abergeron/myia_debug

Debugging tools for Myia

myia_debug

Utility to debug Myia.

You can call this with <buche> python -m myia_debug <args>, or alias python -m myia_debug to a shortcut like dm.

Debug Myia

Usage:
  dm <command>
     [-f FUNCTION...] [-a ARG...] [-g...]
     [-t TYPE...]
     [--shapes SHAPE...]
     [-O OPT...]
     [--config FILE...]
     [--pipeline PIP...]
     [--scalar]
     [--no-beautify]
     [--function-nodes]
     [<rest>...]

Options:
  -f --fn FUNCTION...   The function to run.
  -a --args ARG...      Arguments to feed to the function.
  -g                    Apply gradient once for each occurrence of the flag.
  -t --types TYPE...    Types of the arguments.
  --shapes SHAPE...     Shapes of the arguments.
  -c --config FILE...   Use given configuration.
  -O --opt OPT...       Run given optimizations.
  -p --pipeline PIP...  The pipeline to use.
  --scalar              Use the scalar pipeline.
  --no-beautify         Don't beautify graphs.
  --function-nodes      Show individual nodes for functions called.

This requires buche to be installed: https://github.com/breuleux/buche/releases

The value for most arguments and options can be of the form module:variable, in which case the corresponding module is imported and the variable is read from it. Multiple values for arguments or types are separated by either , or ;.

The -g flag doesn't do anything right now.

Examples

Suppose x.py contains the following code:

import numpy

x1 = 4
x2 = 8.5
x3 = numpy.random.random((4, 6))

def hibou(x, y):
    return x + y

Running the function on test inputs:

dm run x:hibou --args 4,7
dm run x:hibou --args x:x1,x:x2
dm run x:hibou --args x:x2,x:x3

Showing the graph.

buche dm show x:hibou
buche dm show x:hibou --pipeline parse,resolve,infer,specialize,opt --types i64,i64
buche dm show x:hibou --pipeline '!opt' --types i64,i64
buche dm show x:hibou --pipeline '!opt' --args x:x2,x:x3
buche dm show x:hibou --pipeline '!opt' --types i64,i64 -O inline

The !opt pipeline is the same as parse,resolve,infer,specialize,opt. There is also a cconv step.

Custom operations

Suppose y.py contains the following code:

from myia_debug import steps

def inferred(o):
    res = o.run(default=[steps.parse,
                         steps.infer,
                         steps.specialize])
    if 'error' in res:
        raise res['error']
    print(res['inference_results'])

Then you can run this, and it will print the inference results:

dm y:inferred x:hibou --args x:x2,x:x3

The call to o.run handles most parameters for you, so you can provide types, shapes, and customize the pipeline.