Flash fails to load image labels from CSV on Kaggle RANZCR CLiP - Catheter and Line Position Challenge about lightning-flash

[teddykoker]

You need to use ImageClassificationData.from_filepaths(), and then supply the filenames and labels as lists. The image paths will need to be extracted from the csv, something like:

# load in csv
df = pd.read_csv("train.csv")

# extract ids
ids = list(df["StudyInstanceUID"])

# get filenames by appending .jpg to id
filenames = [f"{id}.jpg" for id in ids]

# from https://www.kaggle.com/heyytanay/torch-trainer-augmentations-nfnets
# get labels for each id
labels = [label = df[df['StudyInstanceUID'] == id].values.tolist()[0][1:-1] for id in ids] 


# hypothetically you could do this, if it wasn't multilabel, see note below
data = ImageClassificationData.from_filepaths(train_filepaths=filenames, train_labels=labels, ...)

NOTE: looking at the Kaggle competition this data appears to be multi-label, which the ImageClassificationData does not support (yet?)

[edenlightning]

Looks like we need to update the docs on how to pass in train_labels (need to provide an array of labels, not a filepath)

[aribornstein]

In the code it makes it seem as if you can pass a file. Will also claimed that he had this working is it possible something changed?

See line 304

https://github.com/PyTorchLightning/lightning-flash/blob/master/flash/vision/classification/data.py

[aribornstein]

Yes that's something missing from the prediction interface a way to see the labels and a way to get the prediction class distribution. @williamFalcon

[williamFalcon]

https://github.com/PyTorchLightning/lightning-flash/blob/master/flash/data/data_utils.py#L6