TensorFlow混淆矩阵顺序确认：tf-slim生成6类矩阵的行列含义

Great question—this is a common point of confusion (pun intended) across ML tools! Let's clarify this for your TF-Slim use case:

1. Your Initial Statement Is Correct #

TF-Slim's confusion matrix implementation aligns with TensorFlow's native tf.math.confusion_matrix behavior, which follows:

• Rows = Ground Truth (True) Labels
• Columns = Predicted Labels

Let's verify this with your matrix to make it concrete:

# Your 6-class confusion matrix
[
 [41, 2, 0, 0, 0, 0],  # True class 0: 41 correct, 2 misclassified as class 1
 [1, 11, 4, 1, 0, 0],  # True class 1: 11 correct, 1 misclassified as 0, 4 as 2, 1 as 3
 [0, 1, 12, 0, 0, 0],  # True class 2: 12 correct, 1 misclassified as class 1
 [0, 0, 0, 22, 1, 0],  # True class 3: 22 correct, 1 misclassified as class 4
 [0, 0, 0, 0, 7, 0],   # True class 4: 7 correct, no misclassifications
 [0, 0, 0, 0, 0, 20]   # True class 5: 20 correct, no misclassifications
]

Each row sums to the total number of samples in that true class (e.g., row 0 sums to 43, which is 41+2—total true class 0 samples). Each column sums to the total number of predictions for that class (e.g., column 1 sums to 2+11+1=14—total predictions for class 1). This matches the "rows = true, columns = predicted" structure perfectly.

2. Why Do Some Sources Say the Opposite? #

The mix-up happens because there’s no universal standard across all ML libraries or custom code:

• Some older tools or hand-rolled implementations swap the axes (columns = true labels, rows = predictions)
• Certain visualization libraries might flip the matrix for display readability
• Always double-check the documentation of the specific tool you’re using to confirm the order

For TF-Slim specifically, since it’s a lightweight wrapper around TensorFlow’s core APIs, you can trust that it uses the same axis order as TensorFlow’s official confusion matrix function.

内容的提问来源于stack exchange，提问作者Jame