提交 b131e669 authored 作者: lamblin's avatar lamblin

Merge pull request #977 from nouiz/flags

Flags
......@@ -245,6 +245,22 @@ import theano and print the config variable, as in:
This sets the default compilation mode for theano functions. By default the
mode Mode is equivalent to FAST_RUN. See Config attribute linker and optimizer.
.. attribute:: profile
Bool value: either True or False
Default False
Do the vm/cvm linker profile the execution of Theano function?
.. attribute:: profile_optimizer
Bool value: either True or False
Default False
Do the vm/cvm linker profile the optimization phase when compiling a Theano function?
.. attribute:: config.lib.amdlibm
Bool value: either True or False
......@@ -580,49 +596,6 @@ import theano and print the config variable, as in:
B. log_likelihood_v_given_h
C. log_likelihood_h
.. attribute:: config.time_seq_optimizer
Bool value, default: False
Should each SeqOptimizer object print the time taken by each of its
optimizer. Each SeqOptimizer prints something like this::
SeqOptimizer gpu_opt time 0.014s for 8/9 nodes before/after optimization
0.012573s - ('gpu_local_optimizations', 'EquilibriumOptimizer')
0.001373s - ('gpu_cut_transfers', 'EquilibriumOptimizer')
0.000441s - ('InputToGpuOptimizer', 'InputToGpuOptimizer')
This prints the name of the SeqOptimizer (gpu_opt), the number of
Apply nodes in the graph before (8) and after (9)
optimizations. Then a list of lines, one per optimization
in this SeqOptimizer. The first element is the time
taken by this optimization and then it is a tuple with the name of the
optimization and this class. This list is sorted from the sub
optimization that takes the most time to the optimization that takes
the least time.
.. attribute:: config.time_eq_optimizer
Bool value, default: False
Should each EquilibriumOptimizer print the time taken by each of its
iterations, the total number of times each of its optimizers is applied,
and informations about the total number of nodes in the graph.
Here is an example of output::
EquilibriumOptimizer specialize
time 4.760s for 4 passes, 3801 nodes max
0 - 1.961s (0.079s in global opts) - 3797 nodes
1 - 1.233s (0.080s in global opts) - 3801 nodes
2 - 0.857s (0.071s in global opts) - 3203 nodes
3 - 0.710s (0.066s in global opts) - 3095 nodes
times applied - optimizer:
384 - dimshuffle_as_view
262 - constant_folding
216 - local_subtensor_make_vector
216 - local_shape_to_shape_i
4 - local_mul_specialize
.. attribute:: config.cmodule.warn_no_version
......
......@@ -26,15 +26,6 @@ from theano.configparser import AddConfigVar, BoolParam
_logger = logging.getLogger('theano.gof.opt')
AddConfigVar('time_seq_optimizer',
"Should SeqOptimizer print the time taked by each of its optimizer",
BoolParam(False),
in_c_key=False)
AddConfigVar('time_eq_optimizer',
"Should EquilibriumOptimizer print the time taken by each optimizer",
BoolParam(False),
in_c_key=False)
import destroyhandler as dh
import traceback
......@@ -180,29 +171,6 @@ class SeqOptimizer(Optimizer, list):
else:
raise
if config.time_seq_optimizer:
print "SeqOptimizer",
if hasattr(self,"name"): print self.name,
elif hasattr(self,"__name__"): print self.__name__,
print " time %.3fs for %d/%d nodes before/after optimization"%(sum(l),nb_node_before,len(fgraph.apply_nodes))
print " time %.3fs for validate " % (
fgraph.profile.validate_time - validate_before)
ll=[]
for opt in self:
if hasattr(opt,"__name__"):
ll.append((opt.__name__,opt.__class__.__name__))
else:
ll.append((opt.name,opt.__class__.__name__))
lll=zip(l,ll)
def cmp(a,b):
if a[0]==b[0]: return 0
if a[0]<b[0]: return -1
return 1
lll.sort(cmp)
for (t, opt) in lll[::-1]:
print ' %.6fs - %s' % (t, opt)
print
if fgraph.profile:
validate_time = fgraph.profile.validate_time - validate_before
else:
......@@ -1517,29 +1485,6 @@ class EquilibriumOptimizer(NavigatorOptimizer):
+ "%f with the theano flag 'optdb.max_use_ratio'." %
config.optdb.max_use_ratio)
if config.time_eq_optimizer:
print "EquilibriumOptimizer",
print getattr(self, "name", getattr(self, "__name__", ""))
print " time %.3fs for %d passes, %d nodes max" % (
sum(loop_timing), len(loop_timing), max_nb_nodes)
for i in range(len(loop_timing)):
print '%d - %.3fs (%.3fs in global opts) - %d nodes' % (
i, loop_timing[i], global_opt_timing[i], nb_nodes[i])
print
count_opt = []
for opt, count in process_count.iteritems():
if count > 0:
count_opt.append((count, opt))
if count_opt:
print 'times applied - optimizer:'
count_opt.sort()
for (count, opt) in count_opt[::-1]:
print ' %d - %s' % (count, opt)
print
return (self, loop_timing, process_count, max_nb_nodes,
global_opt_timing, nb_nodes, time_lopts, io_toposort_timing)
......
Markdown 格式
0%
您添加了 0 到此讨论。请谨慎行事。
请先完成此评论的编辑!
注册 或者 后发表评论