张量板直方图上未显示张量流 v2 梯度

Question

我有一个简单的神经网络，我正在尝试使用 tensorboard 通过使用如下回调来绘制梯度：

class GradientCallback(tf.keras.callbacks.Callback):
    console = False
    count = 0
    run_count = 0

    def on_epoch_end(self, epoch, logs=None):
        weights = [w for w in self.model.trainable_weights if 'dense' in w.name and 'bias' in w.name]
        self.run_count += 1
        run_dir = logdir+"/gradients/run-" + str(self.run_count)
        with tf.summary.create_file_writer(run_dir).as_default(),tf.GradientTape() as g:
          # use test data to calculate the gradients
          _x_batch = test_images_scaled_reshaped[:100]
          _y_batch = test_labels_enc[:100]
          g.watch(_x_batch)
          _y_pred = self.model(_x_batch)  # forward-propagation
          per_sample_losses = tf.keras.losses.categorical_crossentropy(_y_batch, _y_pred) 
          average_loss = tf.reduce_mean(per_sample_losses) # Compute the loss value
          gradients = g.gradient(average_loss, self.model.weights) # Compute the gradient

        for t in gradients:
          tf.summary.histogram(str(self.count), data=t)
          self.count+=1
          if self.console:
                print('Tensor: {}'.format(t.name))
                print('{}\n'.format(K.get_value(t)[:10]))

# Set up logging
!rm -rf ./logs/ # clear old logs
from datetime import datetime
import os
root_logdir = "logs"
run_id = datetime.now().strftime("%Y%m%d-%H%M%S")
logdir = os.path.join(root_logdir, run_id)


# register callbacks, this will be used for tensor board latter
callbacks = [
    tf.keras.callbacks.TensorBoard( log_dir=logdir, histogram_freq=1, 
                                   write_images=True, write_grads = True ),
    GradientCallback()
]

然后，我在 fit 期间使用回调：

network.fit(train_pipe, epochs = epochs,batch_size = batch_size, validation_data = val_pipe, callbacks=callbacks)

现在，当我检查张量板时，我可以在左侧过滤器上看到渐变，但直方图选项卡中没有显示任何内容：

我在这里缺少什么？我是否正确记录了梯度？

Answer 1

看起来问题在于您在 tf 摘要编写器的上下文之外编写直方图。 我相应地更改了您的代码。但我没试过。

class GradientCallback(tf.keras.callbacks.Callback):
    console = False
    count = 0
    run_count = 0

    def on_epoch_end(self, epoch, logs=None):
        weights = [w for w in self.model.trainable_weights if 'dense' in w.name and 'bias' in w.name]
        self.run_count += 1
        run_dir = logdir+"/gradients/run-" + str(self.run_count)
        with tf.summary.create_file_writer(run_dir).as_default()
          with tf.GradientTape() as g:
            # use test data to calculate the gradients
            _x_batch = test_images_scaled_reshaped[:100]
            _y_batch = test_labels_enc[:100]
            g.watch(_x_batch)
            _y_pred = self.model(_x_batch)  # forward-propagation
            per_sample_losses = tf.keras.losses.categorical_crossentropy(_y_batch, _y_pred) 
            average_loss = tf.reduce_mean(per_sample_losses) # Compute the loss value
            gradients = g.gradient(average_loss, self.model.weights) # Compute the gradient

          for nr, grad in enumerate(gradients):
            tf.summary.histogram(str(nr), data=grad)
            if self.console:
                  print('Tensor: {}'.format(grad.name))
                  print('{}\n'.format(K.get_value(grad)[:10]))