I have trained 10 epochs of a image dataset. But the the accuracy keeps fluctuating. It doesn’t have a trend. Should I reduce my epochs?
Here is just what I think:
I think you are pretty lucky enough to get a very good result with a learning rate.
The best learning rate should be like this.
That is the first 20 percents goes to the top and then quickly falls down. I guess you used 1e-3. If you want, I hope you can share your lr plot.
This is my lr_curve
And the learning rate that you used? 1e-3? if so, I think you are doing right.
Should I reduce no of epochs?
It is totally normal. As @JonathanSum pointed. Your loss first increase quickly to get an idea of learning space and then decreases slowly. It leads to better optimisation. I see no issues with your epochs. If you see a big jump in loss at start, then you should try reducing the learning rate a bit.
I think vishalydv23 is more correct than me and I am wrong in something. If your loss first increase quickly, you should decrease it slowly. I suggest to use magic number 1e-3 instead of suggestion 2e-3.
Ok. Will try.
Thanks both of u 
Two thoughts came to my mind:
- How large is your validation set? Metrics tend to fluctuate more heavily for small validation sets.
- Looking at the print out of your model training, I would even try to increase the number of epochs and see what happens. In epoch 10, your validation loss and accuracy were still improving, so it can be that you still haven’t reached the highest accuracy.