python - 将 KerasTensor 对象转换为 numpy 数组以在 Callback 中可视化预测

Question

我正在为 tensorflow keras 顺序模型的 model.fit() 方法编写自定义 on_train_end 回调函数。回调函数是关于绘制模型所做的预测，因此它涉及将模型的输入转换为 numpy 数组并将其提供给 model.predict()。我使用 self.model.inputs 访问输入，这是 KerasTensor 对象的列表，第 0 个索引处的对象是我想要的。我尝试了以下方法

class my_visualizer(tf.keras.callbacks.Callback):

    def on_train_end(self, logs=None):

        x = tf.keras.backend.eval(self.model.inputs[0])
        y_predictions = self.model.predict(x)
        

但得到了错误

AttributeError: 'KerasTensor' object has no attribute 'numpy'

所以这个方法是针对另一种类型的张量而不是 KerasTensor。我发现其他解决方案适用于 tensorflow 的 Tensor 对象，但不适用于 keras 的 KerasTensor 对象，并且我没有在 keras 文档中找到任何关于实现所需功能的方法的提及。谢谢你的帮助！

标签： pythontensorflowkerastf.keras

Answer 1

虽然它不完全使用KerasTensor对象，但在 Tensorflow 2.x 上对我有用的是将真实数据传递给Callback：

import tensorflow as tf
import matplotlib.pyplot as plt

def legend_without_duplicate_labels(ax):
    handles, labels = ax.get_legend_handles_labels()
    unique = [(h, l) for i, (h, l) in enumerate(zip(handles, labels)) if l not in labels[:i]]
    ax.legend(*zip(*unique))

class Visualizer(tf.keras.callbacks.Callback):
   def __init__(self, ds):
        self.ds = ds
   def on_train_end(self, epoch, logs=None):
        features, true_labels = next(iter(self.ds.take(1)))
        ynew = self.model.predict(features)
        labels = [1 if y > 0.5 else 0 for y in tf.squeeze(ynew)]
        true_labels = [y.numpy() for y in tf.squeeze(true_labels)]
        fig, axes = plt.subplots(1, 2)
        fig.set_figheight(10)
        fig.set_figwidth(10)
        cdict = {0: 'red', 1: 'blue'}
        titles = ['True Labels', 'Predicted Labels']
        for ax, ls, t in zip(axes.flatten(), [true_labels, labels], titles):
          for i, txt in enumerate(ls):
            ax.scatter(features[i, 0], features[i, 1], c = cdict[txt], marker="o", label = txt, s = 100)
          legend_without_duplicate_labels(ax)
          ax.title.set_text(t)
        plt.show()

inputs = tf.keras.layers.Input((2,))
x = tf.keras.layers.Dense(32, activation='relu')(inputs)
outputs = tf.keras.layers.Dense(units=1, activation='sigmoid')(x)
model = tf.keras.Model(inputs, outputs)
model.compile(optimizer='adam', loss=tf.keras.losses.BinaryCrossentropy())

train_ds = tf.data.Dataset.from_tensor_slices((tf.random.normal((50, 2)), tf.random.uniform((50, 1), maxval=2, dtype=tf.int32))).batch(10)
test_ds = tf.data.Dataset.from_tensor_slices((tf.random.normal((50, 2)), tf.random.uniform((50, 1), maxval=2, dtype=tf.int32))).batch(50)
model.fit(train_ds, epochs=10, callbacks=[Visualizer(test_ds)])

Epoch 1/10
5/5 [==============================] - 0s 3ms/step - loss: 0.7521
Epoch 2/10
5/5 [==============================] - 0s 2ms/step - loss: 0.7433
Epoch 3/10
5/5 [==============================] - 0s 2ms/step - loss: 0.7363
Epoch 4/10
5/5 [==============================] - 0s 2ms/step - loss: 0.7299
Epoch 5/10
5/5 [==============================] - 0s 2ms/step - loss: 0.7239
Epoch 6/10
5/5 [==============================] - 0s 2ms/step - loss: 0.7183
Epoch 7/10
5/5 [==============================] - 0s 2ms/step - loss: 0.7131
Epoch 8/10
5/5 [==============================] - 0s 2ms/step - loss: 0.7082
Epoch 9/10
5/5 [==============================] - 0s 2ms/step - loss: 0.7037
Epoch 10/10
5/5 [==============================] - 0s 2ms/step - loss: 0.6994