def test_nnet_rectangles():
rectangles_eval_fn = partial(eval_fn, protocol_cls=RectanglesVectorXV)
fmin_pass_expr_memo_ctrl(rectangles_eval_fn)
trials = hyperopt.Trials()
hyperopt.fmin(
rectangles_eval_fn,
space=nnet1_preproc_space(sup_min_epochs=20, sup_max_epochs=40),
max_evals=10,
algo=hyperopt.rand.suggest,
trials=trials,
)
def test_nnet_mrbi():
rectangles_eval_fn = partial(eval_fn,
protocol_cls=MNIST_RotatedBackgroundImages_VectorXV)
fmin_pass_expr_memo_ctrl(rectangles_eval_fn)
trials = hyperopt.Trials()
hyperopt.fmin(
rectangles_eval_fn,
space=nnet1_preproc_space(sup_min_epochs=20, sup_max_epochs=40),
max_evals=10,
algo=hyperopt.rand.suggest,
trials=trials,
)
def main_convex(filename='iptrials_convex.pkl'):
from skdata.larochelle_etal_2007.view import ConvexVectorXV as Protocol
iptrials = get_iptrials(filename)
dataset_eval_fn = partial(eval_fn, protocol_cls=Protocol)
for max_evals in range(10, 50, 10):
iptrials.fmin(
fn=dataset_eval_fn,
space=nnet1_preproc_space(),
algo=tpe.suggest,
max_evals=max_evals,
verbose=1,
pass_expr_memo_ctrl=True,
)
iptrials.wait()
iptrials.refresh()
ofile = open(filename, 'w')
cPickle.dump(iptrials, ofile)
ofile.close()
def main_rectangles(filename='iptrials_rectangles.pkl'):
from skdata.larochelle_etal_2007.view import RectanglesVectorXV
iptrials = get_iptrials(filename)
rectangles_eval_fn = partial(eval_fn, protocol_cls=RectanglesVectorXV)
for max_evals in [10, 25, 50]:
iptrials.fmin(
fn=rectangles_eval_fn,
space=nnet1_preproc_space(),
algo=rand.suggest,
max_evals=max_evals,
verbose=1,
pass_expr_memo_ctrl=True,
)
iptrials.wait()
iptrials.refresh()
ofile = open(filename, 'w')
cPickle.dump(iptrials, ofile)
ofile.close()
def main_convex(filename='iptrials_convex.pkl'):
from skdata.larochelle_etal_2007.view import ConvexVectorXV as Protocol
iptrials = get_iptrials(filename)
dataset_eval_fn = partial(eval_fn,
protocol_cls=Protocol)
for max_evals in range(10, 50, 10):
iptrials.fmin(
fn=dataset_eval_fn,
space=nnet1_preproc_space(),
algo=tpe.suggest,
max_evals=max_evals,
verbose=1,
pass_expr_memo_ctrl=True,
)
iptrials.wait()
iptrials.refresh()
ofile = open(filename, 'w')
cPickle.dump(iptrials, ofile)
ofile.close()
def main_MRBI(filename='iptrials_MRBI.pkl'):
from skdata.larochelle_etal_2007.view \
import MNIST_RotatedBackgroundImages_VectorXV as Protocol
iptrials = get_iptrials(filename)
dataset_eval_fn = partial(eval_fn, protocol_cls=Protocol)
for max_evals in range(20, 100, 200):
iptrials.fmin(
fn=dataset_eval_fn,
space=nnet1_preproc_space(),
algo=tpe.suggest,
max_evals=max_evals,
verbose=1,
pass_expr_memo_ctrl=True,
)
iptrials.wait()
iptrials.refresh()
ofile = open(filename, 'w')
cPickle.dump(iptrials, ofile)
ofile.close()
def main_rectangles(filename='iptrials_rectangles.pkl'):
from skdata.larochelle_etal_2007.view import RectanglesVectorXV
iptrials = get_iptrials(filename)
rectangles_eval_fn = partial(eval_fn,
protocol_cls=RectanglesVectorXV)
for max_evals in [10, 25, 50]:
iptrials.fmin(
fn=rectangles_eval_fn,
space=nnet1_preproc_space(),
algo=rand.suggest,
max_evals=max_evals,
verbose=1,
pass_expr_memo_ctrl=True,
)
iptrials.wait()
iptrials.refresh()
ofile = open(filename, 'w')
cPickle.dump(iptrials, ofile)
ofile.close()
def main_MRBI(filename='iptrials_MRBI.pkl'):
from skdata.larochelle_etal_2007.view \
import MNIST_RotatedBackgroundImages_VectorXV as Protocol
iptrials = get_iptrials(filename)
dataset_eval_fn = partial(eval_fn, protocol_cls=Protocol)
for max_evals in range(20, 100, 200):
iptrials.fmin(
fn=dataset_eval_fn,
space=nnet1_preproc_space(),
algo=tpe.suggest,
max_evals=max_evals,
verbose=1,
pass_expr_memo_ctrl=True,
)
iptrials.wait()
iptrials.refresh()
ofile = open(filename, 'w')
cPickle.dump(iptrials, ofile)
ofile.close()
def run_config(config):
argd = config["argd"]
def config_lookup(key):
if key == "scale_mult1":
return argd["W_init_algo_old_multiplier"]
if key == "scale_heur1":
if "old" == argd["W_init_algo"]:
return 0
else:
assert "Xavier" == argd["W_init_algo"]
return 1
if key == "preproc":
return {"raw": 0, "normalize": 1, "pca": 2}[argd["preprocessing"]]
if key == "batch_size":
return 0 if 20 == argd["batchsize"] else 1
if key == "nhid1":
return argd["n_hid"]
if key == "dist1":
return 0 if argd["W_init_algo"] == "old" else 1
if key == "squash":
return 0 if argd["squash"] == "tanh" else 1
if key == "colnorm_thresh":
return 1e-7
if key == "l2_penalty_nz":
return argd["l2_penalty"]
if key == "l2_penalty":
return 0 if argd["l2_penalty"] == 0 else 1
if key == "iseed":
# convert from seed value to choice index
return argd["iseed"] - 5
try:
return argd[key]
except KeyError:
print "Returning GarbageCollected for %s" % key
return hyperopt.pyll.base.GarbageCollected
expr = nnet1_preproc_space()
hps = {}
expr_to_config(expr, None, hps)
print config
memo = {}
for k, v in hps.items():
# print k, v
memo[v["node"]] = config_lookup(k)
print memo
rval = eval_fn(expr=expr, memo=memo, ctrl=None, protocol_cls=RectanglesVectorXV)
print "-" * 80
print "COMPUTED RESULTS IN TERMS OF *ERROR*"
print rval["loss"]
print "-" * 80
print "SAVED RESULTS IN TERMS OF *ACCURACY*"
print config["result"]
print "-" * 80
def run_config(config):
argd = config['argd']
def config_lookup(key):
if key == 'scale_mult1':
return argd['W_init_algo_old_multiplier']
if key == 'scale_heur1':
if 'old' == argd['W_init_algo']:
return 0
else:
assert 'Xavier' == argd['W_init_algo']
return 1
if key == 'preproc':
return {'raw': 0, 'normalize': 1, 'pca': 2}[argd['preprocessing']]
if key == 'batch_size':
return 0 if 20 == argd['batchsize'] else 1
if key == 'nhid1':
return argd['n_hid']
if key == 'dist1':
return 0 if argd['W_init_algo'] == 'old' else 1
if key == 'squash':
return 0 if argd['squash'] == 'tanh' else 1
if key == 'colnorm_thresh':
return 1e-7
if key == 'l2_penalty_nz':
return argd['l2_penalty']
if key == 'l2_penalty':
return 0 if argd['l2_penalty'] == 0 else 1
if key == 'iseed':
# convert from seed value to choice index
return argd['iseed'] - 5
try:
return argd[key]
except KeyError:
print 'Returning GarbageCollected for %s' % key
return hyperopt.pyll.base.GarbageCollected
expr = nnet1_preproc_space()
hps = {}
expr_to_config(expr, None, hps)
print config
memo = {}
for k, v in hps.items():
#print k, v
memo[v['node']] = config_lookup(k)
print memo
rval = eval_fn(expr=expr,
memo=memo,
ctrl=None,
protocol_cls=RectanglesVectorXV)
print '-' * 80
print 'COMPUTED RESULTS IN TERMS OF *ERROR*'
print rval['loss']
print '-' * 80
print 'SAVED RESULTS IN TERMS OF *ACCURACY*'
print config['result']
print '-' * 80
