在TensorFlow 2.0中：使用optimizer.apply_gradients训练错误

Question

I am trying to learn the new TF 2.0 alpha release. I'm training a Sequential model for a binary classification purpose. My datatable is df , which is a numpy array. classification is the one-hot encoding dataframe of the classes I must predict.

The definition of the model is clear, as it is the definition of loss and accuracy functions and the (Adam) optimizer. However, I get an error at the point of training:

loss_history = []
accuracy_history = []

for epoch in range(n_epochs):

    with tf.GradientTape() as tape:

        # compute binary crossentropy loss (bce_loss)
        current_loss = bce_loss(model(df), classification.astype(np.float64))
        loss_history.append(current_loss)

        # train the model based on the gradient of loss function
        gradients = tape.gradient(current_loss, model.trainable_variables)
        optimizer.apply_gradients([gradients, model.trainable_variables])  # optimizer = Adam


    # print the training progress
    print(str(epoch+1) + '. Train Loss: ' + str(metrics) + ', Accuracy: ' + str(current_accuracy))

print('\nTraining complete.')

At this point, I get error pointed at optimizer.apply_gradients() . The error message says:

ValueError: too many values to unpack (expected 2)

Where is my mistake?

I did some research on this type of error, but I found nothing useful related to this particular function. Any help is appreciated.

Answer 1

试试这个：

optimizer.apply_gradients(zip(gradients, model.trainable_variables))