Home > Hadrian >

When you first look at an auto-generated PFA file, it often looks like this:

{"input": {"type": "record", "fields": [{"name": "x1", "type": "string"}, {"name": "x2",
"type": ["null", "double"]}], "name": "Input"}, "output": "Output", "action": [{"let":
{"derived": {"type": {"type": "record", "fields": [{"name": "x1", "type": "string"},
{"name": "x2", "type": "double"}], "name": "Derived"}, "new": {"x1": {"s.lower": [{"attr":
"input", "path": [{"string": "x1"}]}]}, "x2": {"u.imputeX2": [{"attr": "input", "path":
[{"string": "x2"}]}]}}}}}, {"let": {"scores": {"a.map": [{"cell": "forest"}, {"params":
[{"tree": "TreeNode"}], "ret": "string", "do": [{"model.tree.simpleWalk": ["derived",
"tree", {"params": [{"d": "Derived"}, {"node": "TreeNode"}], "ret": "boolean", "do":
[{"model.tree.simpleTest": ["d", "node"]}]}]}]}]}}}, {"let": {"winner": {"a.mode":
["scores"]}}}, {"let": {"breakdown": {"type": {"type": "map", "values": "double"}, "new":
{"bad": {"/": [{"a.count": ["scores", {"string": "bad"}]}, {"a.len": ["scores"]}]},
"good": {"/": [{"a.count": ["scores", {"string": "good"}]}, {"a.len": ["scores"]}]}}}}},
{"type": {"type": "record", "fields": [{"name": "winner", "type": "string"}, {"name":
"breakdown", "type": {"type": "map", "values": "double"}}], "name": "Output"}, "new":
{"winner": "winner", "breakdown": "breakdown"}}], "fcns": {"imputeX2": {"params":
[{"possiblyNull": ["null", "double"]}], "ret": "double", "do": [{"ifnotnull": {"x":
"possiblyNull"}, "then": {"do": [{"cell": "runningAverage", "to": {"params": [{"old":
"RunningAverage"}], "ret": "RunningAverage", "do": [{"stat.sample.updateEWMA": ["x", 0.1,

Technically, it’s human-readable, but it isn’t organized well enough to be practical. Fortunately, Titus provides some tools for navigating PFA and making sense of it.

Before you begin…

Download and install Titus. This article was tested with Titus 0.8.3; newer versions should work with no modification. Python >= 2.6 and < 3.0 is required.

Launch a Python prompt and import json and titus.producer.tools:

Python 2.7.6
Type "help", "copyright", "credits" or "license" for more information.
>>> import json
>>> import titus.producer.tools as t

and download myModel.pfa. Also, verify that you can run the pfainspector script (it gets installed in your executable path when you install Titus).

Direct inspection in Python

PFA documents are pure JSON, so you can load them with Python’s json module.

pfaDocument = json.load(open("myModel.pfa"))

This loads the JSON text into Python objects with the following substitutions:

JSON type Example Python object Example
null null None None
boolean true, false bool True, False
number 3, 3.14 int, long, float 3, 3.14
string "hello" str 'hello'
array [1, 2, 3] list [1, 2, 3]
object {"one": 1, "two": 2} dict {'one': 1, 'two': 2}

You can navigate these types using ordinary Python square brackets, like this:

>>> pfaDocument["cells"]["forest"]["init"][29]["pass"] \
...                ["TreeNode"]["fail"]["TreeNode"]["value"]
{u'array': [u'three']}

or modify them like this:

>>> pfaDocument["cells"]["forest"]["init"][29]["pass"] \
...                ["TreeNode"]["fail"]["TreeNode"]["value"] = \
...                    {"array": ["three", "four"]}

but that still isn’t particularly convenient for deep objects. Also, finding a path involves many interactive steps, asking for the .keys() or len() at each level, to avoid flooding the output with huge PFA fragments.

The titus.producer.tools module (loaded as t here) has a get method that treats Python’s JSON representation as a tree that can be traversed with a path from root to any element. The above example becomes

>>> t.get(pfaDocument, ("cells", "forest", "init", 29, "pass", "TreeNode", \
                        "fail", "TreeNode", "value"))

or even

>>> t.get(pfaDocument, "cells,forest,init,29,pass,TreeNode,fail,TreeNode,value")

There is a corresponding assign for assignment.

t.assign(pfaDocument, "cells,forest,init,29,pass,TreeNode,fail,TreeNode,value", \
             {"array": ["three", "four"]})

Finding the relevant path is still an issue, so Titus has a look function for browsing.

>>> t.look(pfaDocument)
index                          data
action                           "action": [
action,0                           {
action,0,let                         "let": {
action,0,let,derived                   "derived": {
action,0,let,derived,new                 "new": {
action,0,let,derived,new,x2                "x2": {
action,0,let,derived,new,x2...               "u.imputeX2": [
action,0,let,derived,new,x2...                 {...}
action,0,let,derived,new,x1                "x1": {
action,0,let,derived,new,x1...               "s.lower": [
action,0,let,derived,new,x1...                 {...}
action,0,let,derived,type                "type": {
action,0,let,derived,type,f...             "fields": [{"type": "string", "name": "x1"}, {"type": "double", "name": "x2"}],
action,0,let,derived,type,type             "type": "record",
action,0,let,derived,type,name             "name": "Derived"
action,1                           {
action,1,let                         "let": {

This function gives tree-indexes on the left and content on the right, pretty-printed such that two levels of depth are inline, six levels of depth are exploded, and anything deeper is hidden in ellipsis (...). All of these parameters can be controlled.

>>> help(t.look)
Help on function look in module titus.producer.tools:

look(expr, maxDepth=8, inlineDepth=2, indexWidth=30, dropAt=True, stream=<open file '<stdout>', mode 'w'>)
    Print a JSON object on the screen in a readable way.
    maxDepth: maximum depth to show before printing ellipsis (...)
    inlineDepth: maximum depth to show on a single line
    indexWidth: width (in characters) of the index column on the left
    dropAt: don't show "@" keys
    stream: allows the output to be sent to a file or stream.

For very deep objects, it’s often useful to call look once to find part of the path, copy it into a get, and then call look again to go deeper.

>>> t.look(t.get(pfaDocument, "cells,forest,init,29,pass"))
index                          data
TreeNode                         "TreeNode": {
TreeNode,operator                  "operator": "<=",
TreeNode,field                     "field": "x2",
TreeNode,fail                      "fail": {
TreeNode,fail,TreeNode               "TreeNode": {
TreeNode,fail,TreeNode,oper...         "operator": "in",
TreeNode,fail,TreeNode,field           "field": "x1",
TreeNode,fail,TreeNode,fail            "fail": {"string": "bad"},
TreeNode,fail,TreeNode,value           "value": {"array": ["three", "four"]},
TreeNode,fail,TreeNode,pass            "pass": {"string": "good"}
TreeNode,value                     "value": {"double": 7.26247432127833},
TreeNode,pass                      "pass": {
TreeNode,pass,TreeNode               "TreeNode": {
TreeNode,pass,TreeNode,oper...         "operator": "in",
TreeNode,pass,TreeNode,field           "field": "x1",
TreeNode,pass,TreeNode,fail            "fail": {"string": "bad"},
TreeNode,pass,TreeNode,value           "value": {"array": ["three"]},
TreeNode,pass,TreeNode,pass            "pass": {"string": "good"}

These tools are intended for interactive exploration of a PFA document. To find substructures in a way that doesn’t depend on index (to be more robust against changes in the PFA document), use JSON regular expressions, the next topic.