jaberg / pyll Goto Github PK

The AST nodes should be invariant to the calling convention (positional vs. named) used in the function calls that built the nodes.

args should be a dictionary mapping paramname -> argval

e.g. the replacement of

genson_template.sample(seed)

?

Right now,

 stochastic.sample(template, np.random.RandomState(seed))

does NOT yield the same result on each call with a fixed seed, e.g. 0. Is it supposed to? If so, I'll submit a test

In [8]: stochastic.sample(L.template, np.random)

UnboundLocalError Traceback (most recent call last)
/Users/dyamins/eccv12/ in ()
----> 1 stochastic.sample(L.template, np.random)

/Users/dyamins/eccv12/pyll/pyll/stochastic.pyc in sample(expr, rng)
199
200 def sample(expr, rng):
--> 201 foo, newrng = replace_implicit_stochastic_nodes(expr, as_apply(rng))
202 return rec_eval(foo)
203

/Users/dyamins/eccv12/pyll/pyll/stochastic.pyc in replace_implicit_stochastic_nodes(expr, rng, scope)
163 expr = draw
164 lrng = new_lrng
--> 165 return expr, new_lrng
166
167

UnboundLocalError: local variable 'new_lrng' referenced before assignment

This is very high priority if pyll is be used as a way to write lazy functional code.

e.g. the old internal references from genson?

Hey James,

I'd like to package pyll (and later pyautodiff) for Gentoo and having a alpha version on PyPi would help. Would you be interested in pushing 0.0.1 releases of some of these packages on PyPi? It will also make requirements.txt files simpler ;-)

Thanks!

n

This is more of a "how to use pyll" question than an "issue":

Suppose I have a pyll object, e.g. a model spec generator. Suppose I want to programmatically replace one value in it with something determined by a function from some inputs. E.g. I want to set a key "num_filters" = pyll.scope.choice(function(some_inputs)), where the "num_filters" key is a node that appears somewhere down the object tree. I suspect I need to use replace_input? But the exact way to construct the arguments escapes me. Or do I need to do something else?

Similarly, suppose I want to add a key?

I guess I'm in general asking: how do I do simple transformations on pyll ASTs of this kind? If you could point me to the relevant place to figure it out, that would be great.

currently pyll does not play well with reload python modules ... you get lots of 'Cannot override existing symbol' errors

can this be improved?

Crude way to reproduce:

In [36]: pyll.stochastic.sample({1:[s.uniform(0, 1), s.one_of(0, 1, 2)]}, np.random)
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
/home/npinto/<ipython-input-36-6a0b951cbcec> in <module>()
----> 1 pyll.stochastic.sample({1:[s.uniform(0, 1), s.one_of(0, 1, 2)]}, np.random)

/home/npinto/venv/pyll-system/pyll/pyll/stochastic.py in sample(expr, rng)
    154 
    155 def sample(expr, rng):
--> 156     foo = recursive_set_rng_kwarg(clone(as_apply(expr)), as_apply(rng))
    157     return rec_eval(foo)
    158 

/home/npinto/venv/pyll-system/pyll/pyll/base.py in as_apply(obj)
    132         items.sort()
    133         named_args = [(k, as_apply(v)) for (k, v) in items]
--> 134         rval = Apply('dict', [], named_args, len(named_args))
    135     else:
    136         rval = Literal(obj)

/home/npinto/venv/pyll-system/pyll/pyll/base.py in __init__(self, name, pos_args, named_args, o_len)
    154         assert all(isinstance(v, Apply) for v in pos_args)
    155         assert all(isinstance(v, Apply) for k, v in named_args)
--> 156         assert all(isinstance(k, basestring) for k, v in named_args)

There is currently a line mentioning self that cannot possibly be right. Test this function.

For users of genson (such as myself), ... pyll is currently not a complete working replacement, so code that depends heavily on genson and hyperopt must stil use old versions of all that.

It is critical for moving to pyll (which I think would be a good idea sooner rather than later) that pyll support genson completely, with a simple drop-in replacement. Genson have a good interface for a variety of tasks -- especially sampling -- which pyll seems less suited to. These are important tasks for our purposes, I think.

I will assume that I will assign myself to this issue ...