Graph: show output nodes of a model if it is not connected to any other nodes

Version: tensorflow 2.0.0a0

Python code for bug reproduction:

import numpy as np
import tensorflow as tf

print(tf.__version__)

input1 = tf.keras.Input(shape=[3, 4])

lstm = tf.keras.layers.GRU(1, return_state=True)
[output1, output2] = lstm(input1)

dense = tf.keras.layers.Dense(1)
output1 = dense(output1)

model = tf.keras.Model(inputs=input1, outputs=[output1, output2])
model.compile(loss=['mean_squared_error', 'mean_squared_error'],
              optimizer='sgd')

num_examples = 8
xs1 = np.zeros([num_examples, 3, 4])
ys1 = np.ones([num_examples, 1])
ys2 = np.zeros([num_examples, 1])

model.fit(xs1, [ys1, ys2], epochs=10,
          callbacks=[tf.keras.callbacks.TensorBoard('/tmp/foo6

Then start tensorboard:

tensorboard --logdir /tmp/foo6

Navigate to the Graph dashboard and select tag "keras". The graph looks like the following:

This graph is misleading in a number of ways:

1. It doesn't indicate that the model has two outputs (output1 and output2 in the code)
2. It doesn't indicate that the GRU (unified_gru) layer has two outputs
3. It seems to indicate that both outputs of the GRU layer go into the dense layer, whereas in reality only the first one does and the second one is directly routed to the output of the entire model.

Start by running the Python reproduction with TensorBoard and inspect the Graph dashboard's "keras" tag. Trace how the displayed graph represents the model's output nodes and the GRU's two outputs. Done means the graph distinguishes both model outputs and shows that only the first GRU output feeds the dense layer.