def stabilize(operand):
scalar_constant = 4.0
f = Constant.scalar(sanitize_dtype_cntk(np.float32), scalar_constant);
fInv = Constant.scalar(sanitize_dtype_cntk(np.float32), 1.0 / scalar_constant)
beta = element_times(fInv, log(Constant.scalar(sanitize_dtype_cntk(np.float32), 1.0) + exp(element_times(f, parameter(shape=(), value=0.99537863)))))
return element_times(beta, operand)
def create_model(base_model_file, feature_node_name, last_hidden_node_name, num_classes, input_features, freeze=False):
# Load the pretrained classification net and find nodes
base_model = load_model(base_model_file)
feature_node = None
last_node = None
# feature node
for n in cntk.logging.graph.depth_first_search(base_model, (lambda x: True)):
if (n.name.strip() == feature_node_name) or (n.uid.strip() == feature_node_name):
feature_node = n
print("feature node:", feature_node)
if feature_node is None:
raise Exception("Failed to locate feature node: " + feature_node_name)
# last hidden node
for n in cntk.logging.get_node_outputs(base_model):
if (n.name.strip() == last_hidden_node_name) or (n.uid.strip() == last_hidden_node_name):
last_node = n
print("last hidden node:", last_node)
if last_node is None:
raise Exception("Failed to locate last hidden node: " + last_hidden_node_name)
# Clone the desired layers with fixed weights
cloned_layers = combine([last_node.owner]).clone(
CloneMethod.freeze if freeze else CloneMethod.clone,
{feature_node: placeholder(name='features')})
# Add new dense layer for class prediction
feat_norm = input_features - Constant(114)
cloned_out = cloned_layers(feat_norm)
z = Dense(num_classes, activation=None, name=new_output_node_name) (cloned_out)
return z
def test_op_reduce_mean_all_constant(input_data, axis, device_id, precision):
dt = PRECISION_TO_TYPE[precision]
value = AA(input_data, dtype=dt)
from .. import reduce_mean
from cntk import Axis, Constant
a = Constant(value, name='a')
input_op = reduce_mean(a, axis=Axis.all_axes())
expected_forward = AA(np.mean(value))
actual_forward = input_op.eval()
assert np.allclose(actual_forward, expected_forward)
def test_op_reduce_mean_all_constant(input_data, axis, device_id, precision):
# dt = PRECISION_TO_TYPE[precision]
# FIXME: we'd like to do dt = PRECISION_TO_TYPE[precision]
# however there seems to be an issue with actual_forward below
# that gets computed correctly but by the time np.allclose executes
# it contains garbage values. The problem goes away if one uses
# actual_forward = np.copy(input_op.eval())
dt = np.float32
value = AA(input_data, dtype=dt)
from .. import reduce_mean
from cntk import Axis, Constant
a = Constant(value, name='a')
input_op = reduce_mean(a, axis=Axis.all_axes())
expected_forward = AA(np.mean(value))
actual_forward = input_op.eval()
assert np.allclose(actual_forward, expected_forward)
def create_model(base_model_file, feature_node_name, last_hidden_node_name, num_classes, input_features, freeze=False):
# Load the pretrained classification net and find nodes
base_model = load_model(base_model_file)
feature_node = find_by_name(base_model, feature_node_name)
last_node = find_by_name(base_model, last_hidden_node_name)
# Clone the desired layers with fixed weights
cloned_layers = combine([last_node.owner]).clone(
CloneMethod.freeze if freeze else CloneMethod.clone,
{feature_node: placeholder(name='features')})
# Add new dense layer for class prediction
feat_norm = input_features - Constant(114)
cloned_out = cloned_layers(feat_norm)
z = Dense(num_classes, activation=None, name=new_output_node_name) (cloned_out)
return z
def FU(x):
return UF(Constant(1), x)