试图理解 model.predict 的输出

Question

我正在尝试使用 model.predict 方法预测猫/狗图像。因为它是 2 类分类器，所以我得到一个包含 2 个值的数组。根据我的理解，这些值代表属于每个类别的概率（如果我错了，请纠正我）。如果是这样，则概率必须相加为 1。但我得到两个类的概率相同

模型历史

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten (Flatten)            multiple                  0         
_________________________________________________________________
dense (Dense)                multiple                  30848     
_________________________________________________________________
dropout (Dropout)            multiple                  0         
_________________________________________________________________
batch_normalization (BatchNo multiple                  512       
_________________________________________________________________
dense_1 (Dense)              multiple                  12900     
_________________________________________________________________
dropout_1 (Dropout)          multiple                  0         
_________________________________________________________________
batch_normalization_1 (Batch multiple                  400       
_________________________________________________________________
dense_2 (Dense)              multiple                  10100     
_________________________________________________________________
dropout_2 (Dropout)          multiple                  0         
_________________________________________________________________
batch_normalization_2 (Batch multiple                  400       
_________________________________________________________________
dense_3 (Dense)              multiple                  10100     
_________________________________________________________________
dropout_3 (Dropout)          multiple                  0         
_________________________________________________________________
batch_normalization_3 (Batch multiple                  400       
_________________________________________________________________
dense_4 (Dense)              multiple                  10100     
_________________________________________________________________
dropout_4 (Dropout)          multiple                  0         
_________________________________________________________________
batch_normalization_4 (Batch multiple                  400       
_________________________________________________________________
dense_5 (Dense)              multiple                  202       
=================================================================
Total params: 76,362
Trainable params: 75,306
Non-trainable params: 1,056

---

预测代码

class_prob=model.predict(new_array_2.T,batch_size=1)
print(class_prob)
classifications=model.predict_classes(new_array_2.T,batch_size=1)
print(classifications)
print(CATEGORIES[classifications[0]])

输出

[[0.39456758 0.39456758]]
[0]
Dog

Answer 1

model.predict 只返回给定输入的模型计算输出，因此您提到的所有细节都取决于模型的输出，例如最后一层的激活。

模型输出总和为 1 的概率仅由输出层的 softmax 激活产生，在我看来，您的最后一层有不同的激活，可能是 sigmoid，它不会产生总和为一。您应该更喜欢使用softmax 进行多类分类。