I am having problem with prediction with my model trained on resnet50. I have 10 classes of Nepali numbers from (0 ...9). I have trained the model for 100 epochs with around 40,000 data . What is the issue with my model? I am having overfit? Or, the model I have used to train is just too complex for 10 class. I have also tried this model to predict on the training set but the prediction accuracy is very very poor around (10%).
Hi,
Such result is exextended with this huge number of classes. Here are some suggestions:
- Reduce the class number.
- Clean your data properly.
- Try different number of hidden layers, units, epochs, etc.
- Use cross-validation to evaluate the learning process you performed.
NB: Don't use the training set for testing purposes.
HTH.
Dr. Samer Sarsam
Saugat,
I would suggest that you plot the error curve as it is being processed (while it is not a measure for publication, it is a good troubleshooting tool). you should see a steady pattern of reduction. if you see a steady line, most likely it is how you are feeding the data. if you see oscillations it is more likely some parameter that needs adjustment.
Dear Saugat Bhattarai ,
I think you are using few samples to learning a ResNet network. In particular, this network requires a lot of samples during the training, due to its high capacity. I had similar problems in CIFAR-10 using ResNet 50, and when I used data argumentation, the network works fine. Finally, I recommend you to decrease the learning rate, iteratively, after some epochs, for instance, after 80, 120, 160, 180 epochs.
- Badges
- Science topic