Currently when i want to make csv submission to Kaggle, and have the csv loaded as a dataframe, i run
def applied_fn(s):
s[1] = predict_this_id(s[0])
return s
df_predict.apply(applied_fn, axis=1, raw=True)
Where the predict_this_id loads the image from the path etc…
It is however rather slow as predict_this_id
loads the image from disk and does the prediction .
I was wondering if fastai v1 has got a tool for this to load all images, or in bunches, and then one can do the predictions from memory. Or does it not make much difference on an SD anyway.
Checking Github before i post this, i realized that there is ’get_image_files’ in fastai.vision.data that loads a FilePathList, using get_files
from fastai.data_block
How does fastai load images into memory when creating the databunch, although i guess that loads it into GPU memory…