* All theano tests should now have a deterministic seed

theano/compile/tests/test_inplace_opt_for_value.py

@@ -5,6 +5,7 @@ from theano.gof import OpSub, TopoOptimizer
 
 from pylearn.algorithms.minimizer import make_minimizer # minimizer
 from theano.printing import Print
+from theano.tests import unittest_tools
 #import sgd #until Olivier's module-import thing works better
 
 ####################
@@ -177,7 +178,7 @@ class ExampleRNN(Module):
     def _instance_initialize(self, obj):
         n_vis = self.n_vis
 
-        rng = N.random.RandomState(2342)
+        rng = N.random.RandomState(unittest_tools.fetch_seed(2342))
 
         obj.z0 = N.zeros(n_vis)
         obj.w = rng.randn(n_vis, n_vis) * 0.01
@@ -193,7 +194,7 @@ def test_example_rnn():
 
     rnn = rnn_module.make(mode='FAST_RUN')
 
-    rng = N.random.RandomState(7722342)
+    rng = N.random.RandomState(unittest_tools.fetch_seed(7722342))
     x = rng.randn(10,n_vis)
     y = rng.randn(10,n_out)
 
@@ -215,7 +216,7 @@ def test_example_rnn():
 def test_WEIRD_STUFF():
     n_vis = 3
 
-    rng = N.random.RandomState(7722342)
+    rng = N.random.RandomState(unittest_tools.fetch_seed(7722342))
     x = rng.randn(10,n_vis)
     y = rng.randn(10,n_vis)
 

theano/sparse/tests/test_basic.py

@@ -10,6 +10,7 @@ from theano import gof
 
 from theano.sparse.basic import _is_dense, _is_sparse, _is_dense_result, _is_sparse_result
 from theano.sparse.basic import _mtypes, _mtype_to_str
+from theano.tests import unittest_tools
 
 
 def eval_outputs(outputs):
@@ -17,7 +18,8 @@ def eval_outputs(outputs):
 
 class T_transpose(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(44)
+        unittest_tools.seed_rng()
+
     def test_transpose_csc(self):
         sp = sparse.csc_matrix(sparse.eye(5,3))
         a = as_sparse(sp)
@@ -123,7 +125,7 @@ class T_Add(unittest.TestCase):
 
 class T_conversion(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(44)
+        unittest_tools.seed_rng()
 
     def test0(self):
         a = tensor.as_tensor(numpy.random.rand(5))

theano/tensor/deprecated/rmodule.py

@@ -6,6 +6,7 @@ from functools import partial
 from collections import deque
 
 import numpy
+from copy import copy
 
 from ...compile import (SymbolicInputKit, SymbolicInput, 
         Module, module, Method, Member, In, Component)
@@ -32,7 +33,7 @@ class KitComponent(Component):
             r = input.result
             if r not in memo:
                 input = copy(input)
-                input.value = gof.Container(r, storage = [None])
+                input.value = Container(r, storage = [None])
                 memo[r] = input
 
     def build(self, mode, memo):

theano/tensor/tests/test_basic.py

@@ -16,7 +16,10 @@ from theano import gof
 from theano.tensor.elemwise import DimShuffle
 from theano.compile.mode import default_mode
 from theano import function
+from theano.tests import unittest_tools
 
+### seed random number generator so that unittests are deterministic ###
+unittest_tools.seed_rng()
 
 def inplace_func(inputs, outputs, mode=default_mode):
     return function(inputs, outputs, mode=mode, accept_inplace=True)
@@ -604,7 +607,7 @@ class T_Cast(unittest.TestCase):
 
 class T_max_and_argmax(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(123784)
+        unittest_tools.seed_rng()
         MaxAndArgmax.debug = 0
 
     def test0(self):
@@ -620,13 +623,17 @@ class T_max_and_argmax(unittest.TestCase):
         self.failUnless(i == 2)
 
     def test2(self):
-        n = as_tensor(numpy.random.rand(2,3))
+        data = numpy.random.rand(2,3)
+        n = as_tensor(data)
         v,i = eval_outputs(max_and_argmax(n))
-        self.failUnless(numpy.all(i == [0,1]))
+        self.failUnless(numpy.all(v == numpy.max(data,-1)))
+        self.failUnless(numpy.all(i == numpy.argmax(data,-1)))
     def test2b(self):
-        n = as_tensor(numpy.random.rand(2,3))
+        data = numpy.random.rand(2,3)
+        n = as_tensor(data)
         v,i = eval_outputs(max_and_argmax(n,0))
-        self.failUnless(numpy.all(i == [0,1,1]))
+        self.failUnless(numpy.all(v == numpy.max(data,0)))
+        self.failUnless(numpy.all(i == numpy.argmax(data,0)))
     def test2_invalid(self):
         n = as_tensor(numpy.random.rand(2,3))
         try:
@@ -663,7 +670,7 @@ class T_max_and_argmax(unittest.TestCase):
 class T_subtensor(unittest.TestCase):
     def setUp(self):
         Subtensor.debug = False
-        numpy.random.seed(12353123)
+        unittest_tools.seed_rng()
 
     def test0_err_invalid(self):
         #it is impossible to retrieve a view of a 0-d tensor
@@ -1063,6 +1070,9 @@ class test_bitwise(unittest.TestCase):
 
 class T_add(unittest.TestCase):
 
+    def setUp(self):
+        unittest_tools.seed_rng()
+
     def test_complex_all_ops(self):
         for nbits in (64, 128):
             a = value(numpy.ones(3, dtype='complex%i' % nbits)+0.5j)
@@ -1158,7 +1168,7 @@ class T_exp(unittest.TestCase):
 
 # class T_mul(unittest.TestCase):
 #     def setUp(self):
-#         numpy.random.seed([1,2,3,4])
+#         unittest_tools.seed_rng()
 
 #     def test_elemwise(self):
 #         a = as_tensor(0.0)
@@ -1221,7 +1231,7 @@ class T_exp(unittest.TestCase):
 
 # class T_div(unittest.TestCase):
 #     def setUp(self):
-#         numpy.random.seed(9999)
+#         unittest_tools.seed_rng()
 #     def test_grad_e(self):
 #         verify_grad(self, Div, [numpy.random.rand(3), numpy.ones(3)])
 #         verify_grad(self, Div, [numpy.random.rand(3,5), numpy.random.rand(3,5)+0.1])
@@ -1233,10 +1243,14 @@ class T_exp(unittest.TestCase):
 #         verify_grad(self, Div, [numpy.random.rand(3,5), numpy.random.rand(1,1)])
 
 # class T_log2(unittest.TestCase):
+#     def setUp(self):
+#         unittest_tools.seed_rng()
 #     def test0(self):
 #         verify_grad(self, Log2, [numpy.random.rand(3,1)+0.0001])
 
 # class T_log(unittest.TestCase):
+#     def setUp(self):
+#         unittest_tools.seed_rng()
 #     def test0(self):
 #         verify_grad(self, Log, [numpy.random.rand(3,1)+0.0001])
 #     def test1(self):
@@ -1248,7 +1262,7 @@ class T_exp(unittest.TestCase):
 
 # class T_pow(unittest.TestCase):
 #     def setUp(self):
-#         numpy.random.seed(9999)
+#         unittest_tools.seed_rng()
 #     def test_elemwise(self):
 #         verify_grad(self, Div, [numpy.random.rand(3,4), numpy.random.rand(3,4)+0.1])
 #         verify_grad(self, Pow, [numpy.random.rand(3,4), numpy.random.rand(3,4)])
@@ -1264,7 +1278,7 @@ class T_exp(unittest.TestCase):
 class test_matinv(unittest.TestCase):
 
     def setUp(self):
-        numpy.random.seed(1)
+        unittest_tools.seed_rng()
 
     def mat_reciprocal(self,dim):
         # symbolic program
@@ -1303,8 +1317,7 @@ class test_matinv(unittest.TestCase):
 
 class t_dot(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(44)
-
+        unittest_tools.seed_rng()
     @staticmethod
     def rand(*args):
         return numpy.random.rand(*args)
@@ -1409,6 +1422,8 @@ class T_tensorfromscalar(unittest.TestCase):
 
 
 # class T_tensor(unittest.TestCase):
+#     def setUp(self):
+#         unittest_tools.seed_rng()
 #     def test0(self): # allocate from a scalar float
 #         t = _tensor(1.0)
 #         self.failUnless(isinstance(t, Tensor))
@@ -1578,7 +1593,8 @@ class test_grad(unittest.TestCase):
         self.failUnless(g2.data == 0)
 
 class T_op_cache(unittest.TestCase):
-
+    def setUp(self):
+        unittest_tools.seed_rng()
     def test0(self):
         """trigger bug in ticket #162"""
         lr = constant(0.011)
@@ -1707,6 +1723,8 @@ def test_tile():
 
 
 class test_tensordot(unittest.TestCase):
+    def setUp(self):
+        unittest_tools.seed_rng()
 
     def test0(self):
 

theano/tensor/tests/test_blas.py

@@ -5,6 +5,7 @@ import numpy
 from theano.tensor.blas import *
 from theano.tensor.blas import _dot22, res_is_a
 from unittest import TestCase
+from theano.tests import unittest_tools
 from copy import copy
 
 _as_scalar = GemmLocalOptimizer._as_scalar
@@ -17,7 +18,7 @@ from .test_basic import (_approx_eq, as_tensor, inplace_func,
 class t_gemm(TestCase):
     """This test suite is supposed to establish that gemm works as it is supposed to."""
     def setUp(self):
-        numpy.random.seed(44)
+        unittest_tools.seed_rng()
         _approx_eq.debug = 0
         Gemm.debug = False
 
@@ -265,9 +266,9 @@ def just_gemm(i, o, ishapes = [(4,3), (3,5), (4,5), (), ()]):
         for node in g.maker.env.nodes:
             if node.op == gemm: raise Exception('gemm in original graph')
 
-        rng = numpy.random.RandomState(234)
+        rng = numpy.random.RandomState(unittest_tools.fetch_seed(234))
         r0 = f(*[rng.randn(*sh) for sh in ishapes])
-        rng = numpy.random.RandomState(234)
+        rng = numpy.random.RandomState(unittest_tools.fetch_seed(234))
         r1 = g(*[rng.randn(*sh) for sh in ishapes])
         max_abs_err = numpy.max(numpy.abs(r0[0] - r1[0]))
         if  max_abs_err > 1.0e-8:
@@ -329,9 +330,9 @@ def test_gemm_opt_double_gemm():
         #for node in g.maker.env.nodes:
         #    if node.op == gemm: raise Failure('gemm in graph')
 
-        rng = numpy.random.RandomState(234)
+        rng = numpy.random.RandomState(unittest_tools.fetch_seed(234))
         r0 = f(*[rng.randn(*sh) for sh in ishapes])
-        rng = numpy.random.RandomState(234)
+        rng = numpy.random.RandomState(unittest_tools.fetch_seed(234))
         r1 = g(*[rng.randn(*sh) for sh in ishapes])
         max_abs_err = numpy.max(numpy.abs(r0[0] - r1[0]))
         if  max_abs_err > 1.0e-8:

theano/tensor/tests/test_elemwise.py

@@ -9,6 +9,7 @@ from theano.scalar import *
 
 from theano import tensor
 from theano.tensor.elemwise import *
+from theano.tests import unittest_tools
 
 
 def Env(i, o):
@@ -36,6 +37,8 @@ class test_DimShuffle(unittest.TestCase):
 
 
 class test_Broadcast(unittest.TestCase):
+    def setUp(self):
+        unittest_tools.seed_rng()
 
     def with_linker(self, linker):
         for xsh, ysh in [((3, 5), (3, 5)),
@@ -120,6 +123,8 @@ class test_Broadcast(unittest.TestCase):
 
 
 class test_CAReduce(unittest.TestCase):
+    def setUp(self):
+        unittest_tools.seed_rng()
 
     def with_linker(self, linker):
         for xsh, tosum in [((5, 6), None),

theano/tensor/tests/test_naacl09.py

@@ -9,7 +9,7 @@ from theano.compile.mode import default_mode
 from theano import tensor as T, sparse as S
 import numpy as N
 import sys
-
+from theano.tests import unittest_tools
 
 def cross_entropy(target, output, axis=1):
     """
@@ -175,11 +175,9 @@ class QuadraticDenoisingAA(module.Module):
         if (input_size is None) ^ (hidden_size is None):
             raise ValueError("Must specify input_size and hidden_size or neither.")
         super(QuadraticDenoisingAA, self)._instance_initialize(obj, {})
+        
         obj.random.initialize()
-        if seed is not None:
-            R = N.random.RandomState(seed)
-        else:
-            R = N.random
+        R = N.random.RandomState(unittest_tools.fetch_seed(seed))
         if input_size is not None:
             sz = (input_size, hidden_size)
             inf = 1/N.sqrt(input_size)
@@ -210,6 +208,8 @@ class SigmoidXEQuadraticDenoisingAA(QuadraticDenoisingAA):
     @todo: Merge this into the above.
     @todo: Default noise level for all daa levels
     """
+    def setUp(self):
+        unittest_tools.seed_rng()
 
     def build_corrupted_input(self):
         self.noise_level = theano.Member(T.scalar())
@@ -312,10 +312,7 @@ class ConvolutionalMLPInstance(module.FancyModuleInstance, Loss01):
         # ASK JAMES: Is the following necessary?
 #        super(ConvolutionalMLPInstance, self)._instance_initialize(obj, **kwargs)
 
-        if seed is not None:
-            R = N.random.RandomState(seed)
-        else:
-            R = N.random
+        R = N.random.RandomState(unittest_tools.fetch_seed(seed))
 
         self.input_size = input_size
         self.input_representation_size = input_representation_size
@@ -512,7 +509,7 @@ def test_naacl_model(iters_per_unsup=10, iters_per_sup=10,
 
     print "PROGRAM LEN %i HASH %i"% (len(m.pretraining_update.maker.env.nodes), reduce(lambda a, b: hash(a) ^ hash(b),prog_str))
 
-    rng = N.random.RandomState(23904)
+    rng = N.random.RandomState(unittest_tools.fetch_seed(23904))
 
     inputs = [rng.rand(10,9) for i in 1,2,3]
     targets = N.asarray([0,3,4,2,3,4,4,2,1,0])

theano/tensor/tests/test_nnet.py

@@ -5,13 +5,14 @@ from theano import tensor as T
 from theano import gof
 import test_basic as TT
 import numpy
+from theano.tests import unittest_tools
 
 from theano.tensor.nnet import *
 
 
 class T_sigmoid(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(9999)
+        unittest_tools.seed_rng()
     def test_elemwise(self):
         print dir(self)
         assert 0
@@ -19,13 +20,13 @@ class T_sigmoid(unittest.TestCase):
 
 class T_softplus(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(9999)
+        unittest_tools.seed_rng()
     def test_elemwise(self):
         TT.verify_grad(self, softplus, [numpy.random.rand(3,4)])
 
 class T_Softmax(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(9999)
+        unittest_tools.seed_rng()
     def test0(self):
         def f(a):
             return softmax(a)[:,0]
@@ -46,7 +47,7 @@ class T_Softmax(unittest.TestCase):
 
 class T_SoftmaxWithBias(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(9999)
+        unittest_tools.seed_rng()
     def test0(self):
         def f(a, b):
             return softmax_with_bias(a, b)[:,0]
@@ -70,7 +71,7 @@ class T_SoftmaxWithBias(unittest.TestCase):
 
 class T_CrossentropySoftmax1Hot(unittest.TestCase):
     def setUp(self):
-        numpy.random.seed(9999)
+        unittest_tools.seed_rng()
     def test0(self):
         y_idx = [0,1,3]
         def f(a, b):
@@ -84,6 +85,8 @@ class T_CrossentropySoftmax1Hot(unittest.TestCase):
         TT.verify_grad(self, f, [numpy.random.rand(3,4)])
 
 class T_prepend(unittest.TestCase):
+    def setUp(self):
+        unittest_tools.seed_rng()
     def test0(self):
         """basic functionality"""
         x=tensor.matrix('x')
@@ -109,7 +112,7 @@ class T_prepend(unittest.TestCase):
 
 class T_solve(unittest.TestCase):
     def setUp(self):
-        self.rng = numpy.random.RandomState(666)
+        self.rng = numpy.random.RandomState(unittest_tools.fetch_seed(666))
 
     def test0(self):
         A=self.rng.randn(5,5)

theano/tensor/tests/test_raw_random.py

@@ -2,6 +2,7 @@ __docformat__ = "restructuredtext en"
 import sys
 import unittest
 import numpy as N
+from theano.tests import unittest_tools
 
 from theano.tensor.raw_random import *
 

theano/tensor/tests/test_xlogx.py

@@ -8,8 +8,12 @@ import test_basic as TT
 
 import random
 import numpy.random
+from theano.tests import unittest_tools
 
 class T_XlogX(unittest.TestCase):
+    def setUp(self):
+        unittest_tools.seed_rng()
+
     def test0(self):
         x = as_tensor([1, 0])
         y = xlogx(x)

theano/tests/unittest_tools.py

+import unittest
+import numpy
+
+import os, sys
+
+def fetch_seed(pseed=None):
+    seed = os.getenv("THEANO_UNITTEST_SEED", pseed)
+    try:
+        seed = int(seed) if seed else None
+    except ValueError:
+        print >> sys.stderr, 'Error: THEANO_UNITTEST_SEED contains '\
+                'invalid seed, using None instead'
+        seed = None
+
+    return seed
+
+def seed_rng(pseed=None):
+
+    seed = fetch_seed(pseed)
+    if pseed and pseed!=seed:
+        print >> sys.stderr, 'Warning: using seed given by THEANO_UNITTEST_SEED=%i'\
+                'instead of seed %i given as parameter' % (seed, pseed)
+    numpy.random.seed(seed)
+    return seed