python - 如何在 Tensorboard 中通过 Eager Execution 可视化 keras 卷积过滤器

Question

给定以下模型，是否可以从每个卷积滤波器中获取图像？我似乎无法找到解决办法。

    tf.enable_eager_execution()

    l = tf.keras.layers
    max_pool = l.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)

    kmodel = tf.keras.Sequential(
        [
            l.Reshape(target_shape=input_shape, input_shape=(IMAGE_SIZE * IMAGE_SIZE * colors,)),
            l.Conv2D(32, (5,5), padding='same', data_format=data_format, activation=tf.nn.relu),
            max_pool,
            l.Conv2D(64, (5,5), padding='same', data_format=data_format, activation=tf.nn.relu),
            max_pool,
            l.Conv2D(128, (5,5), padding='same', data_format=data_format, activation=tf.nn.relu),
            max_pool,
            l.Conv2D(256, (5,5), padding='same', data_format=data_format, activation=tf.nn.relu),
            max_pool,
            l.Flatten(),
            l.Dense(1024, activation=tf.nn.relu),
            l.Dropout(0.4),
            l.Dense(num_classes) #num_classes
        ]
    )

# how to get tf.contrib.image for each of the 4 filters?

靠近但没有雪茄：

https://stackoverflow.com/a/35858950

https://github.com/InFoCusp/tf_cnnvis

谢谢！

澄清更新：也使用 tf.GradientTape API

标签： pythontensorflowkerastensorboard