rizar

class UnDef():
    pass

UNDEF = UnDef()

class Block(object):

    def __getattribute__(self, name):
        value = object.__getattribute__(self, name)
        if value == UNDEF:

rizar

# INTRODUCTION
# ------------
#
# Bart's logistic regression oneliner has served us well
# forcing us to think how to make the framework 
# suitable for quick-and-dirty experimentation with small 
# standard models. It did not force us to think about 
# scalability and flexibility necessary for big
# experiments with big models.
#

rizar

import theano
from theano.tensor.shared_randomstreams import RandomStreams

r = RandomStreams(1)

y, u = theano.scan(lambda : r.binomial(size=(2, 2)), n_steps=3)
f = theano.function([], [y])

# It still gives me three different sequence of random matrices!
print f()

rizar

............/home/rizar/Dist/theano/theano/scan_module/scan_perform_ext.py:133: RuntimeWarning: numpy.ndarray size changed, may indicate binary incompatibility
  from scan_perform.scan_perform import *
...ERROR (theano.gof.opt): SeqOptimizer apply <theano.gof.opt.EquilibriumOptimizer object at 0x4ac9c50>
ERROR (theano.gof.opt): Traceback:
ERROR (theano.gof.opt): Traceback (most recent call last):
  File "/home/rizar/Dist/theano/theano/gof/opt.py", line 195, in apply
    sub_prof = optimizer.optimize(fgraph)
  File "/home/rizar/Dist/theano/theano/gof/opt.py", line 81, in optimize
    ret = self.apply(fgraph, *args, **kwargs)
  File "/home/rizar/Dist/theano/theano/gof/opt.py", line 1693, in apply

rizar

import numpy

import theano
from theano import tensor
from theano.printing import debugprint

floatX = theano.config.floatX
foo_a, foo_x, r = 3 * [None]

def foo(x, a):

rizar

class Data(object):
   """Super abstract class for all data sources. 
   
   So far does not have neither methods nor attributes, 
   but perhaps it can provide some sort-of signature, 
   e.g. channel names.
   
   """
   pass

rizar

class RecurrentWithFork(Initializable):

    @lazy(allocation=['input_dim'])
    def __init__(self, recurrent, input_dim, **kwargs):
        super(RecurrentWithFork, self).__init__(**kwargs)
        self.recurrent = recurrent
        self.input_dim = input_dim
        self.fork = Fork(
            [name for name in self.recurrent.sequences
             if name != 'mask'],

rizar

# This to illustrate the idea how deep BiRNN will be implemented in Blocks.

class RecurrentWithFork(Initializable):

    @lazy(allocation=['input_dim'])
    def __init__(self, recurrent, input_dim, **kwargs):
        super(RecurrentWithFork, self).__init__(**kwargs)
        self.recurrent = recurrent
        self.input_dim = input_dim
        self.fork = Fork(

rizar

seen = set()                                        
def strip(obj):                                     
    if id(obj) in seen:                             
        return                                      
    seen.add(id(obj))                               
    if isinstance(obj, theano.Variable):            
        obj.tag.annotations = []                    
        obj.tag.roles = []                          
    if isinstance(obj, (list, tuple, set)):         
        for el in obj:                              

rizar

diff --git a/reverse_words/__init__.py b/reverse_words/__init__.py
index b649ab5..ac296e7 100644
--- a/reverse_words/__init__.py
+++ b/reverse_words/__init__.py
@@ -14,7 +14,7 @@ from theano import tensor
 from blocks.bricks import Tanh, Initializable
 from blocks.bricks.base import application
 from blocks.bricks.lookup import LookupTable
-from blocks.bricks.recurrent import SimpleRecurrent, Bidirectional
+from blocks.bricks.recurrent import SimpleRecurrent, LSTM, Bidirectional