Keras flow_from_dataframe

Question

I am trying to use keras flow_from_dataframe in order to augment the input images to my network (I am using the latest keras_preprocessing version). The network has an input image and an image as a label. I have a csv file in the following format that specifies the input-output:

Input,Output
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\1\1.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\1.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\1\2.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\1.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\1\3.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\1.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\1\4.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\1.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\1\5.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\1.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\1\6.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\1.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\1\7.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\1.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\104\1.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\104.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\104\10.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\104.JPG
E:\LTM\SCIE_DB\Dataset_Part1\Dataset\104\11.JPG,E:\LTM\SCIE_DB\Dataset_Part1\Label\104.JPG

I am trying to load the csv to a dataframe by the following code:

df=pd.read_csv(r"E:\LTM\SCIE_DB\dataset.csv",names=("Input","Output"),dtype={"Input":"str","Output":"str"})
df = df.astype(str)
print(df.dtypes)

datagen=image.ImageDataGenerator(rescale=1./255)
train_generator=datagen.flow_from_dataframe(dataframe=df, directory=None, x_col="Input", y_col="Output", class_mode="other", target_size=(32,32), batch_size=32,has_ext = True)

However, I keep getting the following error :

TypeError: y_col column/s must be numeric datatypes.

No matter what I am doing, I am unable to set the y_col as a str and it is always an object type. Any ideas?

Answer 1

Your values for y_col are of "string" type, but in your code you have refereed it as "numeric" since you specified class_mode as:

class_mode="other"

So change the class mode to

class_mode="input"

so you can use "string" values for y_col