Using csv for images and captions with training_clip

[rbunn80110]

This looks like really amazing work!

I have a csv with "file_path" and "caption" already in the dataframe. Which means I don't need to map captions to ids in filenames. Has anyone modified this to get all the data directly from the csv file? I figure my scenario is a lot simpler code wise, but struggling to figure out how to refactor out the lookup dictionaries and whatnot.

[rbunn80110]

I went with this for now:

df = pd.read_csv("captions.csv")

cids = df.index

def read_image(cid):
    img = PILImage.create(df.iloc[[cid]]["filename"].values[0])
    return img

def read_text(cid):
    caption = df.iloc[[cid]]["caption"].values[0]
    return caption

def dummy_targ(o): return 0 # loss func is not called without it

[rbunn80110]

And then use RandomSplitter()