Merge pull request #1804 from nouiz/fix_scan_mem

.travis.yml

@@ -17,6 +17,7 @@ install:
 # So we test with 0.8. Our internal buildbot have 0.7.2.
 # We install it later only for the PART that need it.
 #  - "pip install -q scipy==0.8 --use-mirrors"
+  - "pip install nose-timelimit --use-mirrors"
   - "pip install . --no-deps --use-mirrors"
 # command to run tests
 env:
@@ -36,7 +37,7 @@ script:
   - df -h
   - ulimit -a
   - echo $PART
-  - theano-nose -v $PART
+  - theano-nose --with-timelimit -v $PART
 
 
 #after_script:

theano/compile/mode.py

@@ -189,9 +189,6 @@ optdb.register('Print1.51', PrintCurrentFunctionGraph('Post-stabilize'),
 optdb.register('specialize', gof.EquilibriumDB(),
         2, 'fast_run')
 
-optdb.register('Print2.01', PrintCurrentFunctionGraph('Post-specialize'),
-        2.01,)  # 'fast_run', 'fast_compile')
-
 # misc special cases for speed that break canonicalization
 optdb.register('uncanonicalize', gof.EquilibriumDB(),
         3, 'fast_run')

theano/gof/optdb.py

@@ -248,6 +248,11 @@ class SequenceDB(DB):
                 position_cutoff = tags[0].position_cutoff
 
         opts = [o for o in opts if self.__position__[o.name] < position_cutoff]
+        # We want to sort by position and then if collision by name
+        # for deterministic optimization.  Since Python 2.2, sort is
+        # stable, so sort by name first, then by position. This give
+        # the order we want.
+        opts.sort(key=lambda obj: obj.name)
         opts.sort(key=lambda obj: self.__position__[obj.name])
         ret = opt.SeqOptimizer(opts, failure_callback=self.failure_callback)
         if hasattr(tags[0], 'name'):

theano/scan_module/scan_opt.py

@@ -1681,7 +1681,6 @@ scan_eqopt1 = theano.gof.EquilibriumDB()
 scan_seqopt1 = theano.gof.SequenceDB()
 
 scan_eqopt2 = theano.gof.EquilibriumDB()
-scan_seqopt2 = theano.gof.EquilibriumDB()
 # We run before blas opt at 1.7 and specialize 2.0
 # but after stabilize at 1.5. Should we put it before stabilize?
 optdb.register('scan_eqopt1', scan_eqopt1, .1, 'fast_run', 'scan')
@@ -1694,8 +1693,6 @@ optdb.register('scanOp_make_inplace',
                'inplace',
                'scan')
 
-scan_eqopt2.register(
-    'all_scan_opts', scan_seqopt2, 1, 'fast_run', 'scan')
 scan_eqopt1.register(
     'all_pushout_opt', scan_seqopt1, 1, 'fast_run', 'scan')
 
@@ -1731,7 +1728,7 @@ scan_seqopt1.register('scan_pushout_dot1',
                       'scan')
 
 
-scan_seqopt2.register('constant_folding_for_scan2',
+scan_eqopt2.register('constant_folding_for_scan2',
                       opt.in2out(tensor.opt.constant_folding,
                                  ignore_newtrees=True),
                       1,
@@ -1739,7 +1736,7 @@ scan_seqopt2.register('constant_folding_for_scan2',
                       'scan')
 
 
-scan_seqopt2.register('scanOp_remove_constants_and_unused_inputs1',
+scan_eqopt2.register('scanOp_remove_constants_and_unused_inputs1',
                       opt.in2out(remove_constants_and_unused_inputs_scan,
                                  ignore_newtrees=True),
                       2,
@@ -1751,14 +1748,14 @@ scan_seqopt2.register('scanOp_remove_constants_and_unused_inputs1',
 # after const merge but before stabilize so that we can have identity
 # for equivalent nodes but we still have the chance to hoist stuff out
 # of the scan later.
-scan_seqopt2.register('scanOp_merge',
+scan_eqopt2.register('scanOp_merge',
                       ScanMerge(),
                       4,
                       'fast_run',
                       'scan')
 
 # After Merge optimization
-scan_seqopt2.register('scanop_remove_constants_and_unused_inputs2',
+scan_eqopt2.register('scanop_remove_constants_and_unused_inputs2',
                       opt.in2out(remove_constants_and_unused_inputs_scan,
                                  ignore_newtrees=True),
                       5,
@@ -1766,7 +1763,7 @@ scan_seqopt2.register('scanop_remove_constants_and_unused_inputs2',
                       'fast_run',
                       'scan')
 
-scan_seqopt2.register('scanOp_merge_inouts',
+scan_eqopt2.register('scanOp_merge_inouts',
                       opt.in2out(scan_merge_inouts, ignore_newtrees=True),
                       6,
                       'scan_merge_inouts',
@@ -1776,14 +1773,14 @@ scan_seqopt2.register('scanOp_merge_inouts',
 # Just before specialize to have the other optimization
 # like constant folding being applied
 # This don't introduce inplace.
-scan_seqopt2.register('scanOp_save_mem',
+scan_eqopt2.register('scanOp_save_mem',
                       ScanSaveMem(),
                       7,
                       'fast_run',
                       'scan')
 
 # After everything else
-scan_seqopt2.register('scanOp_remove_constants_and_unused_inputs3',
+scan_eqopt2.register('scanOp_remove_constants_and_unused_inputs3',
                       opt.in2out(remove_constants_and_unused_inputs_scan,
                                  ignore_newtrees=True),
                       8,

theano/tensor/opt.py

@@ -303,7 +303,8 @@ def inplace_elemwise_optimizer_op(OP):
     return inplace_elemwise_optimizer
 
 inplace_elemwise_optimizer = inplace_elemwise_optimizer_op(T.Elemwise)
-compile.optdb.register('inplace_opt', inplace_elemwise_optimizer, 75,
+compile.optdb.register('inplace_elemwise_opt', inplace_elemwise_optimizer, 75,
+                       'inplace_opt',  # for historic reason
                        'inplace_elemwise_optimizer',
                        'fast_run', 'inplace')