使用功能 API 和 tf.GradientTape() 的组合在 Tensorflow 2.0 中进行训练时，如何将模型图记录到张量板？

Question

Can some please guide me on how to log model graph to tensorboard when I am using Keras Functional API or the model sub calling API to create the model and tf.GradientTape() to train the model?

# Get the model.
inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, activation='softmax', name='predictions')(x)
model = keras.Model(inputs=inputs, outputs=outputs)


optimizer = keras.optimizers.SGD(learning_rate=1e-3)
loss_fn = keras.losses.SparseCategoricalCrossentropy()


batch_size = 64
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(batch_size)

# Iterate over epochs.
epochs = 3
for epoch in range(epochs):
    print('Start of epoch %d' % (epoch,))

    for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):   

    with tf.GradientTape() as tape:
        logits = model(x_batch_train)
        loss_value = loss_fn(y_batch_train, logits)

    grads = tape.gradient(loss_value, model.trainable_weights)


    optimizer.apply_gradients(zip(grads, model.trainable_weights))

    if step % 200 == 0:
        print('Training loss (for one batch) at step %s: %s' % (step, float(loss_value)))
        print('Seen so far: %s samples' % ((step + 1) * 64))

Where should I insert the tensorboard logging for the model graph?

Answer 1

The best way to do this (TF 2.3.0, TB 2.3.0) is to use tf.summary and pass the model through a wrapper function with the @tf.function decorator.

In your case, to export the model graph to TensorBoard for inspection:

inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, activation='softmax', name='predictions')(x)
model = keras.Model(inputs=inputs, outputs=outputs)


# Solution Code:
writer = tf.summary.create_file_writer('./logs/')

@tf.function
def init_model(data, model):
    model(data)

tf.summary.trace_on(graph=True)
init_model(tf.zeros((1,784), model)

with writer.as_default():
    tf.summary.trace_export('name', step=0)

Looking at the logs files in TensorBoard you should see a model graph as such.