenario is, if I remove callback of learning rate decay but keeping data augmentation then training accuracy also improves and reaches 99% but not validation. Why is it stucking with current configuration (lr decay + Data Augmentation) ?
What could be the reason to have this problem with data augmentation?
Please check your theory and code dually. By removing unexpected error it will behave correctly.
Augmentation literally means extending the dataset. If your training set is reaching such high numbers and the validation set isn't that means the model is overfitting. Try building a model with lesser augmented samples or just the dataset in itself. You could also try putting in group/batchnorm layers. If nothing works then try going for another model or transfer learning.
what is your question exactly?
the question and the description do not really match
are you asking why your validation accuracy does not improve above 0.14?
Generally speaking I do not see a problem with that
Have you tried any other model in order to check if you have any predictive power in the covariates at all?
Stefano Nembrini Thank you so much for your kind response. The problem is, my both accuracies stop improving when I apply data augmentation.. I am using a pre-trained model.
If the model was already trained, how is a measure computed on the training set, it's confusing
Also, how is your data augmented exactly?
What are the results when you do not augment it?
What are the out of sample values obtained in the two scenarios(with and without data augmentation )?
Stefano Nembrini I am transferring weights of pre-trained models.
The validation and testing data both are not augmented. The test samples are 10K and evenly distributed between all 10 classes. The validation samples are 6000 random samples that I am getting through train_test_split function of sklearn.metrics.
I don't think you are really clear in your explanation, so we can help so much,
If the model was trained (also you do not say what algorithm or software you are using)
You should only have a measure computed on data not include in the training or validation step.
This being said, what are those values and why do you need them to keep improving? Such measures just plateau at some point but again if it's just a test set, you should have a single value and not one for every epoch
What really matters after the model was trained is the performance on a test set, training and validation data are used to train and select parameters in the algorithm
Stefano Nembrini I got what you said. I am using keras and tensorflow. I trained a pre-trained model from keras applications. I have enough data that's why I trained the model by removing fully connected layers and adding my own FC layer and output layer.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training