January 20, 2020

710 words 4 mins read

fastai/fastdot

fastai/fastdot

A simple wrapper over pydot and graphviz which fixes some sharp edges

repo name fastai/fastdot
repo link https://github.com/fastai/fastdot
homepage https://fastai.github.io/fastdot/
language Jupyter Notebook
size (curr.) 1699 kB
stars (curr.) 36
created 2019-12-13
license Apache License 2.0

fastdot

A simple wrapper over pydot to make it more consistent, unsurprising, and pythonic

Acknowledgement: fastdot is heavily influenced by work from David Page, who built a system for drawing graphs based on a highly flexible data structure he designed. It may turn out that his library, when complete, will make it even easier to do the things described here (and his data structure also supports a lot more than just drawing)!

Install

pip install fastdot

Synopsis

Start with some data representing objects and connections between them (e.g. they wouldn’t normally be just strings like in this example, but would be neural net layers, or users and products, or car trips, etc):

layers1 = ['conv','conv','lin']
layers2 = ['conv','lin']
block1,block2 = ['block1','block2']
conns = ((block1, block2),
         (block1, layers2[-1]))

Then map them directly to a visual respresentation:

g = graph_items(seq_cluster(layers1, block1),
                seq_cluster(layers2, block2))
g.add_items(*object_connections(conns))
g

svg

See the symbolic graphs and object graphs sections below for a more complete example.

fastdot overview

fastdot is a thin wrapper over the excellent pydot program (which is in turn a thin wrapper over the absolutely wonderful Graphviz software), designed to make it more consistent, unsurprising, and pythonic. (An example of removing surprise: pydot.Node('node') gives an obscure compilation exception, since node is a keyword in the underlying graphviz program, whereas fastdot.Node('node') works just fine, due to auto-quoting.) In fact, you never need to provide names in fastdot; you can create edges directly between objects.

Here’s a quick example of some of the main functionality:

g = Dot()
c = Cluster('cl', fillcolor='pink')
a1,a2,b = c.add_items('a', 'a', 'b')
c.add_items(a1.connect(a2), a2.connect(b))
g.add_item(Node('Check tooltip', tooltip="I have a tooltip!"))
g.add_item(c)
g

svg

As you see, graphs know how to show themselves in Jupyter notebooks directly and can be exported to HTML (it uses SVG behind the scenes). Tooltips appear in both notebooks and exported HTML pages. Nodes with the same label, by default, are set to the same color. Also, as shown above, you can just use add_item or add_items, regardless of the type of item.

Symbolic graphs

fastdot is particularly designed to make it easier to create graphs symbolically - for instance, for Python dictionaries, PyTorch/TensorFlow models, and so forth. Here’s a simple example with some mock neural network layers and sequential models. First, let’s define our mock classes:

@dataclass(frozen=True)
class Layer: name:str; n_filters:int=1
class Linear(Layer): pass
class Conv2d(Layer): pass

@dataclass(frozen=True)
class Sequential: layers:list; name:str

Here’s our sequential blocks for our “model”:

block1 = Sequential([Conv2d('conv', 5), Linear('lin', 3)], 'block1')
block2 = Sequential([Conv2d('conv1', 8), Conv2d('conv2', 2), Linear('lin')], 'block2')

fastdot can create all node properties directly from objects; you just have to define functions describing how to map the object’s attributes to graph properties. These mappings go in the node_defaults and cluster_defaults dictionaries (although by default labels are set using str(), so we don’t need any special cluster defaults in this case):

node_defaults['fillcolor'] = lambda o: 'greenyellow' if isinstance(o,Linear) else 'pink'
cluster_defaults['label'] = node_defaults['label'] = attrgetter('name')
node_defaults['tooltip'] = str

With that in place, we can directly create nodes from our objects, for instance using the convenient seq_cluster function:

c1 = seq_cluster(block1.layers, block1)
c2 = seq_cluster(block2.layers, block2)
e1,e2 = c1.connect(c2),c1.connect(c2.last())
graph_items(c1,c2,e1,e2)

svg

Note that in this example we didn’t even need to create the Dot object separately - graph_items creates it directly from the graph items provided.

Using object graphs

In the above example, we defined our edges directly between fastdot objects. In practice, however, you’ll most likely have your edges defined directly between python objects, for instance like this:

conns = (
    (block1, block2),
    (block1, block2.layers[-1]),
)

In this case, you’ll want some way to connect your python objects to the fastdot graph items that represent them. A mapping is stored automatically by fastdot, and is made available through the object2graph function:

g = graph_items(seq_cluster(block1.layers, block1), seq_cluster(block2.layers, block2))
object2graph(block1.layers[-1])
<pydot.Node at 0x7f013180c310>

You can use this to graph your connections without needing access to the graph items:

g.add_items(*[object2graph(a).connect(object2graph(b))
              for a,b in conns])
g

svg

There’s a helper function, object_connections, which creates these connections for you. So the above can be simplified to:

g = graph_items(seq_cluster(block1.layers, block1), seq_cluster(block2.layers, block2))
g.add_items(*object_connections(conns))
g

svg

comments powered by Disqus