No graph is built for visualization

[leom97]

To debug my dolfin-adjoint code I want to visualize the compute graph.

I receive this error after calling tape.visualize():

File "/home/anaconda3/envs/masters_thesis/lib/python3.10/site-packages/tensorflow/python/ops/script_ops.py", line 632, in py_func_common
    result = func(*[np.array(x) for x in inp])
TypeError: Tape._tf_add_blocks.<locals>.<lambda>() takes 0 positional arguments but 2 were given

This is fixed by the following:

    try:
      result = func(*[np.array(x) for x in inp])
    except:
      result = None

inspired from the following lines of tape.py, where one sees that a lambda with zero args is eventually called with more than zero inputs:

tf.py_func(lambda: None, in_tensors, [tf.float64],
                                            name=self._valid_tf_scope_name(str(block)))

Anyhow, tensorboard now says empty graph, so that nothing is visualized. How can this be solved?

[dham]

I'm not familiar with the tensorflow-based visualisation so I'll leave the direct part to others. However, if you are using pyadjoint master then direct PDF visualisation is also supported and could be a workaround. If you pass a filename ending in .pdf then you get a PDF visualisation. E.g. tape.visualise("tape.pdf"). This requires the python package networkx and the system package graphviz (e.g. sudo apt install graphviz).

[leom97]

@dham This does indeed work, although the resulting pdf visualization if quite hard to read, without the possibility to expand and collapse blocks. Thank you for the answer in the meantime.