def test_hdf5meansubtract(backend_default, meansubhdf):
NervanaObject.be.bsz = 128
bsz = 128
datit = HDF5Iterator(meansubhdf[0])
datit.allocate()
typ = meansubhdf[1]
mn = datit.mean.get()
assert typ in ['chan_mean', 'full_mean']
cnt_image = 0
max_len = datit.ndata
MAX_CNT = max_len * datit.inp.shape[1]
for x, t in datit:
x_ = x.get().flatten()
x_exp = (np.arange(len(x_)) + cnt_image) % MAX_CNT
x_exp = x_exp.reshape((-1, np.prod(datit.lshape))).T
if typ == 'chan_mean':
x_exp = x_exp.reshape((datit.lshape[0], -1)) - mn
elif typ == 'full_mean':
x_exp = x_exp.reshape((-1, bsz)) - mn
x_exp = x_exp.flatten()
assert allclose_with_out(x_, x_exp, atol=0.0, rtol=1.0e-7)
cnt_image += len(x_)
datit.cleanup()
def test_multidimout(backend_default, multidimout):
NervanaObject.be.bsz = 128
fn = multidimout
datit = HDF5Iterator(fn)
max_len = datit.ndata
MAX_CNT = max_len * datit.inp.shape[1]
for (x, t) in datit:
x_ = x.get()
t_ = t.get()
assert x_.shape == t_.shape
x_ = x_.T % MAX_CNT
t_ = t_.T % MAX_CNT
assert np.all(x_ == t_[:, ::-1])
datit.cleanup()
def test_reset(backend_default, hdf5datafile):
NervanaObject.be.bsz = 128
fn = hdf5datafile
datit = HDF5Iterator(fn)
for (x, t) in datit:
break
x_1 = x.get()
t_1 = t.get()
for cnt_end in range(2):
cnt = 0
for (x, t) in datit:
cnt += 1
if cnt > cnt_end:
break
datit.reset()
for (x, t) in datit:
break
assert np.all(x.get() == x_1)
assert np.all(t.get() == t_1)
datit.cleanup()
def test_h5iterator(backend_default, hdf5datafile, onehot):
NervanaObject.be.bsz = 128
bsz = 128
fn = hdf5datafile
if onehot:
datit = HDF5IteratorOneHot(fn)
else:
datit = HDF5Iterator(fn)
cnt_image = 0
cnt_target = 0
max_len = datit.ndata
mb_cnt = 0
MAX_CNT = max_len * datit.inp.shape[1]
for (x, t) in datit:
x_ = x.get()
t_ = t.get()
assert x_.shape[1] == t_.shape[1]
assert not np.all(x_ == 0)
assert not np.all(t_ == 0)
x_ = x_.T.flatten()
x_exp = (np.arange(len(x_)) + cnt_image) % MAX_CNT
assert np.all(x_ == x_exp)
cnt_image += len(x_)
if onehot:
t_ = np.argmax(t_, axis=0).flatten()
else:
t_ = t_.flatten()
t_exp = (np.arange(len(t_)) + cnt_target) % max_len
assert np.all(t_ == t_exp)
cnt_target += len(t_)
mb_cnt += 1
assert mb_cnt == int(math.ceil(datit.inp.shape[0] / float(bsz)))
datit.cleanup()
print('Using test file: {}'.format(testFileName))
# Next line gets rid of the deterministic warning
args.deterministic = None
if (args.rng_seed is None):
args.rng_seed = 16
print('Batch size = {}'.format(args.batch_size))
# setup backend
be = gen_backend(**extract_valid_args(args, gen_backend))
# We don't need to one hot encode this since it is only 2 classes
# and the sklearn stuff is just looking at the probability of class 1
test_set = HDF5Iterator(testFileName)
#test_set = HDF5IteratorOneHot(testFileName)
model_filename= 'LUNA16_resnetHDF_subset{}.prm'.format(subset)
#model_filename= 'LUNA16_resnetHDF_subset0.prm'
print('Using model: {}'.format(model_filename))
lunaModel = Model(model_filename)
prob, target = lunaModel.get_outputs(test_set, return_targets=True)
np.set_printoptions(precision=3, suppress=True)
from neon.layers import GeneralizedCost, Affine, Sequential, MergeMultistream, Linear, Pooling, Conv,Dropout
from neon.models import Model
from neon.optimizers import GradientDescentMomentum, RMSProp, Adam
from neon.transforms import Rectlin, Logistic, CrossEntropyBinary, SumSquared, ObjectDetection
from neon.callbacks.callbacks import Callbacks
from neon.backends import gen_backend
# parser = NeonArgparser(__doc__)
# args = parser.parse_args(gen_be=False)
be = gen_backend(batch_size=128, backend='gpu')
traindir = 'train'
imwidth = 256
train_set = HDF5Iterator('whale_train.h5')
eval_set = HDF5Iterator('whale_eval.h5')
test_set = HDF5Iterator('whale_test.h5')
# weight initialization
init_norm = Gaussian(loc=0.0, scale=0.01)
# setup model layers
layers = [Conv((7, 7, 16), init=init_norm, activation=Rectlin()),
Pooling((2, 2)),
Conv((3, 3, 32), init=init_norm, activation=Rectlin()),
Conv((3, 3, 32), init=init_norm, activation=Rectlin()),
Conv((3, 3, 32), init=init_norm, activation=Rectlin()),
