Skip to main content

# Slipform

Project description

🏗 slipform

pythonflow decorator for generating dataflow graphs from raw python.


  • Syntax is natural, you can use a simple decorator to obtain the dataflow graph. No need to rewrite your code for pythonflow.

  • Slipform allows you to write and test code as you normally would, debugging it using the debugger of your choice during runtime. When you are happy with your code, finally, at runtime you can generate the dataflow graph.


Slipform was born out of a desire to learn more about the python AST, and potentially use it for my own personal projects if it works out.

It is not actively developed, nor should it be considered stable.



  • naming from assignments
  • placeholders from args
  • constant support
  • functions to operations
  • import support
  • custom operation support
  • ignore comments #[ignore] or #[slipform:ignore]


  • module / import detection from function scope
  • sequences (map, list, tuple, zip, sum, filter)
  • for loop replacement?
  • conditional expression replacement?
  • assertion replacement?
  • try/catch replacement?
  • explicit dependencies?

Examples based on Using Pythonflow

  1. Get started by importing slipform
from slipform import slipform
# "pf" must be part of scope of any @slipfrom annotated
# function. This limitation will be relaxed in future.
import pythonflow as pf
  1. A simple example is as follows:
def add_graph(x):
  a = 5
  b = 32
  z = a + b + x

With the equivalent pythonflow version:

with pf.Graph() as add_graph:
    x = pf.placeholder('x')
    a = pf.constant(5, name='a')
    b = pf.constant(32, name='b')
    z = (a + b + x).set_name('z')

We can evaluate the graphs like usual using pythonflow:

add_graph(['b', 'z'], x=5)
>>> (32, 42)
  1. A more complicated example
def vae(x, x_target, encoder, decoder, mse):
    # import ... from ... (as ...) are all supported
    import torch
    import torch.nn.functional as F
    # get the encoding!
    z_params = encoder(x)
    # deterministic, we dont reparameterize here
    z = z_params.z_mean
    # reconstruct here
    x_pre_recon = decoder(z)
    # final activation
    x_recon = x_pre_recon if mse else torch.sigmoid(x_pre_recon)
    # compute loss
    loss = F.mse_loss(x_recon, x_target) if mse else F.binary_cross_entropy_with_logits(x_pre_recon, x_target)

The above will generate code equivalent in functionality to:

with pf.Graph() as add_graph:
    x        = pf.placeholder('x')
    x_target = pf.placeholder('x_target')
    encoder  = pf.placeholder('encoder')
    decoder  = pf.placeholder('decoder')
    mse      = pf.placeholder('mse')
    # import everything
    torch = pf.import_('torch')
    F = pf.import_('torch.nn.functional')
    # get the encoding!
    z_params = encoder(x)
    # deterministic, we dont reparameterize here
    z = z_params.z_mean
    # reconstruct here
    x_pre_recon = decoder(z)
    # final activation
    x_recon = pf.conditional(mse, x_pre_recon, torch.sigmoid(x_pre_recon))
    # compute loss
    loss = pf.conditional(mse, F.mse_loss(x_recon, x_target), F.binary_cross_entropy_with_logits(x_pre_recon, x_target))

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

slipform-0.0.1a2.tar.gz (13.9 kB view hashes)

Uploaded source

Built Distribution

slipform-0.0.1a2-py2.py3-none-any.whl (7.9 kB view hashes)

Uploaded py2 py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page