How to select a specific number of each class from the MNIST dataset

Question

I'm using tensorflow to work on Mnist. I need to train my network with a specific number of data of each class (for example 500 number of sample of each digit). I've found how to sort the DB with class labels .

idx = np.argsort(y_train)
x_train_sorted = x_train[idx]
y_train_sorted = y_train[idx]

but how can I select 500 number of each one and then combine them and the shuffle?

Answer 1

If you have all in one DataFrame then you can groupby label and then get head or tail

import pandas as pd

df = pd.DataFrame({
    'X1': [1,2,3,4,5,6,7,8,9,10,11,12],
    'X2': [21,22,23,24,25,26,27,28,29,30,31,32],
    'label': ['a','a','a','a','b','b','b','b','c','c','c','c']
})

groups = df.groupby('label')

df2 = groups.head(2)    
#df2 = groups.apply(lambda x:x[:2]) # the same as head(2)
#df2 = groups.apply(lambda x:x.sample(frac=1)[:2]) # shuffled before get values

print(df2)

Result

   X1  X2 label
0   1  21     a
1   2  22     a
4   5  25     b
5   6  26     b
8   9  29     c
9  10  30     c

And after that you can shuffle it and split it into X_train , y_train

df2 = df2.sample(frac=1).reset_index(drop=True)

X_train = df2[['X1','X2']]
y_train = df2['label']

print(X_train)
print(y_train)