If we fixed a seed for random generation we should be able to get the same results.
Every time you run a deep learning optimization there is randomness involve almost everywhere. For example, we use randomness in the cropping of images, in the initialization of weights etc… I think 0.993 is a very good result. You can run the
@jeremy something happened from last week to this week.
Before the learner was giving feedback when it was downloading the model from fast.ai.
Now it takes a lot of time to start the fit process??
Looks like the whole thing is taking all the memory on the system (16 GB)
– You can use cross-validation, one of the notebooks have an example on how to use it.
– You can add your own metrics. Look for the argument “metrics”.
– As Jeremy mentioned the best approach is to oversample the minority class. You can also under-sample the majority class.
@yinterian which notebook shows an example of how to use cross-validation? Jeremy had mentioned that it wasn’t needed unless the dataset was so small that we couldn’t afford to set aside the validation set. Is there a good benchmark/threshold for deciding when its “too small”? Or is it simply a matter of checking whether or not k-fold cv improves accuracy?
I think it was also mentioned that at the end of validation, you can actually run training one last time with the whole dataset (validation included) to get a better accuracy which again I guess would mean cv isn’t needed in that case.
When center cropping an image (see red boarder), we may loss the important details (head and paws in this case). Should we use image re-sizing from rectangular to square instead?
Excellent question. This type of resizing is what keras does by default. I’ve found that it seems to generally work less well, since it has to learn how images look different depending on how they’re squeezed. But we do have the ability to use this squeezing approach in fastai - maybe @yinterian could show an example?
This is going to depend on the problem. For many problems center cropping will be fine for other problems you may want to resize. You can do both with the fast.ai library.
Curious if anyone has tried ResNet style architecture for language models. Would be interesting to see what kind of features initial layers would capture and how starting with smaller sequences and retraining on larger ones (like how we did in class with starting with smaller size images and changing to larger ones) would affect the model.