def test_get_inputs(self):
d = MNIST(repo_path=self.tmp_repo)
d.backend = CPU(rng_seed=0)
d.backend.actual_batch_size = 128
inputs = d.get_inputs(train=True)
# TODO: make this work (numpy import errors at the moment)
assert inputs['train'] is not None
def test_coarse_labels(self):
data = CIFAR100(coarse=True, repo_path=self.tmp_repo)
data.backend = CPU(rng_seed=0)
data.backend.actual_batch_size = 128
data.load()
assert len(data.inputs['train']) == 50000
assert len(data.targets['train'][0]) == 20
def test_softmax_cputensor():
sftmx = Softmax()
inputs = np.array([0, 1, -2]).reshape((3, 1))
be = CPU(rng_seed=0)
temp = be.zeros((3, 1))
outputs = np.exp(inputs - 1) / np.sum(np.exp(inputs - 1))
sftmx.apply_function(be, CPUTensor(inputs), temp)
assert_tensor_near_equal(CPUTensor(outputs), temp)
def test_logistic_cputensor():
lgstc = Logistic()
inputs = np.array([0, 1, -2]).reshape((3, 1))
be = CPU(rng_seed=0)
temp = be.zeros((3, 1))
outputs = 1.0 / (1.0 + np.exp(-inputs))
lgstc.apply_function(be, CPUTensor(inputs), temp)
assert_tensor_near_equal(CPUTensor(outputs), temp)
def test_rectleaky_slope_zero_rectlin_equiv():
be = CPU()
inputs = be.uniform(low=-5.0, high=10.0, size=(10, 10))
lin_buf = be.empty(inputs.shape)
leaky_buf = be.empty(inputs.shape)
be.rectlin(inputs, out=lin_buf)
be.rectleaky(inputs, slope=0.0, out=leaky_buf)
assert_tensor_equal(lin_buf, leaky_buf)
def test_cross_entropy_derivative_cputensor():
be = CPU(rng_seed=0)
outputs = CPUTensor([0.5, 0.9, 0.1, 0.0001])
targets = CPUTensor([0.5, 0.99, 0.01, 0.2])
temp = [be.zeros(outputs.shape), be.zeros(outputs.shape)]
expected_result = ((outputs.asnumpyarray() - targets.asnumpyarray()) /
(outputs.asnumpyarray() * (1 - outputs.asnumpyarray())))
assert_tensor_near_equal(
expected_result, cross_entropy_derivative(be, outputs, targets, temp))
def test_fine_labels(self):
data = CIFAR100(coarse=False, repo_path=self.tmp_repo)
data.backend = CPU(rng_seed=0)
data.backend.actual_batch_size = 128
par = NoPar()
par.associate(data.backend)
data.load()
assert len(data.inputs['train']) == 50000
assert len(data.targets['train'][0]) == 100
def test_tanh_cputensor():
tntest = Tanh()
be = CPU(rng_seed=0)
CPUTensor([0, 1, -2])
inputs = np.array([0, 1, -2])
temp = be.zeros(inputs.shape)
outputs = np.array([true_tanh(0), true_tanh(1), true_tanh(-2)])
tntest.apply_function(be, CPUTensor(inputs), temp)
assert_tensor_near_equal(CPUTensor(outputs), temp)
def test_logistic_derivative_cputensor():
lgstc = Logistic()
inputs = np.array([0, 1, -2]).reshape((3, 1))
be = CPU(rng_seed=0)
outputs = 1.0 / (1.0 + np.exp(-inputs))
outputs = outputs * (1.0 - outputs)
temp = be.zeros(inputs.shape)
lgstc.apply_derivative(be, CPUTensor(inputs), temp)
assert_tensor_near_equal(CPUTensor(outputs), temp)
def test_tanh_derivative_cputensor():
tntest = Tanh()
inputs = np.array([0, 1, -2])
be = CPU(rng_seed=0)
outputs = np.array(
[1 - true_tanh(0)**2, 1 - true_tanh(1)**2, 1 - true_tanh(-2)**2])
temp = be.zeros(inputs.shape)
tntest.apply_derivative(be, CPUTensor(inputs), temp)
assert_tensor_near_equal(CPUTensor(outputs), temp)
def compare_cpu_tensors(inputs, outputs, deriv=False):
rlin = RectLeaky()
be = CPU()
temp = be.zeros(inputs.shape)
if deriv is True:
rlin.apply_derivative(be, CPUTensor(inputs), temp)
else:
rlin.apply_function(be, CPUTensor(inputs), temp)
be.subtract(temp, CPUTensor(outputs), temp)
assert_tensor_equal(temp, be.zeros(inputs.shape))
def test_cross_entropy_cputensor():
be = CPU(rng_seed=0)
outputs = CPUTensor([0.5, 0.9, 0.1, 0.0001])
targets = CPUTensor([0.5, 0.99, 0.01, 0.2])
temp = [be.zeros(outputs.shape), be.zeros(outputs.shape)]
expected_result = np.sum(
(-targets.asnumpyarray()) * np.log(outputs.asnumpyarray()) -
(1 - targets.asnumpyarray()) * np.log(1 - outputs.asnumpyarray()),
keepdims=True)
assert_tensor_near_equal(expected_result,
cross_entropy(be, outputs, targets, temp))
def test_softmax_derivative_cputensor():
sftmx = Softmax()
inputs = np.array([0, 1, -2]).reshape((3, 1))
be = CPU(rng_seed=0)
outputs = np.exp(inputs) / np.sum(np.exp(inputs))
errmat = np.ones(inputs.shape)
a = np.einsum('ij,ji->i', errmat.T, outputs)
outputs = outputs * (errmat - a[np.newaxis, :])
temp = be.zeros(inputs.shape)
sftmx.apply_derivative(be, CPUTensor(inputs), temp)
assert_tensor_near_equal(CPUTensor(outputs), temp)
def __setstate__(self, state):
"""
Defines how we go about deserializing and loading an instance of this
class from a specified state.
Arguments:
state (dict): attribute values to be loaded.
"""
self.__dict__.update(state)
if self.backend is None:
# use CPU as a default backend
self.backend = CPU()
def test_cross_entropy_limits():
be = CPU(rng_seed=0)
outputs = CPUTensor([0.5, 1.0, 0.0, 0.0001])
targets = CPUTensor([0.5, 0.0, 1.0, 0.2])
eps = 2**-23
temp = [be.zeros(outputs.shape), be.zeros(outputs.shape)]
expected_result = np.sum(
(-targets.asnumpyarray()) * np.log(outputs.asnumpyarray() + eps) -
(1 - targets.asnumpyarray()) *
np.log(1 - outputs.asnumpyarray() + eps),
keepdims=True)
assert_tensor_near_equal(expected_result,
cross_entropy(be, outputs, targets, temp, eps))
def test_round_split(self):
split = 10
batch_size = 32
ntrain, ntest, nin, nout = 100, 10, 10, 5
data = UniformRandom(ntrain, ntest, nin, nout, validation_pct=split)
data.backend = CPU(rng_seed=0)
data.backend.batch_size = batch_size
data.load()
split /= 100.0
nb_batches = ntrain // batch_size
expected_nb_train = floor((1.0 - split) * nb_batches)
expected_nb_valid = floor(split * nb_batches)
assert expected_nb_train == len(data.inputs['train'])
assert expected_nb_train == len(data.targets['train'])
assert expected_nb_valid == len(data.inputs['validation'])
assert expected_nb_valid == len(data.targets['validation'])
def test_xcov_cputensor():
np.random.seed(0)
n = 10
k = 8
(k1, k2) = (3, 5)
a = np.array(np.random.randn(k, n) * 10, dtype='float32', order='C')
acc = xcc(a[:k1], a[k1:])
expected_result = 0.5 * (acc**2.).sum()
be = CPU(rng_seed=0)
outputs = CPUTensor(a.copy())
tempbuf1 = be.empty((k1, n))
tempbuf2 = be.empty((k2, n))
tempbuf3 = be.empty((k1, k2))
tempbuf4 = be.empty(outputs.shape)
temp = [tempbuf1, tempbuf2, tempbuf3, tempbuf4]
my_result = xcov_cost(be, outputs, [], temp, k1)
assert_tensor_near_equal(expected_result, my_result)
def test_xcov_derivative_cputensor():
np.random.seed(0)
n = 10
k = 8
(k1, k2) = (3, 5)
a = np.array(np.random.randn(k, n), dtype='float32', order='C')
s = np.zeros_like(a)
acc = xcc(a[:k1], a[k1:]) # k1 x k2
c1 = a[k1:] - a[k1:].mean(1, keepdims=True) # k2 x n
c2 = a[:k1] - a[:k1].mean(1, keepdims=True) # k1 x n
s[:k1] = acc.dot(c1) / n
s[k1:] = acc.T.dot(c2) / n
be = CPU(rng_seed=0)
outputs = CPUTensor(a.copy())
tempbuf1 = be.empty((k1, n))
tempbuf2 = be.empty((k2, n))
tempbuf3 = be.empty((k1, k2))
tempbuf4 = be.empty(outputs.shape)
temp = [tempbuf1, tempbuf2, tempbuf3, tempbuf4]
my_result = xcov_cost_derivative(be, outputs, [], temp, k1)
expected_result = CPUTensor(s)
assert_tensor_near_equal(expected_result, my_result)
def gen_backend(model=None,
gpu=None,
nrv=False,
datapar=False,
modelpar=False,
flexpoint=False,
rng_seed=None,
numerr_handling=None,
half=False,
stochastic_round=0,
device_id=None):
"""
Construct and return a backend instance of the appropriate type based on
the arguments given. With no parameters, a single CPU core, float32
backend is returned.
Arguments:
model (neon.models.model.Model): The instantiated model upon which we
will utilize this backend.
gpu (string, optional): Attempt to utilize a CUDA capable GPU if
installed in the system. Defaults to None which
implies a CPU based backend. If 'cudanet',
utilize a cuda-convnet2 based backed, which
supports Kepler and Maxwell GPUs with single
precision. If 'nervanagpu', attempt to utilize
the NervanaGPU Maxwell backend with float16 and
float32 support.
nrv (bool, optional): If True, attempt to utilize the Nervana Engine
for computation (must be installed on the
system). Defaults to False which implies a CPU
based backend.
datapar (bool, optional): Set to True to ensure that data is
partitioned and each chunk is processed in
parallel on different compute cores. Requires
mpi4py. Defaults to False which implies that
all data will be processed sequentially on a
single compute core.
modelpar (bool, optional): Set to True to ensure that the nodes in each
model layer are partitioned and distributed
across multiple compute cores. Requires
mpi4py. Defaults to False which implies
that all nodes in all model layers will be
processed by the same single compute core.
flexpoint (bool, optional): If True, attempt to use FlexPoint(TM)
element typed data instead of the default
float32 which is in place if set to False.
rng_seed (numeric, optional): Set this to a numeric value which can be
used to seed the random number generator
of the instantiated backend. Defaults to
None, which doesn't explicitly seed (so
each run will be different)
stochastic_round (numeric, optional): Only affects the max backend. If
1, perform stochastic rounding.
If 0, round to nearest.
numerr_handling (dict, optional): Dictate how numeric errors are
displayed and handled. The keys and
values permissible for this dict
match that seen in numpy.seterr.
If set to None (the default),
behavior is equivalent to
{'all': 'warn'}
device_id (numeric, optional): Set this to a numeric value which can be
used to select which device to run the
process on
Returns:
Backend: newly constructed backend instance of the specifed type.
Notes:
* Attempts to construct a GPU instance without a CUDA capable card or
without cudanet or nervanagpu package installed will cause the
program to display an error message and exit.
* Attempts to construct a parallel instance without mpi4py installed
will cause the program to display an error message and exit.
* The returned backend will still need to call its par.init_model()
at some point after the model has been linked, in order for parallel
training to proceed.
"""
logger = logging.getLogger(__name__)
gpuflag = False
if datapar and modelpar:
raise NotImplementedError('Hybrid parallelization scheme not '
'implemented yet. Try with at most one of'
'datapar or modelpar')
if modelpar:
par = ModelPar()
elif datapar:
par = DataPar()
else:
par = NoPar()
if par.device_id is not None:
if device_id is not None:
logger.warn('Ignoring device id specified in command line.')
device_id = par.device_id
if gpu is not None:
gpu = gpu.lower()
if sys.platform.startswith("linux"):
gpuflag = (os.system("nvidia-smi > /dev/null 2>&1") == 0)
elif sys.platform.startswith("darwin"):
gpuflag = (
os.system("kextstat | grep -i cuda > /dev/null 2>&1") == 0)
if gpuflag and gpu == 'cudanet':
try:
import cudanet # noqa
from neon.backends.cc2 import GPU
be_name = 'Cudanet'
be = GPU(rng_seed=rng_seed, device_id=device_id)
except ImportError:
logger.warning("cudanet not found, can't run via GPU")
gpuflag = False
elif gpuflag and gpu == 'nervanagpu':
try:
import nervanagpu # noqa
try:
# import pycuda.autoinit
import pycuda.driver as drv
drv.init()
device_id = device_id if device_id is not None else 0
global ctx
ctx = drv.Device(device_id).make_context()
import atexit
atexit.register(ctx.pop)
from neon.backends.gpu import GPU
be_name = 'NervanaGPU'
be = GPU(rng_seed=rng_seed,
stochastic_round=stochastic_round,
device_id=device_id)
except ImportError:
logger.warning("pycuda error, can't run via GPU")
gpuflag = False
except ImportError:
logger.warning("nervanagpu not found, can't run via GPU")
gpuflag = False
if gpuflag is False:
raise RuntimeError("Can't find CUDA capable GPU")
elif nrv:
nrv = False
try:
from umd.nrv_backend import NRVBackend
nrv = True
except ImportError:
logger.warning("Nervana Engine system software not found")
if flexpoint:
logger.warning("Flexpoint(TM) backend not currently available")
if nrv:
be_name = 'NRV'
be = NRVBackend(rng_seed=rng_seed,
seterr_handling=numerr_handling,
device_id=device_id)
elif not gpuflag:
be_name = 'CPU'
be = CPU(rng_seed=rng_seed, seterr_handling=numerr_handling)
logger.info("{} backend, RNG seed: {}, numerr: {}".format(
be_name, rng_seed, numerr_handling))
par.associate(be)
return be
def test_argmax_noaxis(self):
be = CPU()
tsr = be.array([[-1, 0], [1, 92]])
out = be.empty([1, 1])
be.argmax(tsr, None, out)
assert_tensor_equal(out, CPUTensor(3))
def gen_backend(model=None,
gpu=None,
nrv=False,
flexpoint=False,
rng_seed=None,
numerr_handling=None,
half=False,
stochastic_round=0,
device_id=None):
"""
Construct and return a backend instance of the appropriate type based on
the arguments given. With no parameters, a single CPU core, float32
backend is returned.
Arguments:
model (neon.models.model.Model): The instantiated model upon which we
will utilize this backend.
gpu (string, optional): Attempt to utilize a CUDA capable GPU if
installed in the system. Defaults to None which
implies a CPU based backend. If 'cudanet',
utilize a cuda-convnet2 based backed, which
supports Kepler and Maxwell GPUs with single
precision. If 'nervanagpu', attempt to utilize
the NervanaGPU Maxwell backend with float16 and
float32 support.
nrv (bool, optional): If True, attempt to utilize the Nervana Engine
for computation (must be installed on the
system). Defaults to False which implies a CPU
based backend.
rng_seed (numeric, optional): Set this to a numeric value which can be
used to seed the random number generator
of the instantiated backend. Defaults to
None, which doesn't explicitly seed (so
each run will be different)
stochastic_round (numeric, optional): Only affects the max backend. If
1, perform stochastic rounding.
If 0, round to nearest.
numerr_handling (dict, optional): Dictate how numeric errors are
displayed and handled. The keys and
values permissible for this dict
match that seen in numpy.seterr.
If set to None (the default),
behavior is equivalent to
{'all': 'warn'}
device_id (numeric, optional): Set this to a numeric value which can be
used to select which device to run the
process on
Returns:
Backend: newly constructed backend instance of the specifed type.
Notes:
* Attempts to construct a GPU instance without a CUDA capable card or
without cudanet or nervanagpu package installed will cause the
program to display an error message and exit.
* Attempts to construct a parallel instance without mpi4py installed
will cause the program to display an error message and exit.
* The returned backend will still need to call its par.init_model()
at some point after the model has been linked, in order for parallel
training to proceed.
"""
logger = logging.getLogger(__name__)
gpuflag = False
if gpu is not None:
gpu = gpu.lower()
if sys.platform.startswith("linux"):
gpuflag = (os.system("nvcc --version > /dev/null 2>&1") == 0)
elif sys.platform.startswith("darwin"):
gpuflag = (
os.system("kextstat | grep -i cuda > /dev/null 2>&1") == 0)
if gpuflag and gpu == 'cudanet':
try:
import cudanet # noqa
from neon.backends.cc2 import GPU
be_name = 'Cudanet'
be = GPU(rng_seed=rng_seed, device_id=device_id)
except ImportError:
raise RuntimeError("cudanet not found, can't run via GPU")
elif gpuflag and gpu.startswith('nervanagpu'):
try:
import nervanagpu # noqa
try:
be_name = 'NervanaGPU'
if gpu == 'nervanagpu':
device_id = 0 if device_id is None else device_id[0]
from neon.backends.gpu import GPU
be = GPU(rng_seed=rng_seed,
stochastic_round=stochastic_round,
device_id=device_id)
else:
from neon.backends.mgpu import MGPU
try:
num_dev = int(gpu.strip('nervanagpu'))
except (ValueError):
raise ValueError("invalid number of GPUs" +
" specified")
if not device_id:
device_id = range(num_dev)
if len(device_id) != num_dev:
raise RuntimeError(
"Incorrect number of devices"
" specified ", device_id, num_dev)
be = MGPU(rng_seed=rng_seed,
stochastic_round=stochastic_round,
device_id=device_id,
num_dev=num_dev)
except ImportError:
logger.warning("pycuda error, can't run via GPU")
gpuflag = False
except ImportError:
logger.warning("nervanagpu not found, can't run via GPU")
gpuflag = False
if gpuflag is False:
raise RuntimeError("Can't find CUDA capable GPU")
elif nrv:
nrv = False
try:
from umd.nrv_backend import NRVBackend
nrv = True
except ImportError:
logger.warning("Nervana Engine system software not found")
if flexpoint:
logger.warning("Flexpoint(TM) backend not currently available")
if nrv:
be_name = 'NRV'
be = NRVBackend(rng_seed=rng_seed,
seterr_handling=numerr_handling,
device_id=device_id)
elif not gpuflag:
be_name = 'CPU'
be = CPU(rng_seed=rng_seed, seterr_handling=numerr_handling)
logger.info("{} backend, RNG seed: {}, numerr: {}".format(
be_name, rng_seed, numerr_handling))
return be
def test_cpu_bprop(self):
backend = CPU(rng_seed=0)
layer = self.create_layer(backend=backend)
check_bprop(layer, backend)
def __init__(self):
# this code gets called prior to each test
self.be = CPU()
def test_argmax_axis1(self):
be = CPU()
tsr = be.array([[-1, 10], [11, 9]])
out = be.empty((2, ))
be.argmax(tsr, 1, out)
assert_tensor_equal(out, CPUTensor([1, 0]))
def test_argmin_axis0(self):
be = CPU()
tsr = be.array([[-1, 0], [1, 92]])
out = be.empty((1, 2))
be.argmin(tsr, 0, out)
assert_tensor_equal(out, CPUTensor([[0, 0]]))
