def setup_model(system_dict):
if (system_dict["model"]["type"] == "pretrained"):
model_name = system_dict["model"]["params"]["model_name"]
use_pretrained = system_dict["model"]["params"]["use_pretrained"]
freeze_base_network = system_dict["model"]["params"][
"freeze_base_network"]
custom_network = system_dict["model"]["custom_network"]
final_layer = system_dict["model"]["final_layer"]
num_classes = system_dict["dataset"]["params"]["num_classes"]
finetune_net, model_name = get_base_model(model_name, use_pretrained,
num_classes,
freeze_base_network)
if (len(custom_network)):
if (final_layer):
if (model_name in set1):
finetune_net = create_final_layer(finetune_net,
custom_network,
num_classes,
set=1)
elif (model_name in set2):
finetune_net = create_final_layer(finetune_net,
custom_network,
num_classes,
set=2)
elif (model_name in set3):
finetune_net = create_final_layer(finetune_net,
custom_network,
num_classes,
set=3)
else:
print("Final layer not assigned")
return 0
else:
if (model_name in set1):
with finetune_net.name_scope():
finetune_net.output = nn.Dense(
num_classes, weight_initializer=init.Xavier())
finetune_net.output.initialize(init.Xavier(), ctx=ctx)
elif (model_name in set2):
net = nn.HybridSequential()
with net.name_scope():
net.add(
nn.Conv2D(num_classes,
kernel_size=(1, 1),
strides=(1, 1),
weight_initializer=init.Xavier()))
net.add(nn.Flatten())
with finetune_net.name_scope():
finetune_net.output = net
finetune_net.output.initialize(init.Xavier(), ctx=ctx)
elif (model_name in set3):
with finetune_net.name_scope():
finetune_net.fc = nn.Dense(
num_classes, weight_initializer=init.Xavier())
finetune_net.fc.initialize(init.Xavier(), ctx=ctx)
if (not use_pretrained):
finetune_net.initialize(init.Xavier(), ctx=ctx)
system_dict["local"]["model"] = finetune_net
return system_dict
else:
count = []
for i in range(len(names)):
count.append(1)
network_stack = system_dict["custom_model"]["network_stack"]
G = nx.DiGraph()
G.add_node("Net", pos=(1, 1))
sequential_first = "data"
sequential_second, count = get_layer_uid(network_stack[0], count)
count = []
for i in range(len(names)):
count.append(1)
position = 1
G.add_node(sequential_first, pos=(2, 1))
position += 1
net = nn.HybridSequential()
max_width = 1
for i in range(len(network_stack)):
if (type(network_stack[i]) == list):
branch_end_points = []
branch_lengths = []
branches = []
branch_net = []
if (max_width < len(network_stack[i]) - 2):
max_width = len(network_stack[i]) - 2
for j in range(len(network_stack[i]) - 1):
small_net = []
branch_net.append(nn.HybridSequential())
branch_first = sequential_first
branch_position = position
column = j + 2
for k in range(len(network_stack[i][j])):
branch_second, count = get_layer_uid(
network_stack[i][j][k], count)
small_net.append(
custom_model_get_layer(network_stack[i][j][k]))
branch_net[j].add(
custom_model_get_layer(network_stack[i][j][k]))
G.add_node(branch_second,
pos=(column, branch_position))
branch_position += 1
G.add_edge(branch_first, branch_second)
branch_first = branch_second
if (k == len(network_stack[i][j]) - 1):
branch_end_points.append(branch_second)
branch_lengths.append(len(network_stack[i][j]))
branches.append(small_net)
position += max(branch_lengths)
position += 1
sequential_second, count = get_layer_uid(
network_stack[i][-1], count)
if (network_stack[i][-1]["name"] == "concatenate"):
subnetwork = contrib_nn.HybridConcurrent(axis=1)
for j in range(len(network_stack[i]) - 1):
#print(branch_net[j])
subnetwork.add(branch_net[j])
else:
subnetwork = addBlock(branches)
G.add_node(sequential_second, pos=(2, position))
position += 1
for i in range(len(branch_end_points)):
G.add_edge(branch_end_points[i], sequential_second)
sequential_first = sequential_second
net.add(subnetwork)
else:
sequential_second, count = get_layer_uid(
network_stack[i], count)
net.add(custom_model_get_layer(network_stack[i]))
G.add_node(sequential_second, pos=(2, position))
position += 1
G.add_edge(sequential_first, sequential_second)
sequential_first = sequential_second
net = initialize_network(
net, system_dict["custom_model"]["network_initializer"])
if (max_width == 1):
G.add_node("monk", pos=(3, position))
else:
G.add_node("monk", pos=(max_width + 3, position))
pos = nx.get_node_attributes(G, 'pos')
plt.figure(3, figsize=(8, 12 + position // 6))
nx.draw_networkx(G,
pos,
with_label=True,
font_size=16,
node_color="yellow",
node_size=100)
plt.savefig("graph.png")
system_dict["local"]["model"] = net
return system_dict
def create_block(network_stack, count, G, sequential_first, position,
current_width):
'''
Recursively create sub-blocks when designing custom networks
Args:
network_stack (list): List of lists containing information on layers for the sub-branch in the network
count (dict): A dictionary mapping to a count of every type of layer in the network
G (directed graph): NetworkX object
sequential_first (str): NAme of the current input layer
position (int): Vertical position on the directed graph
current_width (int): Horizontal position on the directed graph
Returns:
neural network: The required sub-branch
directed graph: Updated directed graph
str: Name of the outermost layer in the sub-network
int: Vertical position of the outer most layer in the sub-network
int: Horizontal position of the outer most layer in the sub-network
'''
position += 1
max_width = current_width
net = nn.HybridSequential()
for i in range(len(network_stack)):
if (type(network_stack[i]) == list):
is_block = True
if (type(network_stack[i][-1]) != list):
if (network_stack[i][-1]["name"] in ["add", "concatenate"]):
is_block = False
if (is_block):
block, G, count, sequential_second, position, _ = create_block(
network_stack[i], count, G, sequential_first, position,
current_width)
sequential_first = sequential_second
net.add(block)
else:
branch_end_points = []
branch_max_length = 0
branches = []
branch_net = []
#if(max_width < len(network_stack[i])-2):
# max_width = len(network_stack[i])-2;
max_width = current_width
width = current_width
for j in range(len(network_stack[i]) - 1):
small_net = []
branch_net.append(nn.HybridSequential())
branch_first = sequential_first
branch_position = position
column = max((j + 1) * 2 + current_width, width)
max_width = column
for k in range(len(network_stack[i][j])):
if type(network_stack[i][j][k]) == list:
is_block2 = True
if (type(network_stack[i][j][k][-1]) != list):
if (network_stack[i][j][k][-1]["name"]
in ["add", "concatenate"]):
is_block2 = False
if (is_block2):
block, G, count, branch_second, branch_position, width = create_block(
network_stack[i][j][k], count, G,
branch_first, branch_position,
column - 2) #j+k+width
else:
block, G, count, branch_second, branch_position, width = create_block(
[network_stack[i][j][k]], count, G,
branch_first, branch_position,
column - 2) #j+k+width
branch_first = branch_second
small_net.append(block)
branch_net[j].add(block)
else:
branch_second, count = get_layer_uid(
network_stack[i][j][k], count)
small_net.append(
custom_model_get_layer(network_stack[i][j][k]))
branch_net[j].add(
custom_model_get_layer(network_stack[i][j][k]))
G.add_node(branch_second,
pos=(column, branch_position))
branch_position += 1
G.add_edge(branch_first, branch_second)
branch_first = branch_second
branch_max_length = max(branch_position,
branch_max_length)
if (k == len(network_stack[i][j]) - 1):
branch_end_points.append(branch_second)
branches.append(small_net)
position = branch_max_length
position += 1
max_width += 2
sequential_second, count = get_layer_uid(
network_stack[i][-1], count)
if (network_stack[i][-1]["name"] == "concatenate"):
subnetwork = contrib_nn.HybridConcurrent(axis=1)
for j in range(len(network_stack[i]) - 1):
subnetwork.add(branch_net[j])
else:
subnetwork = addBlock(branches)
G.add_node(sequential_second,
pos=(2 + current_width, position))
position += 1
for i in range(len(branch_end_points)):
G.add_edge(branch_end_points[i], sequential_second)
sequential_first = sequential_second
net.add(subnetwork)
else:
sequential_second, count = get_layer_uid(network_stack[i], count)
net.add(custom_model_get_layer(network_stack[i]))
G.add_node(sequential_second, pos=(2 + current_width, position))
position += 1
G.add_edge(sequential_first, sequential_second)
sequential_first = sequential_second
return net, G, count, sequential_second, position, max_width
def create_network(network_stack, debug=True):
'''
Main function to create network when designing custom networks
Args:
network_stack (list): List of lists containing information on layers in the network
Returns:
neural network: The required complete network
'''
count = []
for i in range(len(names)):
count.append(1)
G = nx.DiGraph()
G.add_node("Net", pos=(1, 1))
sequential_first = "data"
#sequential_second, count = get_layer_uid(network_stack[0], count);
count = []
for i in range(len(names)):
count.append(1)
position = 1
G.add_node(sequential_first, pos=(2, 1))
position += 1
net = nn.HybridSequential()
max_width = 1
width = 0
for i in range(len(network_stack)):
if (type(network_stack[i]) == list):
is_block = True
if (type(network_stack[i][-1]) != list):
if (network_stack[i][-1]["name"] in ["add", "concatenate"]):
is_block = False
if (is_block):
block, G, count, sequential_second, position, _ = create_block(
network_stack[i], count, G, sequential_first, position, 0)
sequential_first = sequential_second
net.add(block)
else:
branch_end_points = []
branch_max_length = 0
branches = []
branch_net = []
if (max_width < len(network_stack[i]) - 2):
max_width = len(network_stack[i]) - 2
width = 0
for j in range(len(network_stack[i]) - 1):
small_net = []
branch_first = sequential_first
branch_net.append(nn.HybridSequential())
branch_position = position
if (width > 0):
if (column == width):
column += 2
else:
column = width
else:
column = (j + 1) * 2
for k in range(len(network_stack[i][j])):
if type(network_stack[i][j][k]) == list:
is_block2 = True
if (type(network_stack[i][j][k][-1]) != list):
if (network_stack[i][j][k][-1]["name"]
in ["add", "concatenate"]):
is_block2 = False
if (is_block2):
block, G, count, branch_second, branch_position, width = create_block(
network_stack[i][j][k], count, G,
branch_first, branch_position,
column - 2) #j+k+width
else:
block, G, count, branch_second, branch_position, width = create_block(
[network_stack[i][j][k]], count, G,
branch_first, branch_position, column - 2)
branch_first = branch_second
small_net.append(block)
branch_net[j].add(block)
else:
branch_second, count = get_layer_uid(
network_stack[i][j][k], count)
small_net.append(
custom_model_get_layer(network_stack[i][j][k]))
branch_net[j].add(
custom_model_get_layer(network_stack[i][j][k]))
G.add_node(branch_second,
pos=(column, branch_position))
branch_position += 1
G.add_edge(branch_first, branch_second)
branch_first = branch_second
branch_max_length = max(branch_position,
branch_max_length)
if (k == len(network_stack[i][j]) - 1):
branch_end_points.append(branch_second)
branches.append(small_net)
position = branch_max_length
position += 1
max_width += width
sequential_second, count = get_layer_uid(
network_stack[i][-1], count)
if (network_stack[i][-1]["name"] == "concatenate"):
subnetwork = contrib_nn.HybridConcurrent(axis=1)
for j in range(len(network_stack[i]) - 1):
subnetwork.add(branch_net[j])
else:
subnetwork = addBlock(branches)
sequential_second, count = get_layer_uid(
network_stack[i][-1], count)
G.add_node(sequential_second, pos=(2, position))
position += 1
for i in range(len(branch_end_points)):
G.add_edge(branch_end_points[i], sequential_second)
sequential_first = sequential_second
net.add(subnetwork)
else:
sequential_second, count = get_layer_uid(network_stack[i], count)
G.add_node(sequential_second, pos=(2, position))
net.add(custom_model_get_layer(network_stack[i]))
position += 1
G.add_edge(sequential_first, sequential_second)
sequential_first = sequential_second
max_width = max(max_width, width)
if (max_width == 1):
G.add_node("monk", pos=(3, position))
else:
G.add_node("monk", pos=(max_width + 3, position))
pos = nx.get_node_attributes(G, 'pos')
if (not debug):
plt.ioff()
plt.figure(3, figsize=(12, 12 + position // 6))
nx.draw_networkx(G,
pos,
with_label=True,
font_size=16,
node_color="yellow",
node_size=100)
plt.savefig("graph.png")
plt.clf()
else:
plt.figure(3, figsize=(12, 12 + position // 6))
nx.draw_networkx(G,
pos,
with_label=True,
font_size=16,
node_color="yellow",
node_size=100)
plt.savefig("graph.png")
return net
def create_block(network_stack, count, G, sequential_first, position,
current_width):
position += 1
max_width = current_width
net = nn.HybridSequential()
for i in range(len(network_stack)):
if (type(network_stack[i]) == list):
is_block = True
if (type(network_stack[i][-1]) != list):
if (network_stack[i][-1]["name"] in ["add", "concatenate"]):
is_block = False
if (is_block):
block, G, count, sequential_second, position, _ = create_block(
network_stack[i], count, G, sequential_first, position,
current_width)
sequential_first = sequential_second
net.add(block)
else:
branch_end_points = []
branch_max_length = 0
branches = []
branch_net = []
#if(max_width < len(network_stack[i])-2):
# max_width = len(network_stack[i])-2;
max_width = current_width
width = current_width
for j in range(len(network_stack[i]) - 1):
small_net = []
branch_net.append(nn.HybridSequential())
branch_first = sequential_first
branch_position = position
column = max((j + 1) * 2 + current_width, width)
max_width = column
for k in range(len(network_stack[i][j])):
if type(network_stack[i][j][k]) == list:
is_block2 = True
if (type(network_stack[i][j][k][-1]) != list):
if (network_stack[i][j][k][-1]["name"]
in ["add", "concatenate"]):
is_block2 = False
if (is_block2):
block, G, count, branch_second, branch_position, width = create_block(
network_stack[i][j][k], count, G,
branch_first, branch_position,
column - 2) #j+k+width
else:
block, G, count, branch_second, branch_position, width = create_block(
[network_stack[i][j][k]], count, G,
branch_first, branch_position,
column - 2) #j+k+width
branch_first = branch_second
small_net.append(block)
branch_net[j].add(block)
else:
branch_second, count = get_layer_uid(
network_stack[i][j][k], count)
small_net.append(
custom_model_get_layer(network_stack[i][j][k]))
branch_net[j].add(
custom_model_get_layer(network_stack[i][j][k]))
G.add_node(branch_second,
pos=(column, branch_position))
branch_position += 1
G.add_edge(branch_first, branch_second)
branch_first = branch_second
branch_max_length = max(branch_position,
branch_max_length)
if (k == len(network_stack[i][j]) - 1):
branch_end_points.append(branch_second)
branches.append(small_net)
position = branch_max_length
position += 1
max_width += 2
sequential_second, count = get_layer_uid(
network_stack[i][-1], count)
if (network_stack[i][-1]["name"] == "concatenate"):
subnetwork = contrib_nn.HybridConcurrent(axis=1)
for j in range(len(network_stack[i]) - 1):
subnetwork.add(branch_net[j])
else:
subnetwork = addBlock(branches)
G.add_node(sequential_second,
pos=(2 + current_width, position))
position += 1
for i in range(len(branch_end_points)):
G.add_edge(branch_end_points[i], sequential_second)
sequential_first = sequential_second
net.add(subnetwork)
else:
sequential_second, count = get_layer_uid(network_stack[i], count)
net.add(custom_model_get_layer(network_stack[i]))
G.add_node(sequential_second, pos=(2 + current_width, position))
position += 1
G.add_edge(sequential_first, sequential_second)
sequential_first = sequential_second
return net, G, count, sequential_second, position, max_width