def create_final_layer(finetune_net, custom_network, num_classes, set=1):
'''
Create final sub-network
Args:
finetune_net (network): Initial base network
custom_network (list): List of dicts containing details on appeded layers to base netwoek in transfer learning
num_classes (int): Number of classes in the dataset
set (int): Select the right set to find the details of outermost layer
Returns:
network: Updated base network with appended custom additions
'''
last_layer_name = custom_network[len(custom_network) - 1]["name"]
if (set == 1):
if (last_layer_name == "linear"):
with finetune_net.name_scope():
for i in range(len(custom_network) - 1):
layer = get_layer(custom_network[i])
finetune_net.features.add(layer)
finetune_net.features[len(finetune_net.features) -
1].initialize(init.Xavier(), ctx=ctx)
finetune_net.output = get_layer(
custom_network[len(custom_network) - 1])
finetune_net.output.initialize(init.Xavier(), ctx=ctx)
else:
with finetune_net.name_scope():
for i in range(len(custom_network) - 2):
layer = get_layer(custom_network[i])
finetune_net.features.add(layer)
finetune_net.features[len(finetune_net.features) -
1].initialize(init.Xavier(), ctx=ctx)
finetune_net.output = get_final_layer(
custom_network[len(custom_network) - 2],
activation=custom_network[len(custom_network) - 1]['name'])
finetune_net.output.initialize(init.Xavier(), ctx=ctx)
if (set == 2):
net = nn.HybridSequential()
with net.name_scope():
for i in range(len(custom_network)):
layer = get_layer(custom_network[i])
net.add(layer)
with finetune_net.name_scope():
finetune_net.output = net
finetune_net.output.initialize(init.Xavier(), ctx=ctx)
if (set == 3):
msg = "Custom model addition for - Set 3 models: Not Implemented.\n"
msg += "Set 3 models - {}\n".format(set3)
msg += "Ignoring added layers\n"
ConstraintWarning(msg)
return finetune_net
def create_final_layer(finetune_net, custom_network, num_classes, set=1):
last_layer_name = custom_network[len(custom_network) - 1]["name"]
if (set == 1):
if (last_layer_name == "linear"):
with finetune_net.name_scope():
for i in range(len(custom_network) - 1):
layer = get_layer(custom_network[i])
finetune_net.features.add(layer)
finetune_net.features[len(finetune_net.features) -
1].initialize(init.Xavier(), ctx=ctx)
finetune_net.output = get_layer(
custom_network[len(custom_network) - 1])
finetune_net.output.initialize(init.Xavier(), ctx=ctx)
else:
with finetune_net.name_scope():
for i in range(len(custom_network) - 2):
layer = get_layer(custom_network[i])
finetune_net.features.add(layer)
finetune_net.features[len(finetune_net.features) -
1].initialize(init.Xavier(), ctx=ctx)
finetune_net.output = get_final_layer(
custom_network[len(custom_network) - 2],
activation=custom_network[len(custom_network) - 1]['name'])
finetune_net.output.initialize(init.Xavier(), ctx=ctx)
if (set == 2):
net = nn.HybridSequential()
with net.name_scope():
for i in range(len(custom_network)):
layer = get_layer(custom_network[i])
net.add(layer)
with finetune_net.name_scope():
finetune_net.output = net
finetune_net.output.initialize(init.Xavier(), ctx=ctx)
if (set == 3):
msg = "Custom model addition for - Set 3 models: Not Implemented.\n"
msg += "Set 3 models - {}\n".format(set3)
msg += "Ignoring added layers\n"
ConstraintWarning(msg)
return finetune_net
