Use dataframe instead of folder structure about action_recognition HOT 3 CLOSED

Hey,
This should be pretty straight forward using the DataBlock API. Just define your get_x and get_y, with the corresponding columns in the dataframe. For each row of your dataframe, you get a pair (folder_path, target_value). Something like:

def get_x(row): 
    return Path(row[path_col_name]).ls().sorted()

def get_y(row):
    return row[target_col]

You could also use the fastai method:

Init signature: ColReader(self, cols, pref='', suff='', label_delim=None)
Docstring:      Read `cols` in `row` with potential `pref` and `suff`
File:           ~/miniconda3/envs/fastai/lib/python3.8/site-packages/fastai/data/transforms.py
Type:           _TfmMeta

There is also a more detailed version of the action notebooks on a PR that I am making for fastai :
https://github.com/fastai/fastai/blob/0c18d134099b62dcc9f9b105e2295088f38abab4/nbs/24_tutorial.image_sequence.ipynb

from action_recognition.

yes, that was it. I was making it way too complicated. Thanks!

from action_recognition.

I the last update, I removed the Block API. Sorry if that hurts your dev.
The mid level API is now the preferred way, and it is very well documented.

from action_recognition.