Multilabel regression - pick labems


I am looking at a multi label regression model.
The labels are face landmarks (left_eye_center, left_eye_left … etc).
I am trying to create a databunch using ImageDataBunch. In this case, train_df_path has image path as the first column and the remaining facial attribute coordinates as the y value (labels). I have tried below DataBunch and I get an error

This is the ImageDataBunch code

tfms = get_transforms(do_flip=False)
data = ImageDataBunch.from_df(path, train_df_path, ds_tfms=tfms,  size=24, label_delim= ' ') \
   .split_by_rand_pct(0.2) \
   .label_from_df(cols=cols[1:]) \

Appreciate any help on this.


Forgot to mention the error

AssertionError: You have NaN values in column(s) [‘left_eye_center_x’, …] of your dataframe, please fix it.
Before replacing NaN,

True     28
False     3
dtype: int64

I get it despite having train_df.fillna(method = 'ffill',inplace = True)

After running above, I get

False    31
dtype: int64

@ctippur this is late but were you able to solve this? I know that typically you use points like -1,-1 to replace missing values. Are you doing something like that here?