def handle_cross_entropy(node, name, c_name):
new_tmp = []
nm = nodes.handler.entitiesHandler.node_map
outer_inputs = []
for node_name in nodes.handler.entitiesHandler.node_map.keys():
if (name in node_name.split("/") or all(x in node_name.split("/"))
for x in name.split("/")) and 'gradients' not in node_name:
for inp in nm[node_name].get_inputs():
if name not in inp.get_name():
outer_inputs.append(inp)
while outer_inputs != []:
for inp in outer_inputs:
#print("Searching for ",inp.get_name()," operation ",inp.get_op())
if inp.get_op() == "Placeholder":
for dim in inp.get_attr()["shape"].shape.dim:
if dim.size != -1:
if dim.size > 2:
return cost_function(
c_name, categorical_cross_entropy(name, node))
else:
return cost_function(
c_name, binary_cross_entropy(name, node))
print(
"ERROR:Found placeholder,but the sizes are all negatives.")
return cost_function(c_name,
categorical_cross_entropy(name, node))
elif inp.get_op(
) in nodes.handler.entitiesHandler.intermediate_operations:
for inp_ in inp.get_inputs():
new_tmp.append(inp_)
outer_inputs = new_tmp
new_tmp = []
return ""
def handle_mul_as_cross_entropy(node, c_name):
children = node.get_inputs()
log = False
placeholder = False
type = None
for elem in children:
if elem.get_op() == "Log":
log = True
if elem.get_op() == "Placeholder":
placeholder = True
for dim in elem.get_attr()["shape"].shape.dim:
if dim.size != -1:
if dim.size > 2:
type = "c"
else:
type = "b"
if log == True and placeholder == True:
if type == "b":
return cost_function(c_name,
binary_cross_entropy(node.get_name(), node))
else:
return cost_function(
c_name, categorical_cross_entropy(node.get_name(), node))
else:
return ""
def handle_sparse_cross_entropy(node, c_name):
for dim in node.get_output():
for i, elem in enumerate(dim.dim):
if elem.size != -1:
if elem.size > 2:
return cost_function(
c_name,
categorical_cross_entropy(node.get_name(), node))
else:
return cost_function(
c_name, binary_cross_entropy(node.get_name(), node))
def handle_log(node):
return cost_function("_cost",
categorical_cross_entropy(node.get_name(), node))
def handle_sigmoid_entropy(node, c_name, name):
#TODO SIGMOID
return cost_function(c_name, categorical_cross_entropy(name, node, True))