No, inference is done on unseen data.
You will already have trained a model by the time you’re doing inference, so you won’t need any additional training data, no.
I did not see any API in fastai, that would training data independent.
Where does fastai accept constructive criticism?
Metrics is just stuff that gets printed out. So this prints the error rate as it trains.
CPUs and GPUs might give different floating-point results, e.g., due to better IEEE support on CPUs etc. Is that a concern when a model is trained on GPUs and classifies on CPUs? E.g., I have heard of studies where model performance deteriorates when rounding modes are changed.
Anybody else having trouble SSH
'ing into their instances? I think it may be an issue with the firewall at PG&E. I can curl google.com
but cannot ping google.com
.
The pull request feature on GitHub.
about the course not GitHub
Or you can create an issue on github, if you don’t have a solution to offer via a pull request.
I am using google now. No problems.
Why is ds_tfms=get_transforms()
used in the ImageDataBunch for training and tfms=get_transforms()
in the one for for inference?
It’s not exactly the same API that is used. But I agree it’s confusing so we’ll try to make it better.
data = ImageDataBunch.from_folder(
path, train=".", valid_pct=0.2,
ds_tfms=get_transforms(), size=224, num_workers=4
).normalize(imagenet_stats)
I don’t fully understand the path
and train
arguments in the above.
How to know the confidence the model has on this prediction? How to know the percentage of confidence for other predictions?
would be interesting to know on the top of it which one are done by default and which one can be on epxlicit
'create_cnn(data2, models.resnet34) ’ will create the model with the same architecture as ‘data2’ loaded weights is it?
Path is path of your data folder and ‘.’ means current directory.
path is the directory where is will look for folder named “.” and where a folder named “models” is. train is normally the name of the folder of the train images, in this case, as he didn’t have a folder for validation. it is the current folder.
i believe this creates a model with same architecture, will then have to use the .load() method on the learner to load the weights that have been saved.