Do we have to reinstall everytime it changes version? I assume git pull just gets the code not the update of fastai
Might be different on tabular data, but for CNNs nine times out of ten validation error still goes below training error by the end
Name: fastai Version: 1.0.18 Summary: fastai makes deep learning with PyTorch faster, more accurate, and easier
can learn.recorder.plot_losses() be explained in a bit more detail? why train_loss is plotted for each iteration whereas val_loss is plotted after each epoch?
Thanks! I wonder if there is anyone who looks at this rigorously…
Git pull is just for the nbs/doc/materials. The libraries should be updated with pip or conda.
Validation loss is calculated for the entire validation set at the end of every epoch.
How do we know that our images were not used to train resnet. does it matter ?
When I train a learner for 1 epoch and then repeat that step (e.g. train for another epoch in another fit_one_cycle
call), is that equivalent to training for 2 epochs right away?
I think your point about 3e-4 being “halfway” between e-5 and e-4 on a log scale is right. I’ve seen this used elsewhere.
What models can be used (proven to work) for large scale Image classification (ResNext,InceptionResNet,WideResNet) except ResNet (more number of parameters) if our dataset exceeds 10M images?
What do you do if there are unbalanced classes in your dataset? Such as 200 teddy bears and 50 grizzlies.
how does ’ Using the FileDeleter
widget from fastai.widgets
we can prune our top losses, removing photos that don’t belong.’ this prune our dataset knowing what is correct or not?
Suppose that I am building a classifier to detect oranges and apples in an image. Suppose I have 100 images for each of the class. Also I want to predict onjecte which doesn’t have apple or oranges in it and I want to classify it as ‘other’. So does this requires me to get images which belong to other class like mango and other fruits and then train the model. How can I achieve this ?
yes, it is the same.
Most of the time you want to oversample the smaller class to balance it out in training, at least I think that’s what the current best recommendation is
for ‘from fastai.widgets import *’ I get ModuleNotFoundError: No module named ‘fastai.widgets’
It’s a manual process. FileDeleter
shows you an interface in your notebook where you can look at some images and delete any that you think look incorrect.
That means you need to git pull
to get the latest version of the library, or use conda. Instructions here: https://course-v3.fast.ai/update_gcp.html#update-the-fastai-library
The question was already asked somewhere on forums. However, I would like to duplicate this issue here if someone else had it:
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-4-440e075cecad> in <module>
----> 1 a = tensor(3.,2); a
TypeError: tensor() takes 1 positional argument but 2 were given
When trying to run cells from SGD notebook. You can solve it with:
a = tensor([3.,2]); a
Probably something with a version of pytorch
?