def create_frcn_model(frcn_fine_tune=False):
b1 = BranchNode(name="b1")
imagenet_layers = add_vgg_layers()
HW = (7, 7)
frcn_layers = [
RoiPooling(layers=imagenet_layers, HW=HW,
bprop_enabled=frcn_fine_tune),
Affine(nout=4096,
init=Gaussian(scale=0.005),
bias=Constant(.1),
activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=4096,
init=Gaussian(scale=0.005),
bias=Constant(.1),
activation=Rectlin()),
Dropout(keep=0.5), b1,
Affine(nout=21,
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Softmax())
]
bb_layers = [
b1,
Affine(nout=84,
init=Gaussian(scale=0.001),
bias=Constant(0),
activation=Identity())
]
return Model(layers=Tree([frcn_layers, bb_layers]))
def __init__(self,
overlapping_classes=None,
exclusive_classes=None,
analytics_input=True,
network_type='conv_net',
num_words=60,
width=100,
lookup_size=0,
lookup_dim=0,
optimizer=Adam()):
assert (overlapping_classes is not None) or (exclusive_classes
is not None)
self.width = width
self.num_words = num_words
self.overlapping_classes = overlapping_classes
self.exclusive_classes = exclusive_classes
self.analytics_input = analytics_input
self.recurrent = network_type == 'lstm'
self.lookup_size = lookup_size
self.lookup_dim = lookup_dim
init = GlorotUniform()
activation = Rectlin(slope=1E-05)
gate = Logistic()
input_layers = self.input_layers(analytics_input, init, activation,
gate)
if self.overlapping_classes is None:
output_layers = [
Affine(len(self.exclusive_classes), init, activation=Softmax())
]
elif self.exclusive_classes is None:
output_layers = [
Affine(len(self.overlapping_classes),
init,
activation=Logistic())
]
else:
output_branch = BranchNode(name='exclusive_overlapping')
output_layers = Tree([[
SkipNode(), output_branch,
Affine(len(self.exclusive_classes), init, activation=Softmax())
],
[
output_branch,
Affine(len(self.overlapping_classes),
init,
activation=Logistic())
]])
layers = [
input_layers,
# this is where inputs meet, and where we may want to add depth or
# additional functionality
Dropout(keep=0.8),
output_layers
]
super(ClassifierNetwork, self).__init__(layers, optimizer=optimizer)
def make_tree(trunk, branch1, branch2, alphas):
# Make one copy that is the Tree version
_trunk = [l['layer'](**l['config']) for l in trunk]
bnode = [BranchNode(name='bnode')]
_branch1 = [l['layer'](**l['config']) for l in branch1]
_branch2 = [l['layer'](**l['config']) for l in branch2]
v1 = Tree([_trunk + bnode + _branch1, bnode + _branch2], alphas)
# Now a second copy with no sharing as the reference version
_trunkb = [l['layer'](**l['config']) for l in trunk]
_branch1b = [l['layer'](**l['config']) for l in branch1]
_branch2b = [l['layer'](**l['config']) for l in branch2]
return (v1, _trunkb, _branch1b, _branch2b)
Affine(nout=16, linear_name="b1_l1", **normrelu),
Affine(nout=10, linear_name="b1_l2", **normsigm)]
p3 = [b2,
Affine(nout=16, linear_name="b2_l1", **normrelu),
Affine(nout=10, linear_name="b2_l2", **normsigm)]
# setup cost function as CrossEntropy
cost = Multicost(costs=[GeneralizedCost(costfunc=CrossEntropyMulti()),
GeneralizedCost(costfunc=CrossEntropyBinary()),
GeneralizedCost(costfunc=CrossEntropyBinary())],
weights=[1, 0., 0.])
# setup optimizer
optimizer = GradientDescentMomentum(0.1, momentum_coef=0.9, stochastic_round=args.rounding)
# initialize model object
alphas = [1, 0.25, 0.25]
mlp = Model(layers=Tree([p1, p2, p3], alphas=alphas))
# setup standard fit callbacks
callbacks = Callbacks(mlp, train_set, eval_set=valid_set, **args.callback_args)
# run fit
mlp.fit(train_set, optimizer=optimizer, num_epochs=args.epochs, cost=cost, callbacks=callbacks)
logging.getLogger('neon').info("Misclassification error = %.1f%%",
(mlp.eval(valid_set, metric=Misclassification())*100))
print('Misclassification error = %.1f%%' % (mlp.eval(valid_set, metric=Misclassification())*100))
def build_model(dataset,
frcn_rois_per_img,
train_pre_nms_N=12000,
train_post_nms_N=2000,
test_pre_nms_N=6000,
test_post_nms_N=300,
inference=False):
"""
Returns the Faster-RCNN model. For inference, also returns a reference to the
proposal layer.
Faster-RCNN contains three modules: VGG, the Region Proposal Network (RPN),
and the Classification Network (ROI-pooling + Fully Connected layers), organized
as a tree. Tree has 4 branches:
VGG -> b1 -> Conv (3x3) -> b2 -> Conv (1x1) -> CrossEntropyMulti (objectness label)
b2 -> Conv (1x1) -> SmoothL1Loss (bounding box targets)
b1 -> PropLayer -> ROI -> Affine -> Affine -> b3 -> Affine -> CrossEntropyMulti
b3 -> Affine -> SmoothL1Loss
When the model is constructed for inference, several elements are different:
- The number of regions to keep before and after non-max suppression is (6000, 300) for
training and (12000, 2000) for inference.
- The out_shape of the proposalLayer of the network is equal to post_nms_N (number of rois
to keep after performaing nms). This is configured by passing the inference flag to the
proposalLayer constructor.
Arguments:
dataset (objectlocalization): Dataset object.
frcn_rois_per_img (int): Number of ROIs per image considered by the classification network.
inference (bool): Construct the model for inference. Default is False.
Returns:
model (Model): Faster-RCNN model.
proposalLayer (proposalLayer): Reference to proposalLayer in the model.
Returned only for inference=True.
"""
num_classes = dataset.num_classes
# define the branch points
b1 = BranchNode(name="conv_branch")
b2 = BranchNode(name="rpn_branch")
b3 = BranchNode(name="roi_branch")
# define VGG
VGG = util.add_vgg_layers()
# define RPN
rpn_init = dict(strides=1, init=Gaussian(scale=0.01), bias=Constant(0))
# these references are passed to the ProposalLayer.
RPN_3x3 = Conv((3, 3, 512), activation=Rectlin(), padding=1, **rpn_init)
RPN_1x1_obj = Conv((1, 1, 18),
activation=PixelwiseSoftmax(c=2),
padding=0,
**rpn_init)
RPN_1x1_bbox = Conv((1, 1, 36),
activation=Identity(),
padding=0,
**rpn_init)
# inference uses different network settings
if not inference:
pre_nms_N = train_pre_nms_N # default 12000
post_nms_N = train_post_nms_N # default 2000
else:
pre_nms_N = test_pre_nms_N # default 6000
post_nms_N = test_post_nms_N # default 300
proposalLayer = ProposalLayer([RPN_1x1_obj, RPN_1x1_bbox],
dataset,
pre_nms_N=pre_nms_N,
post_nms_N=post_nms_N,
num_rois=frcn_rois_per_img,
inference=inference)
# define ROI classification network
ROI = [
proposalLayer,
RoiPooling(HW=(7, 7)),
Affine(nout=4096,
init=Gaussian(scale=0.005),
bias=Constant(.1),
activation=Rectlin()),
Dropout(keep=0.5),
Affine(nout=4096,
init=Gaussian(scale=0.005),
bias=Constant(.1),
activation=Rectlin()),
Dropout(keep=0.5)
]
ROI_category = Affine(nout=num_classes,
init=Gaussian(scale=0.01),
bias=Constant(0),
activation=Softmax())
ROI_bbox = Affine(nout=4 * num_classes,
init=Gaussian(scale=0.001),
bias=Constant(0),
activation=Identity())
# build the model
# the four branches of the tree mirror the branches listed above
frcn_tree = Tree([ROI + [b3, ROI_category], [b3, ROI_bbox]])
model = Model(layers=Tree([
VGG + [b1, RPN_3x3, b2, RPN_1x1_obj],
[b2, RPN_1x1_bbox],
[b1] + [frcn_tree],
]))
if inference:
return (model, proposalLayer)
else:
return model
GeneralizedCost(costfunc=CrossEntropyMulti()),
GeneralizedCost(costfunc=CrossEntropyMulti()),
GeneralizedCost(costfunc=CrossEntropyMulti())
],
weights=[1, 0., 0.]) # We only want to consider the CE of the main path
if not args.resume:
# build the model from scratch and run it
# Now construct the model
branch_nodes = [BranchNode(name='branch{}'.format(i)) for i in (0, 1)]
main1 = main_branch(branch_nodes)
aux1 = aux_branch(branch_nodes[0], ind=1)
aux2 = aux_branch(branch_nodes[1], ind=2)
model = Model(layers=Tree([main1, aux1, aux2], alphas=[1.0, 0.3, 0.3]))
else:
# load up the save model
model = Model('serialize_test_2.pkl')
model.initialize(train, cost=cost)
# configure callbacks
callbacks = Callbacks(model,
progress_bar=True,
output_file='temp1.h5',
serialize=1,
history=3,
save_path='serialize_test.pkl')
lr_sched = PolySchedule(total_epochs=10, power=0.5)
bb_layers = [
b1,
bbox_pred,
]
# setup optimizer
opt_w = GradientDescentMomentum(0.001 * learning_rate_scale,
0.9,
wdecay=0.0005)
opt_b = GradientDescentMomentum(0.002 * learning_rate_scale, 0.9)
optimizer = MultiOptimizer({'default': opt_w, 'Bias': opt_b})
# setup model
model = Model(layers=Tree([frcn_layers, bb_layers]))
# if training a new model, seed the Alexnet conv layers with pre-trained weights
# otherwise, just load the model file
if args.model_file is None:
load_imagenet_weights(model, args.data_dir)
cost = Multicost(costs=[
GeneralizedCost(costfunc=CrossEntropyMulti()),
GeneralizedCostMask(costfunc=SmoothL1Loss())
],
weights=[1, 1])
callbacks = Callbacks(model, **args.callback_args)
model.fit(train_set,
def test_branch_model_fork():
from neon.layers import BranchNode, Tree
NervanaObject.be = gen_backend("gpu", batch_size=64)
be = NervanaObject.be
bnode = BranchNode()
i1 = inception([(32, ), (32, 32), ('max', 16)])
top1 = top_branch()
top2 = top_branch()
p1 = Sequential(main_branch() + [bnode, i1] + top1)
p2 = [bnode] + top2
alpha2 = 0.3
neon_layer = Tree([p1, p2], alphas=[1.0, alpha2])
inshape = (3, 224, 224)
insize = np.prod(inshape)
inpa = np.random.random((insize, batch_size))
neon_layer.configure(inshape)
inp = neon_layer.be.array(inpa)
neon_layer.allocate()
print neon_layer.nested_str()
neon_layer.layers[0].layers[0].prev_layer = True
neon_layer.allocate_deltas()
neon_layer.layers[0].layers[0].set_deltas([be.iobuf(inshape)])
neon_out_dev = neon_layer.fprop(inp)
neon_out = [d.get() for d in neon_out_dev]
# Now make the reference pathways:
main_trunk2 = Sequential(main_branch())
main_trunk2.configure(inshape)
main2 = main_trunk2.layers
main2[0].prev_layer = True
main2[0].set_deltas([be.iobuf(inshape)])
branch2 = Sequential(top_branch())
lbranch2 = branch2.layers
(b1, b2, b3) = inception_bare(i1, [(32, ), (32, 32), ('max', 16)])
for bb in (b1, b2, b3, lbranch2):
oshape = inshape
for ll in main2 + bb:
oshape = ll.configure(oshape)
main1_trunk = neon_layer.layers[0].layers[:8]
for ll, lo in zip(main2, main1_trunk):
if ll.has_params:
ll.set_params({'params': {'W': lo.W.get()}})
ll.allocate()
ll.set_deltas([be.iobuf(ll.in_shape)])
for ll, lo in zip(lbranch2, neon_layer.layers[1].layers[1:]):
if ll.has_params:
ll.set_params({'params': {'W': lo.W.get()}})
for bb in (b1, b2, b3, lbranch2):
for ll in bb:
ll.allocate()
ll.set_deltas([be.iobuf(ll.in_shape)])
# Create the combined output buffer
merge_output = be.empty_like(neon_layer.layers[0].layers[9].outputs)
x = inp
for ll in main2:
x = ll.fprop(x)
main2_out = x
start = 0
for bb in (b1, b2, b3):
xb = main2_out
for ll in bb:
xb = ll.fprop(xb)
end = start + xb.shape[0]
merge_output[start:end] = xb
start = end
x = merge_output
top_trunk = Sequential(top1).layers
for ll in top_trunk:
x = ll.fprop(x)
neon_out_ref = x.get()
difference = neon_out_ref - neon_out[0]
assert np.max(np.abs(difference)) < 1e-7
print np.max(np.abs(difference))
# Now do second branch
neon_out_ref2 = branch2.fprop(main2_out).get()
difference = neon_out_ref2 - neon_out[1]
assert np.max(np.abs(difference)) < 1e-7
print np.max(np.abs(difference))
print "Beginning Back prop"
erra = [np.random.random(d.shape) for d in neon_out]
err = [be.array(d) for d in erra]
neon_layer.layers[0].layers[0].deltas = be.iobuf(inshape)
neon_layer.bprop(err)
bottom_neon_deltas = neon_layer.layers[0].layers[1].deltas.get()
middle_neon_deltas = neon_layer.layers[1].layers[1].deltas.get()
err0 = err[0]
for ll in reversed(top_trunk):
err0 = ll.bprop(err0)
err1 = err[1]
for ll in reversed(lbranch2):
err1 = ll.bprop(err1)
for bb, errb in zip((b1, b2, b3),
neon_layer.layers[0].layers[-5].error_views):
for ll in reversed(bb):
errb = ll.bprop(errb)
# Now sum up the deltas at the root of the branch layer and compare
ref_deltas = be.zeros_like(b1[0].deltas)
ref_deltas[:] = b1[0].deltas + b2[0].deltas + b3[
0].deltas + alpha2 * lbranch2[0].deltas
neon_ref_deltas = ref_deltas.get()
difference = middle_neon_deltas - neon_ref_deltas
print np.max(np.abs(difference))
assert np.max(np.abs(difference)) < 1e-8
x = ref_deltas
main2[0].deltas = be.iobuf(inshape)
for ll in reversed(main2):
x = ll.bprop(x)
bottom_neon_ref_deltas = main2[1].deltas.get()
difference = bottom_neon_deltas - bottom_neon_ref_deltas
print np.max(np.abs(difference))
assert np.max(np.abs(difference)) < 1e-8
def test_branch_model_fork_cpu(backend_cpu64):
from neon.layers import BranchNode, Tree
np.random.seed(0)
be = NervanaObject.be
be.bsz = 32
bnode = BranchNode()
i1 = inception([(32,), (32, 32), ('max', 16)])
top1 = top_branch()
top2 = top_branch()
p1 = Sequential(main_branch() + [bnode, i1] + top1)
p2 = [bnode] + top2
alpha2 = 0.3
neon_layer = Tree([p1, p2], alphas=[1.0, alpha2])
inshape = (4, 224, 224)
insize = np.prod(inshape)
inpa = np.random.random((insize, batch_size))
neon_layer.configure(inshape)
inp = neon_layer.be.array(inpa)
neon_layer.allocate()
neon_layer.layers[0].layers[0].prev_layer = True
neon_layer.allocate_deltas()
neon_out_dev = neon_layer.fprop(inp)
neon_out = [d.get() for d in neon_out_dev]
# Now make the reference pathways:
main_trunk2 = Sequential(main_branch())
main_trunk2.configure(inshape)
main2 = main_trunk2.layers
main2[0].prev_layer = True
main2[0].deltas = be.iobuf(inshape)
branch2 = Sequential(top_branch())
lbranch2 = branch2.layers
(b1, b2, b3) = inception_bare(i1, [(32,), (32, 32), ('max', 16)])
for bb in (b1, b2, b3, lbranch2):
oshape = inshape
for ll in main2 + bb:
oshape = ll.configure(oshape)
main1_trunk = neon_layer.layers[0].layers[:8]
for ll, lo in zip(main2, main1_trunk):
if ll.has_params:
ll.set_params({'params': {'W': lo.W.get()}})
ll.allocate()
temp_deltas = DeltasTree()
temp_deltas.proc_layer(ll)
temp_deltas.allocate_buffers()
ll.set_deltas(temp_deltas)
for ll, lo in zip(lbranch2, neon_layer.layers[1].layers[1:]):
if ll.has_params:
ll.set_params({'params': {'W': lo.W.get()}})
for bb in (b1, b2, b3, lbranch2):
for ll in bb:
ll.allocate()
temp_deltas = DeltasTree()
temp_deltas.proc_layer(ll)
temp_deltas.allocate_buffers()
ll.set_deltas(temp_deltas)
# Create the combined output buffer
merge_output = be.empty_like(neon_layer.layers[0].layers[9].outputs)
x = inp
for ll in main2:
x = ll.fprop(x)
main2_out = x
start = 0
for bb in (b1, b2, b3):
xb = main2_out
for ll in bb:
xb = ll.fprop(xb)
end = start + xb.shape[0]
merge_output[start:end] = xb
start = end
x = merge_output
top_trunk = Sequential(top1).layers
for ll in top_trunk:
x = ll.fprop(x)
neon_out_ref = x.get()
assert allclose_with_out(neon_out_ref, neon_out[0], rtol=0)
# Now do second branch
neon_out_ref2 = branch2.fprop(main2_out).get()
assert allclose_with_out(neon_out_ref2, neon_out[1])
neon_logger.display("Beginning Back prop")
erra = [np.random.random(d.shape) for d in neon_out]
err = [be.array(d) for d in erra]
neon_layer.layers[0].layers[0].deltas = be.iobuf(inshape)
neon_layer.bprop(err)
bottom_neon_deltas = neon_layer.layers[0].layers[1].deltas.get()
middle_neon_deltas = neon_layer.layers[1].layers[1].deltas.get()
err0 = err[0]
for ll in reversed(top_trunk):
err0 = ll.bprop(err0)
err1 = err[1]
for ll in reversed(lbranch2):
err1 = ll.bprop(err1)
for bb, errb in zip((b1, b2, b3), neon_layer.layers[0].layers[-5].error_views):
for ll in reversed(bb):
errb = ll.bprop(errb)
# Now sum up the deltas at the root of the branch layer and compare
ref_deltas = be.zeros_like(b1[0].deltas)
ref_deltas[:] = alpha2 * lbranch2[0].deltas
ref_deltas[:] = ref_deltas + b3[0].deltas + b2[0].deltas + b1[0].deltas
neon_ref_deltas = ref_deltas.get()
assert allclose_with_out(middle_neon_deltas, neon_ref_deltas, rtol=0)
x = ref_deltas
main2[0].deltas = be.iobuf(inshape)
for ll in reversed(main2):
x = ll.bprop(x)
bottom_neon_ref_deltas = main2[1].deltas.get()
assert allclose_with_out(bottom_neon_deltas, bottom_neon_ref_deltas, rtol=0)
def test_branch_model_fork(backend_gpu):
from neon.layers import BranchNode, Tree
np.random.seed(0)
be = NervanaObject.be
if be.gpu_memory_size < 6.1 * 1024 * 1024 * 1024:
pytest.skip(msg='Test requires more than 6.1GB')
be.bsz = 64
bnode = BranchNode()
i1 = inception([(32,), (32, 32), ('max', 16)])
top1 = top_branch()
top2 = top_branch()
p1 = Sequential(main_branch() + [bnode, i1] + top1)
p2 = [bnode] + top2
alpha2 = 0.3
neon_layer = Tree([p1, p2], alphas=[1.0, alpha2])
inshape = (4, 224, 224)
insize = np.prod(inshape)
inpa = np.random.random((insize, batch_size))
neon_layer.configure(inshape)
inp = neon_layer.be.array(inpa)
neon_layer.allocate()
neon_layer.layers[0].layers[0].prev_layer = True
neon_layer.allocate_deltas()
neon_out_dev = neon_layer.fprop(inp)
neon_out = [d.get() for d in neon_out_dev]
# Now make the reference pathways:
main_trunk2 = Sequential(main_branch())
main_trunk2.configure(inshape)
main2 = main_trunk2.layers
main2[0].prev_layer = True
main2[0].deltas = be.iobuf(inshape)
branch2 = Sequential(top_branch())
lbranch2 = branch2.layers
(b1, b2, b3) = inception_bare(i1, [(32,), (32, 32), ('max', 16)])
for bb in (b1, b2, b3, lbranch2):
oshape = inshape
for ll in main2 + bb:
oshape = ll.configure(oshape)
main1_trunk = neon_layer.layers[0].layers[:8]
for ll, lo in zip(main2, main1_trunk):
if ll.has_params:
ll.set_params({'params': {'W': lo.W.get()}})
ll.allocate()
temp_deltas = DeltasTree()
temp_deltas.proc_layer(ll)
temp_deltas.allocate_buffers()
ll.set_deltas(temp_deltas)
for ll, lo in zip(lbranch2, neon_layer.layers[1].layers[1:]):
if ll.has_params:
ll.set_params({'params': {'W': lo.W.get()}})
for bb in (b1, b2, b3, lbranch2):
for ll in bb:
ll.allocate()
temp_deltas = DeltasTree()
temp_deltas.proc_layer(ll)
temp_deltas.allocate_buffers()
ll.set_deltas(temp_deltas)
# Create the combined output buffer
merge_output = be.empty_like(neon_layer.layers[0].layers[9].outputs)
x = inp
for ll in main2:
x = ll.fprop(x)
main2_out = x
start = 0
for bb in (b1, b2, b3):
xb = main2_out
for ll in bb:
xb = ll.fprop(xb)
end = start + xb.shape[0]
merge_output[start:end] = xb
start = end
x = merge_output
top_trunk = Sequential(top1).layers
for ll in top_trunk:
x = ll.fprop(x)
neon_out_ref = x.get()
assert allclose_with_out(neon_out_ref, neon_out[0], rtol=0)
# Now do second branch
neon_out_ref2 = branch2.fprop(main2_out).get()
assert allclose_with_out(neon_out_ref2, neon_out[1])
neon_logger.display("Beginning Back prop")
erra = [np.random.random(d.shape) for d in neon_out]
err = [be.array(d) for d in erra]
neon_layer.layers[0].layers[0].deltas = be.iobuf(inshape)
neon_layer.bprop(err)
bottom_neon_deltas = neon_layer.layers[0].layers[1].deltas.get()
middle_neon_deltas = neon_layer.layers[1].layers[1].deltas.get()
err0 = err[0]
for ll in reversed(top_trunk):
err0 = ll.bprop(err0)
err1 = err[1]
for ll in reversed(lbranch2):
err1 = ll.bprop(err1)
for bb, errb in zip((b1, b2, b3), neon_layer.layers[0].layers[-5].error_views):
for ll in reversed(bb):
errb = ll.bprop(errb)
# Now sum up the deltas at the root of the branch layer and compare
ref_deltas = be.zeros_like(b1[0].deltas)
ref_deltas[:] = alpha2 * lbranch2[0].deltas
ref_deltas[:] = ref_deltas + b3[0].deltas + b2[0].deltas + b1[0].deltas
neon_ref_deltas = ref_deltas.get()
assert allclose_with_out(middle_neon_deltas, neon_ref_deltas, rtol=0)
x = ref_deltas
main2[0].deltas = be.iobuf(inshape)
for ll in reversed(main2):
x = ll.bprop(x)
bottom_neon_ref_deltas = main2[1].deltas.get()
assert allclose_with_out(bottom_neon_deltas, bottom_neon_ref_deltas, rtol=0)
}),
)), ('m_l3', 0, 0, 0, 0, 0, {}), (share, 'shared', (
(0, 16, 0, 0, 'b2_l1', 0, {'dense'}),
(0, 10, 0, 0, 'b2_l2', 0, {
'dense': True,
'nonlinearity': Logistic(shortcut=True),
'equal': ['target', 'b2', CrossEntropyBinary()]
}),
)), ('m_l4', )), {'target': l_y})
cost = GeneralizedCost(costfunc=CrossEntropyBinary())
cost, layers, tagslice = deepstacks.neon.get_loss(errors, watchpoints, cost)
print layers
network = Tree([network] + layers)
layers = [network]
inputs = deepstacks.neon.get_inputs(network)
targets = deepstacks.neon.get_targets(cost)
print inputs, targets
#assert tuple(inputs)==('image',)
#p1 = [l_in,
# Affine(nout=100, name="m_l1", **normrelu),
# b1,
# Affine(nout=32, name="m_l2", **normrelu),
# Affine(nout=16, name="m_l3", **normrelu),
# b2,
def test_branch_model_fork():
from neon.layers import BranchNode, Tree
NervanaObject.be = gen_backend("gpu", batch_size=64)
be = NervanaObject.be
bnode = BranchNode()
i1 = inception([(32,), (32, 32), ('max', 16)])
top1 = top_branch()
top2 = top_branch()
p1 = Sequential(main_branch() + [bnode, i1] + top1)
p2 = [bnode] + top2
alpha2 = 0.3
neon_layer = Tree([p1, p2], alphas=[1.0, alpha2])
inshape = (3, 224, 224)
insize = np.prod(inshape)
inpa = np.random.random((insize, batch_size))
neon_layer.configure(inshape)
inp = neon_layer.be.array(inpa)
neon_layer.allocate()
print neon_layer.nested_str()
neon_layer.layers[0].layers[0].prev_layer = True
neon_layer.allocate_deltas()
neon_layer.layers[0].layers[0].set_deltas([be.iobuf(inshape)])
neon_out_dev = neon_layer.fprop(inp)
neon_out = [d.get() for d in neon_out_dev]
# Now make the reference pathways:
main_trunk2 = Sequential(main_branch())
main_trunk2.configure(inshape)
main2 = main_trunk2.layers
main2[0].prev_layer = True
main2[0].set_deltas([be.iobuf(inshape)])
branch2 = Sequential(top_branch())
lbranch2 = branch2.layers
(b1, b2, b3) = inception_bare(i1, [(32,), (32, 32), ('max', 16)])
for bb in (b1, b2, b3, lbranch2):
oshape = inshape
for ll in main2 + bb:
oshape = ll.configure(oshape)
main1_trunk = neon_layer.layers[0].layers[:8]
for ll, lo in zip(main2, main1_trunk):
if ll.has_params:
ll.set_params(lo.W.get())
ll.allocate()
ll.set_deltas([be.iobuf(ll.in_shape)])
for ll, lo in zip(lbranch2, neon_layer.layers[1].layers[1:]):
if ll.has_params:
ll.set_params(lo.W.get())
for bb in (b1, b2, b3, lbranch2):
for ll in bb:
ll.allocate()
ll.set_deltas([be.iobuf(ll.in_shape)])
# Create the combined output buffer
merge_output = be.empty_like(neon_layer.layers[0].layers[9].outputs)
x = inp
for ll in main2:
x = ll.fprop(x)
main2_out = x
start = 0
for bb in (b1, b2, b3):
xb = main2_out
for ll in bb:
xb = ll.fprop(xb)
end = start + xb.shape[0]
merge_output[start:end] = xb
start = end
x = merge_output
top_trunk = Sequential(top1).layers
for ll in top_trunk:
x = ll.fprop(x)
neon_out_ref = x.get()
difference = neon_out_ref - neon_out[0]
assert np.max(np.abs(difference)) < 1e-7
print np.max(np.abs(difference))
# Now do second branch
neon_out_ref2 = branch2.fprop(main2_out).get()
difference = neon_out_ref2 - neon_out[1]
assert np.max(np.abs(difference)) < 1e-7
print np.max(np.abs(difference))
print "Beginning Back prop"
erra = [np.random.random(d.shape) for d in neon_out]
err = [be.array(d) for d in erra]
neon_layer.layers[0].layers[0].deltas = be.iobuf(inshape)
neon_layer.bprop(err)
bottom_neon_deltas = neon_layer.layers[0].layers[1].deltas.get()
middle_neon_deltas = neon_layer.layers[1].layers[1].deltas.get()
err0 = err[0]
for ll in reversed(top_trunk):
err0 = ll.bprop(err0)
err1 = err[1]
for ll in reversed(lbranch2):
err1 = ll.bprop(err1)
for bb, errb in zip((b1, b2, b3), neon_layer.layers[0].layers[-5].error_views):
for ll in reversed(bb):
errb = ll.bprop(errb)
# Now sum up the deltas at the root of the branch layer and compare
ref_deltas = be.zeros_like(b1[0].deltas)
ref_deltas[:] = b1[0].deltas + b2[0].deltas + b3[0].deltas + alpha2 * lbranch2[0].deltas
neon_ref_deltas = ref_deltas.get()
difference = middle_neon_deltas - neon_ref_deltas
print np.max(np.abs(difference))
assert np.max(np.abs(difference)) < 1e-8
x = ref_deltas
main2[0].deltas = be.iobuf(inshape)
for ll in reversed(main2):
x = ll.bprop(x)
bottom_neon_ref_deltas = main2[1].deltas.get()
difference = bottom_neon_deltas - bottom_neon_ref_deltas
print np.max(np.abs(difference))
assert np.max(np.abs(difference)) < 1e-8
l_in, (
(0, 100, 0, 0, 0, 0, {'dense', 'nobias'}),
(0, 10, 0, 0, 0, 0, {
'dense': True,
'nonlinearity': Logistic(shortcut=True),
'nobias': True
}),
))
# setup cost function as CrossEntropy
cost = GeneralizedCost(costfunc=CrossEntropyBinary())
cost, extra_layers, tagslice = deepstacks.neon.get_loss(
errors, watchpoints, cost)
network = Tree([network] + extra_layers)
deepstacks.neon.utils.walk(network)
inputs = deepstacks.neon.get_inputs(network)
assert tuple(inputs) == ('image', )
#print network.get_description()
layers = network
# setup optimizer
optimizer = GradientDescentMomentum(0.1,
momentum_coef=0.9,
stochastic_round=args.rounding)