从训练有素的 Keras 模型打印结果概率答案

【问题标题】：Printing outcome probabilities from a trained Keras Model从训练有素的 Keras 模型打印结果概率
【发布时间】：2021-08-04 13:04:18
【问题描述】：

当我提供新原始数据时，我正尝试从我训练的模型中打印每个类别结果的预测概率。这是一个多类分类问题，有 8 个输出和 21 个输入。

当我展示新数据时，我可以打印 1 个结果，例如：

 "Example 0 prediction: 1 (15.0%)"

相反，我希望看到类似于下面的内容。其中显示了每个类（0、1、2、3、4、6、Wide、Out）的概率：

Example 0 prediction 0: (12.5%), prediction 1: (12.5%), prediction 2: (12.5%), prediction 3: (12.5%), prediction 4: (12.5%), prediction 6: (12.5%), prediction Wide: (12.5%), prediction Out: (12.5%)

请注意，我已尝试搜索类似问题，包括 here、here 和 here，并查阅了 TensorFlow 文档。但是，这些主要讨论模型本身的更改，例如在最后一层激活softmax，将分类交叉熵作为损失函数等，从而生成概率。

我已经包含了模型架构以及预测代码以实现完全可见性。

型号：

earlystopping = callbacks.EarlyStopping(monitor ="val_loss", 
                                        mode ="min", patience = 125, 
                                        restore_best_weights = True)
  
#define Keras
model = Sequential()
model.add(Dense(50, input_dim=21))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dropout(0.5,input_shape=(50,)))
model.add(Dense(50))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dropout(0.5,input_shape=(50,)))
model.add(Dense(8, activation='softmax'))

#compile the keras model
model.compile(loss='categorical_crossentropy', optimizer='Adam', metrics=['accuracy'])   

model.fit(X, dummy_y, validation_split=0.25, epochs=1000, batch_size=100, verbose=1, callbacks=[earlystopping])

_, accuracy3 = model.evaluate(X, dummy_y, verbose=0)
print('Accuracy: %.2f' % (accuracy3*100))

做出预测：

class_names = ['0', '1', '2','3','4','6','Wide','Out']

predict_dataset = tf.convert_to_tensor([
  [1,5,1,0.459,0.322,0.041,0.002,0.103,0.032,0.041,14,0.404,0.284,0.052,0.008,0.128,0.044,0.037,0.043,54,0,],
    [1,18,5,0.512,0.286,0,0,0.083,0.024,0.095,13,0.24,0.44,0.08,0,0.08,0.08,0,0.08,173,3],
    [2,11,13,0.5,0.417,0,0,0.083,0,0.083,82,0.35,0.36,0.042,0.003,0.135,0.039,0.051,0.02,51,7]
])  

predictions = model(predict_dataset, training=False)

for i, logits in enumerate(predictions):
    class_idx = tf.argmax(logits).numpy()
    p = tf.nn.softmax(logits)[class_idx]
    name = class_names[class_idx]
    print("Example {} prediction: {} ({:4.1f}%)".format(i, name,100*p))

输出：

Example 0 prediction: 1 (15.0%)
Example 1 prediction: 1 (16.0%)
Example 2 prediction: 0 (16.9%)

我已经尝试对使用 TensorFlow 的 logits 的 for 循环进行更改，但我仍然无法让它打印每个结果和相关的概率。

非常感谢任何指导。

【问题讨论】：

标签： tensorflow machine-learning keras

【解决方案1】：

最后，我没有尝试实现 For 循环，而是直接打印了 numpy 数组中的每个结果。

不是最干净的方式，但它可以完成工作。希望将来对某人有用。

predict_dataset = tf.convert_to_tensor([
    [1,5,1,0.459,0.322,0.041,0.002,0.103,0.032,0.041,14,0.404,0.284,0.052,0.008,0.128,0.044,0.037,0.043,54,0,155]
])

predictions = model3(predict_dataset, training=False)
       
predictions2 = predictions.numpy()

prob_0 = predictions2[0,0]
prob_1 = predictions2[0,1]
prob_2 = predictions2[0,2]
prob_3 = predictions2[0,3]
prob_4 = predictions2[0,4]
prob_wide = predictions2[0,5]
prob_6 = predictions2[0,6]
prob_wicket = predictions2[0,7]

print(prob_0)
print(prob_1)
print(prob_2)
print(prob_3)
print(prob_4)
print(prob_wide)
print(prob_6)
print(prob_wicket)

输出

0.28349978
0.32451397
0.06382967
0.0053077294
0.20397986
0.07999096
6.386134e-08
0.038877998

【讨论】：