def log_queue(log_to=None):
if log_to:
# standard logfile
jlog = munk.file_sink(log_to + ".log")
jlog = munk.jsonify(jlog)
jlog = munk.timify(jlog, tag="timestamp")
jlog = munk.exclude(jlog, "params")
# parameter logfile
paraml = munk.file_sink(log_to + ".params")
paraml = munk.jsonify(paraml)
paraml = munk.timify(paraml, tag="timestamp")
paraml = munk.include(paraml, "params")
jplog = munk.broadcast(*[jlog, paraml])
# finally a pretty printer for some immediate feedback
pp = munk.timify(munk.prettyprint_sink())
pp = munk.dontkeep(pp, "tags")
pp = munk.include_tags_only(pp, "pretty")
jplog = munk.exclude_tags(jplog, "pretty")
log = munk.broadcast(*[jplog, pp])
else:
pp = munk.timify(munk.prettyprint_sink())
pp = munk.dontkeep(pp, "tags")
log = munk.include_tags_only(pp, "pretty")
return log
def log_queue(log_to=None):
if log_to:
# standard logfile
jlog = munk.file_sink(log_to+".log")
jlog = munk.jsonify(jlog)
jlog = munk.timify(jlog, tag="timestamp")
jlog = munk.exclude(jlog, "params")
# parameter logfile
paraml = munk.file_sink(log_to+".params")
paraml = munk.jsonify(paraml)
paraml = munk.timify(paraml, tag="timestamp")
paraml = munk.include(paraml, "params")
jplog = munk.broadcast(*[jlog, paraml])
# finally a pretty printer for some immediate feedback
pp = munk.timify(munk.prettyprint_sink())
pp = munk.dontkeep(pp, "tags")
pp = munk.include_tags_only(pp, "pretty")
jplog = munk.exclude_tags(jplog, "pretty")
log = munk.broadcast(*[jplog, pp])
else:
pp = munk.timify(munk.prettyprint_sink())
pp = munk.dontkeep(pp, "tags")
log = munk.include_tags_only(pp, "pretty")
return log
def test_dontkeep():
lst = []
pipe = L.list_sink(lst)
pipe = L.dontkeep(pipe, ['blabla'])
pipe.send({'message': 'take care', 'blabla': 'blubb'})
assert len(lst) == 1, 'wrong number of messages got through'
assert lst[0] == {'message': 'take care'}, 'wrong keys got through'
def reload(depot, folder, tag, layer):
"""
"""
import notebook as nb
model, schedule = nb.reload(depot, folder, tag, layer)
log = munk.prettyprint_sink()
log = munk.dontkeep(log, "tags")
log = munk.include_tags_only(log, "pretty")
schedule['logging'] = log
lab = schedule['__lab__']
lab = __import__(lab.split('.')[0])
lab.no_training(model, schedule)
def reload(depot, folder, tag, layer):
"""
"""
import notebook as nb
model, schedule = nb.reload(depot, folder, tag, layer)
log = munk.prettyprint_sink()
log = munk.dontkeep(log, "tags")
log = munk.include_tags_only(log, "pretty")
schedule['logging'] = log
lab = schedule['__lab__']
lab = __import__(lab.split('.')[0])
lab.no_training(model, schedule)
print '#pars:', P.data.size
## import numericalGradientChecker
## checker = numericalGradientChecker.numericalGradientChecker(fraw, fprime, inputDim=P.data.size, outputDim=batchsize, args=args, bounds=None)
## for i, info in enumerate(checker):
## print "errors in gradient", info['errors']
## if i>0:
## break
import chopmunk
ignore = ['args', 'kwargs', 'gradient', 'Hp']
console_sink = chopmunk.prettyprint_sink()
console_sink = chopmunk.dontkeep(console_sink, ignore)
file_sink = chopmunk.file_sink('mnist.log')
file_sink = chopmunk.jsonify(file_sink)
file_sink = chopmunk.dontkeep(file_sink, ignore)
logger = chopmunk.broadcast(console_sink, file_sink)
logfunc = logger.send
#size of blocks used for the diagonal approximation
blocksizes = np.ones(n_hidden + n_output)
blocksizes[:n_hidden] *= (n_inpt+1)
blocksizes[n_hidden:] *= (n_hidden+1)
opt = tonga(P.data, fprime, damping=1e-2, blocksizes=blocksizes, nb_estimates=50, args=args, cov_args=cov_args, logfunc=logfunc)