Keras，从DataFrameIterator中获取对应标签的numpy数组

Question

I've this issue with Keras, My test_set is defined as follows:

My test set

test_datagen = ImageDataGenerator(rescale = 1./255)
test_set = test_datagen.flow_from_dataframe(dataframe=X_test,
                                            x_col='image_path',
                                            y_col='category_id',
                                            #imagepath_test,
                                            target_size = (64, 64),
                                            batch_size = 32,
                                            class_mode = 'categorical')

To resume shortly, I use ImageDataGenerator from Keras, and then I assign the datagenerator to my test_set.

My dataframe X_test has 2 columns,

x_col='image_path' ## The path to my image files
y_col='category_id' ## My categorical features - Labels 

I need to extract the values in y_col from my test_set, because,

test_set['category_id'] is not in same order neither same shape as X_test['category_id'] 
X_test['category_id'].shape
(315,)

When I look at test_set type i get:

type(test_set)
keras_preprocessing.image.dataframe_iterator.DataFrameIterator

The reason:

When we make predictions we need to predict on this "test_set"

#_________________________________
# Making Predictions
y_preds = classifier.predict(test_set)

So, When I want to display my Classification report I can't use "test_set" because of the wrong format and I can't use my X_test['categorical_id'] because the true values are not in the same order as test_set.

Below an example  of classification_report with test_set and the result:

print(classification_report(test_set, predicted, 
  target_names=df_data['product_cat1'].unique()))

As result I get an error:

ValueError: Found input variables with inconsistent numbers of samples: [10, 315]

Remember, my 'X_test' shape is:

 X_test['category_id'].shape
    (315,)

Whatever I've tried to convert this 'test_set' into array or dataframe didn't work.

If I use X_test['category_id'] in my classifiction_report, it works but the scores are fake,

Otherwise, Multiclass Classification using Keras is funny but useless, if we can't identify precision and recall score for each class and f1_score accuracy model, I mean we just get a global accuracy model and that's all, only good to competitions.

Any ideas, workarounds are welcome.

Answer 1

Since it is test_set and since you do not train on test_set you do not have to shuffle it so that the order is preserved. That way you will know the order of labels (ground truth) from X_test['category_id'] and use the same for the classification_report

Fix

test_set = test_datagen.flow_from_dataframe(dataframe=X_test,
                                            x_col='image_path',
                                            y_col='category_id',
                                            shuffle=False,        ### Do not shuffle
                                            target_size = (64, 64),
                                            batch_size = 32,
                                            class_mode = 'categorical')

y_preds = classifier.predict(test_set)
print(classification_report(test_set, X_test['category_id']))

In case you want to shuffle the test_set then you can seed it with a value and make the prediction. Then use the same seed and iterate over the datagenerator and collect the labels (ground truth). With the same seed value set you will get the same order.