Awesome

syntaxis

Note: this tool used to be called saplings but has sinced been renamed to syntaxis. Saplings now refers to an LLM agent library.

syntaxis is a static analysis tool for Python. Given a program, syntaxis will build object hierarchies for every module imported in the program. Object hierarchies are dependency trees where the root node represents a module and each child represents an attribute of its parent. These can be useful for making inferences about a module's API, mining patterns in how a module is used, and duck typing.

Installation

Requires Python 3.X.

You can install syntaxis with pip:

$ pip install syntaxis

Usage

Using syntaxis takes only two steps. First, convert your input program into an Abstract Syntax Tree (AST) using the ast module. Then, import the Syntaxis object and initialize it with the root node of the AST.

import ast
from syntaxis import Syntaxis

my_program = open("path_to_your_program.py", "r").read()
program_ast = ast.parse(my_program)
my_syntaxis = Syntaxis(program_ast)

That's it. To access the object hierarchies, simply call the get_trees method in your Syntaxis object, like so:

my_syntaxis.get_trees() # => [ObjectNode(), ObjectNode(), ..., ObjectNode()]

For more advanced usage of the Syntaxis object, read the docstring here.

Printing an Object Hierarchy

get_trees returns a list of ObjectNodes, each representing the root node of an object hierarchy and which has the following attributes:

• name (str): Name of the object
• is_callable (bool): Whether the object is callable (i.e. has __call__ defined)
• order (int): Indicates the type of connection to the parent node (e.g. 0 is an attribute of the parent, 1 is an attribute of the output of the parent when called, etc.); -1 if node is root
• frequency (int): Number of times the object is used in the program
• children (list): List of child nodes

To pretty-print a tree, simply pass its root node into the render_tree generator, like so:

from syntaxis import render_tree

trees = my_syntaxis.get_trees()
root_node = trees[0]
for branches, node in render_tree(root_node):
  print(f"{branches}{node}")

numpy (NC, -1)
 +-- random (NC, 0)
 |   +-- randn (C, 0)
 |       +-- __sub__ (C, 1)
 |       |   +-- shape (NC, 1)
 |       |   +-- __index__ (C, 1)
 |       +-- sum (C, 1)
 +-- matmul (C, 0)
 +-- expand_dims (C, 0)
     +-- T (NC, 1)

(Here, NC means indicates a non-callable node and C a callable node. -1/0/1 indicate the order of the node's connection to its parent).

To create a dictionary representation of a tree, pass its root node into the dictify_tree function, like so:

from syntaxis import dictify_tree

dictify_tree(root_node)

{
  "numpy": {
    "is_callable": False,
    "order": -1,
    "frequency": 1,
    "children": [
      {"random": ...},
      {"matmul": ...},
      {"expand_dims": ...}
    ]
  }
}

Interpreting the Object Hierarchy

Each node is an object and an object can either be callable (i.e. has __call__ defined) or non-callable. Links between nodes each have an order –– a number which describes the relationship between a node and its parent. If a node is a 0th-order child of its parent object, then it's an attribute of that object. If it's a 1st-order child, then it's an attribute of the output of the parent object when it's called, and so on. For example:

my_parent = module.my_obj

my_parent.attr # attr is a 0th-order child of my_obj
my_parent().attr # attr is a 1st-order child of my_obj
my_parent()().attr # attr is a 2nd-order child of my_obj

What counts as a function?

In Python, subscripts, comparisons, and binary operations are all just syntactic sugar for function calls, and are treated by syntaxis as such. Here are some common "translations:"

my_obj['my_sub'] # => my_obj.__index__('my_sub')
my_obj + 10 # => my_obj.__add__(10)
my_obj == None # => my_obj.__eq__(None)

Limitations

Syntaxis statically analyzes the usage of a module in a program, meaning it doesn't actually execute any code. Instead, it traverses the program's AST and tracks "object flow," i.e. how an object is passed through a program via variable assignments and calls of user-defined functions and classes. To demonstrate this idea, consider this example of currying and the tree syntaxis produces:

import torch

def compose(g, f):
  def h(x):
    return g(f(x))

  return h

def F(x):
  return x.T

def G(x):
  return x.sum()

composed_func = compose(F, G)
composed_func(torch.tensor())

Syntaxis identifies tensor as an attribute of torch, then follows the object as it's passed into composed_func. Because syntaxis has an understanding of how composed_func is defined, it can analyze the object flow within the function and capture the T and sum sub-attributes.

While syntaxis can track object flow through many complex paths in a program, I haven't tested every edge case, and there are some situations where syntaxis produces inaccurate trees. Below is a list of all the failure modes I'm aware of (and currently working on fixing). If you discover a bug or missing feature that isn't listed here, please create an issue for it.

Data Structures

As of right now, syntaxis can't track assignments to comprehensions, generator expressions, dictionaries, lists, tuples, or sets. It can, however, track object flow inside these data structures. For example, consider the following:

import numpy as np

vectors = [np.array([0]), np.array([1]), np.array([2])]
vectors[0].mean()

Syntaxis can capture array and add it to the numpy object hierarchy, but it cannot capture mean, and thus produces the following tree:

This limitation can have some unexpected consequences. For example, functions that return multiple values with one return statement (e.g. return a, b, c) are actually returning tuples. Therefore, the output of those functions won't be tracked by syntaxis. The same logic applies to variable unpacking with * and **.

Control Flow

Handling control flow is tricky. Tracking object flow in loops and conditionals requires making assumptions about what code actually executes. For example, consider the following:

import numpy as np

for x in np.array([]):
  print(x.mean())

Because syntaxis only does static analysis and doesn't do type inference, it doesn't know that np.array([]) is an empty list, and that therefore the loop never executes. In this situation, capturing mean and adding the __index__ -> mean subtree to numpy -> array would be a false positive, since x (i.e. the output of np.array().__index__()) is never defined. To handle this, syntaxis should branch out and produce two possible trees for this module –– one that assumes the loop doesn't execute, and one that assumes it does:

But as of right now, syntaxis will only produce the tree on the right –– that is, we assume the bodies of for loops are always executed (because they usually are).

Below are the assumptions syntaxis makes for other control flow elements.

while loops

while loops are processed under the same assumption as for loops –– that is, the body of the loop is assumed to execute.

if/else blocks

Syntaxis processes if and else blocks more conservatively than loops. It tracks object flow within these blocks but doesn't allow changes to the namespace to persist into the parent scope. For example, given:

import numpy as np

X = np.array([1, 2, 3])

if condition:
  X = np.matrix([1, 2, 3])
else:
  print(X.mean())
  X = None
  Y = np.array([1, 2, 3])

print(X.sum())
print(Y.max())

syntaxis will produce the following tree:

Notice how the value of X is unreliable since we don't know if condition is True or False. To handle this, syntaxis simply stops tracking any variable that's defined in the outer scope, like X, if it's modified inside an if/else block. Similarly, notice how there exists an execution path where Y is never defined and Y.max() throws an error. To handle this, syntaxis assumes that any variable defined inside an if/else block, such as Y, doesn't persist into the outer scope.

Both of these assumptions are made in attempt to reduce false positives and false negatives. But ideally, syntaxis would branch out and produce two separate trees for this module –– one that assumes the if block executes and another that assumes the else block executes, like so:

try/except blocks

try/except blocks are handled similarly to if/else blocks –– that is, changes to the namespace made in either block do not persist in the outer scope.

Notably, try and else blocks are treated as a single block, since else is only executed if try executes without exceptions. And finally blocks are treated as separate from the control flow, since code in here always executes regardless of whether an exception is thrown.

return, break, and continue statements

All code underneath a return, break, or continue statement is assumed not to execute and will not be analyzed. For example, consider this:

import numpy as np

for x in range(10):
  y = np.array([x])
  continue
  y.mean()

It may be the case that mean is actually an attribute of np.array, but syntaxis will not capture this since y.mean() is never executed.

Notably, syntaxis doesn't apply this assumption to statements inside control flow blocks. For example, if the continue statement above was changed to:

if condition:
  continue

Then mean would be captured by syntaxis as an attribute of np.array.

Functions

Recursion

Syntaxis cannot process recursive function calls. Consider the following example:

import some_module

def my_recursive_func(input):
  if input > 5:
    return my_recursive_func(input - 1)
  elif input > 1:
    return some_module.foo
  else:
    return some_module.bar

output = my_recursive_func(5)
output.attr()

We know this function returns some_module.foo, but syntaxis cannot tell which base case is hit, and therefore can't track the output. To avoid false positives, we assume this function returns nothing, and thus attr will not be captured and added to the object hierarchy. The tree syntaxis produces is:

Generators

Generators aren't processed as iterables. Instead, syntaxis ignores yield/yield from statements and treats the generator like a normal function. For example, given:

import some_module

def my_generator():
  yield from some_module.some_items

for item in my_generator():
  print(item.name)

__index__ -> name won't be added as a subtree to some_module -> some_items, and so the tree produced by syntaxis will look like this:

Notably, this limitation will only produce false negatives –– not false positives.

Anonymous Functions

While the bodies of anonymous (lambda) functions are processed, object flow through assignments and calls of those functions is not tracked. For example, given:

import numpy as np

trans_diag = lambda x: np.diagonal(x.T)
trans_diag(np.random.randn(5, 5))

syntaxis will produce the following tree:

Notice that T is not captured as an attribute of numpy.random.randn, but diagonal is captured as an attribute of numpy. This is because the body of the lambda function is processed by syntaxis, but the assignment to trans_diag is not recognized, and therefore the function call is not processed.

Classes

Syntaxis can track object flow in static, class, and instance methods, getter and setter methods, class and instance variables, classes defined within classes, and class closures (i.e. functions that return classes). Notably, it can keep track of the state of each instance of a user-defined class. Consider the following program and the tree syntaxis produces:

import torch.nn as nn
from torch import tensor

class Perceptron(nn.Module):
  loss = None

  def __init__(self, in_channels, out_channels):
    super(Perceptron, self).__init__()
    self.layer = nn.Linear(in_channels, out_channels)
    self.output = Perceptron.create_output_layer()

  @staticmethod
  def create_output_layer():
    def layer(x):
      return x.mean()

    return layer

  @classmethod
  def calculate_loss(cls, output, target):
    cls.loss = output - target
    return cls.loss

  def __call__(self, x):
    x = self.layer(x)
    return self.output(x)

model = Perceptron(1, 8)
output = model(tensor([10]))
loss = Perceptron.calculate_loss(output, 8)

While syntaxis can handle many common usage patterns for user-defined classes, such as the ones above, there are some things syntaxis can't handle yet. Below are all the limitations I'm aware of:

Class Modifications

In the example above, calling the class method Perceptron.calculate_loss should change the value of the class variable loss. However, syntaxis cannot track modifications to a class when it's passed into a function. Syntaxis can handle when a class is modified in the scope in which it was defined, like so:

Perceptron.loss = tensor()
Perceptron.loss.item()

Here, item would be captured and added to the tree as an attribute of tensor. But if the class is modified via an alias, like so:

NeuralNet = Perceptron
NeuralNet.loss = tensor()
Perceptron.loss.item()

Then syntaxis won't capture item. Syntaxis also can't propagate class modifications to existing instances of the class. For example, continuing the code above:

model = Perceptron(1, 8)
Perceptron.loss = tensor()
model.loss.item()

Because the change to loss, a class variable, won't propagate to model, an instance of Perceptron, item won't be captured as an attribute of tensor.

Inheritance

Syntaxis cannot recognize inherited methods or variables in user-defined classes. For example, given:

import some_module

class MyClass(module.Foo):
  def __init__(self, x):
    self.bar(x)

syntaxis will not recognize bar as an attribute of module.Foo, despite bar being an inherited method. This limitation also holds true when the base class is user-defined.

Metaclasses

Once I learn what metaclasses actually are and how to use them, I'll get around to handling them in syntaxis. But for now this is on the bottom of my to-do list since 99.9% of Python users also don't know what the hell metaclasses are.

Miscellaneous

global and nonlocal statements

global statement are used inside functions to declare a variable to be in the global namespace. But syntaxis doesn't recognize these statements and change the namespace accordingly. For example, given:

import some_module

my_var = some_module.foo

def my_func():
  global my_var
  my_var = None

my_func()
my_var.bar()

syntaxis will produce a tree with bar as an attribute of foo. This would be a false positive since calling my_func sets my_var to None, and of course None doesn't have bar as an attribute.

nonlocal statements are similar to globals, except they allow you to modify variables declared in outer scopes. And like globals, syntaxis doesn't recognize nonlocal statements.

Built-in functions

None of Python's built-in functions are recognized by syntaxis. For example, consider the enumerate function:

import some_module

for index, item in enumerate(some_module.items):
  print(item.some_attr)

syntaxis won't capture attr as an attribute of some_module.items.__iter__, which it would have if some_module.items wasn't wrapped by enumerate.