def create_objects(root_yaml,
be_type='gpu',
batch_size=128,
rng_seed=None,
device_id=0,
default_dtype=np.float32,
stochastic_rounding=False):
"""
Instantiate objects as per the given specifications.
Arguments:
root_yaml (dict): Model definition dictionary parse from YAML file
be_type (str): backend either 'gpu', 'mgpu' or 'cpu'
rng_seed (None or int): random number generator seed
device_id (int): for GPU backends id of device to use
default_dtype (type): numpy data format for default data types,
stochastic_rounding (bool or int): number of bits for stochastic rounding
use False for no rounding
Returns:
tuple: Contains model, cost and optimizer objects.
"""
assert NervanaObject.be is not None, 'Must generate a backend before running this function'
# can give filename or parse dictionary
if type(root_yaml) is str:
with open(root_yaml, 'r') as fid:
root_yaml = yaml.safe_load(fid.read())
# in case references were used
root_yaml = deepcopy(root_yaml)
# initialize layers
yaml_layers = root_yaml['layers']
# currently only support sequential in yaml
layer_dict = {'layers': yaml_layers}
layers = Sequential.gen_class(layer_dict)
# initialize model
model = Model(layers=layers)
# cost (before layers for shortcut derivs)
cost_name = root_yaml['cost']
cost = GeneralizedCost.gen_class({'costfunc': {'type': cost_name}})
# create optimizer
opt = None
if 'optimizer' in root_yaml:
yaml_opt = root_yaml['optimizer']
typ = yaml_opt['type']
opt = getattr(neon.optimizers, typ).gen_class(yaml_opt['config'])
return model, cost, opt
def test_gan_container(backend_default):
"""
Set up a GenerativeAdversarial container and make sure generator
and discriminator layers get configured correctly.
"""
init_norm = Gaussian(loc=0.0, scale=0.01)
# set up container and ensure layers get wired up correctly
generator = Sequential([Affine(nout=10, init=init_norm), Affine(nout=100, init=init_norm)])
discriminator = Sequential([Affine(nout=100, init=init_norm), Affine(nout=1, init=init_norm)])
layers = GenerativeAdversarial(generator, discriminator)
assert len(layers.layers) == 4
assert layers.layers[0].nout == 10
assert layers.layers[1].nout == 100
assert layers.layers[2].nout == 100
assert layers.layers[3].nout == 1
assert layers.generator.layers == layers.layers[0:2]
assert layers.discriminator.layers == layers.layers[2:4]
def sub_handler(layers, flags, stacks, this_model):
head, layers = split_merge_layers(layers)
if len(layers) > 2:
left = layers[0]
right = sequential(
layers=(
MergeSum(layers[1:]),
Activation(neon.transforms.Normalizer(divisor=-1))))
network = Sequential(layers=head+(MergeSum(layers=(left, right)),))
elif len(layers) == 2:
left = layers[0]
right = sequential(
layers=(
layers[1],
Activation(neon.transforms.Normalizer(divisor=-1))))
network = Sequential(layers=head+(MergeSum(layers=(left, right)),))
else:
network = layers[0]
return network
def sequential(layers):
a = ()
for t in layers:
if type(t) == Sequential:
a += tuple(t.layers)
else:
a += (t, )
res=Sequential(layers=a)
#print 'in_shape:',a[0].in_shape
#res.configure(a[0].in_shape)
return res
def split_merge_layers(layers):
bs = []
ls = []
for layer in layers:
if type(layer) == BranchNode:
b = layer
l = SkipNode()
elif type(layer.layers[0]) == BranchNode:
b = layer.layers[0]
l = Sequential(tuple(layer.layers[1:]))
else:
b = None
l = None
bs += [b]
ls += [l]
bset = set(bs)-set({None})
if len(bset) > 1:
print bset
assert len(bset) <= 1
if len(bset) == 1:
for b in bset:
pass
print 'bs:', bs
print 'ls:', ls
head = ()
for i, layer in enumerate(layers):
if ls[i] is None:
ll = split_list(layers[i].layers, b)
assert len(ll) <= 2
if len(ll) == 2:
assert head == ()
head, l = ll
head = tuple(head)
l = tuple(l)
ls[i] = l
else:
ls[i] = layers[i].layers
print 'bs:', bs
print 'head:', head
print 'ls:', ls
return head, tuple(ls)
def add_handler(layers, flags, stacks, this_model):
head, ls = split_merge_layers(layers)
return Sequential(layers=head+(MergeSum(ls),))
def concat_handler(layers, flags, stacks, this_model):
head, ls = split_merge_layers(layers)
return Sequential(layers=head+(MergeBroadcast(ls, merge="depth"),))