def parse_caffe_net(caffe_net):
"""
Create and populate a Topology object, based on a given Caffe protobuf network object
Todo: fix Input assignment
"""
graph = Topology()
# Input BLOBs
for i in range(len(caffe_net.input)):
if len(caffe_net.input_shape) > 0:
graph.add_blob(caffe_net.input[i], caffe_net.input_shape[i].dim,
None)
elif len(caffe_net.input_dim) > 0:
# graph.add_blob(caffe_net.input[i], caffe_net.input_dim[i], None)
graph.add_blob(caffe_net.input[i], caffe_net.input_dim, None)
if len(caffe_net.layer) < 1:
exit(
"Something went wrong - the parser can't find any layers in the network"
)
for layer in caffe_net.layer:
debug('evaluating layer: ' + layer.name)
# filter away layers used only in training phase
phase = 1 # caffe_pb2.Phase.TEST
if phase is not None:
included = False
if len(layer.include) == 0:
included = True
if len(layer.include) > 0 and len(layer.exclude) > 0:
raise ValueError('layer ' + layer.name + ' has both include '
'and exclude specified.')
for layer_phase in layer.include:
included = included or layer_phase.phase == phase
for layer_phase in layer.exclude:
included = included and not layer_phase.phase == phase
if not included:
continue
node_role = 'Producer'
if (len(layer.bottom) == 1 and len(layer.top) == 1
and layer.bottom[0] == layer.top[0]):
# We have an in-place neuron layer.
node_role = 'Modifier'
new_node = graph.add_node(layer.name, layer.type, layer, node_role)
# Iterate over BLOBs consumed by this layer and create edges to them
for caffe_bottom_blob in layer.bottom:
blob = graph.find_blob_by_name(caffe_bottom_blob)
if blob == None:
raise ValueError(layer.name + ' - could not find BLOB:' +
caffe_bottom_blob)
edge = graph.add_edge(src=blob.producer, dst=new_node, blob=blob)
# Add the BLOBs produced by this layer to the topology
for caffe_top_blob in layer.top:
if new_node.type == "Input":
graph.add_blob(caffe_top_blob,
layer.input_param.shape[0].dim,
producer=new_node)
else:
graph.add_blob(caffe_top_blob, None, producer=new_node)
# Add fake output edges
output_blobs = graph.find_output_blobs()
for blob_name in output_blobs:
blob = graph.find_blob_by_name(blob_name)
graph.add_edge(src=blob.producer, dst=None, blob=blob)
return graph
def parse_caffe_net(caffe_net):
"""
Create and populate a Topology object, based on a given Caffe protobuf network object
Todo: fix Input assignment
"""
graph = Topology()
modifiers = {}
# Input BLOBs
for i in range(len(caffe_net.input)):
if len(caffe_net.input_shape) > 0:
graph.add_blob("b_"+caffe_net.input[i], caffe_net.input_shape[i].dim, None)
elif len(caffe_net.input_dim) > 0:
# graph.add_blob(caffe_net.input[i], caffe_net.input_dim[i], None)
graph.add_blob("b_"+caffe_net.input[i], caffe_net.input_dim, None)
if len(caffe_net.layer) < 1:
exit("Something went wrong - the parser can't find any layers in the network")
for layer in caffe_net.layer:
log().debug('evaluating layer: ' + layer.name)
#print('evaluating layer: ' + layer.name)
# filter away layers used only in training phase
phase = 1 # caffe_pb2.Phase.TEST
if phase is not None:
included = False
if len(layer.include) == 0:
included = True
if len(layer.include) > 0 and len(layer.exclude) > 0:
raise ValueError('layer ' + layer.name + ' has both include '
'and exclude specified.')
for layer_phase in layer.include:
included = included or layer_phase.phase == phase
for layer_phase in layer.exclude:
included = included and not layer_phase.phase == phase
if not included:
continue
if layer.type == "Input":
graph.add_blob("b_"+layer.name, layer.input_param.shape[0].dim, producer=None)
continue
else:
new_node = graph.add_op(layer.name, layer.type, layer)
if (len(layer.bottom) == 1 and
len(layer.top) == 1 and
layer.bottom[0] == layer.top[0]):
# We have an in-place neuron layer.
if layer.bottom[0] not in modifiers:
modifiers[layer.bottom[0]] = []
modifiers[layer.bottom[0]].append(new_node)
continue
# Iterate over BLOBs consumed by this layer and create edges to them
for caffe_bottom_blob in layer.bottom:
# Because the Caffe Tensors and Layers have the same name, I add a b_ prefix to
# BLOB names, to make them distinguishable
blob = graph.find_blob_by_name("b_"+caffe_bottom_blob)
if blob == None:
raise ValueError(layer.name + ' - could not find BLOB:' + caffe_bottom_blob)
edge = graph.add_edge(src=blob, dst=new_node)
# Add the BLOBs produced by this layer to the topology
for caffe_top_blob in layer.top:
if new_node==None: #.type == "Input":
new_blob = graph.add_blob("b_"+caffe_top_blob, layer.input_param.shape[0].dim, producer=None)#producer=new_node)
else:
new_blob = graph.add_blob("b_"+caffe_top_blob, None, producer=new_node)
# Add producer edges
edge = graph.add_edge(src=new_node, dst=new_blob)
''' Below is some ugly code to handle Caffe's modifier layers (e.g. ReLU, Dropout).
This code "normailizes" these nodes by "flattening" them.
'''
for modified_node in modifiers:
src = graph.find_op_by_name(modified_node)
# now inject the modifier
for modifier in modifiers[modified_node]:
outgoing_edges = graph.find_outgoing_edges(src)
assert len(outgoing_edges) == 1
assert type(outgoing_edges[0].dst)==BLOB
replicated_blob = copy.deepcopy(outgoing_edges[0].dst)
replicated_blob.name += '_copy'
graph.add_blob2(replicated_blob)
graph.add_edge(src=src, dst=replicated_blob)
graph.add_edge(src=replicated_blob, dst=modifier)
graph.add_edge(src=modifier, dst=outgoing_edges[0].dst)
graph.del_edge(outgoing_edges[0])
return graph
dt = sim_env.dt
### Network Topology
topo_file = open('topology/default.txt', 'r')
observ_topology = Topology(sim_env.N)
comm_topology = Topology(sim_env.N)
edge_num_str = topo_file.readline()
for i in range(int(edge_num_str)):
line = topo_file.readline()
edge = line.split(", ")
observ_topology.add_edge(int(edge[0]), int(edge[1]))
edge_num_str = topo_file.readline()
for i in range(int(edge_num_str)):
line = topo_file.readline()
edge = line.split(", ")
comm_topology.add_edge(int(edge[0]), int(edge[1]))
topo_file.close()
gs_ci_rmse_arr = [0] * total_T
gs_ci_rmte_arr = [0] * total_T
gs_ci_1_rmse_arr = [0] * total_T
gs_ci_1_rmte_arr = [0] * total_T
