def mmdnn_to_annette(args):
"""Convert MMDNN graph .pb-File to annette json format and stores to Annette graph std path.
Args:
network name (str): network name
Returns:
:obj:`annette.AnnetteGraph`: AnnetteGraph object
"""
graphfile = get_database( 'graphs', 'mmdnn', args.network+'.pb')
if(os.path.exists(graphfile)):
print("Graph found")
elif(os.path.exists(args.network)):
print("Graph-file detected")
# extract network name
args.network = os.path.split(args.network)[1].split('.pb')[0]
graphfile = get_database( 'graphs', 'mmdnn', args.network+'.pb')
else:
logging.error("File not found")
weightfile = None
mmdnn_graph = MMGraph(graphfile, weightfile)
annette_graph = mmdnn_graph.convert_to_annette(args.network)
json_file = get_database( 'graphs', 'annette',
annette_graph.model_spec["name"]+'.json')
annette_graph.to_json(json_file)
return annette_graph
def estimate(args):
"""Estimate example function
Args:
network name (str): network name
Returns:
float: estimated time in ms
"""
model = AnnetteGraph(
args.network, get_database('graphs', 'annette',
args.network + '.json'))
if args.mapping != "none":
opt = Mapping_model.from_json(
get_database('models', 'mapping', args.mapping + '.json'))
opt.run_optimization(model)
# LOAD MODELS
mod = Layer_model.from_json(
get_database('models', 'layer', args.layer + '.json'))
# APPLY ESTIMATION
res = mod.estimate_model(model)
write_result(args.network, res, args.mapping, args.layer,
get_database('results'))
return res[0], res[2]
def test_read_ncs2_report(network="benchmark_average_counters_report"):
report_file = get_database('benchmarks', 'tmp', network + '.csv')
ncs2.read_report(report_file)
annette_report = ncs2.r2a(report_file)
annette_report.to_pickle(
get_database('benchmarks', 'tmp', 'annette_bench1.pkl'))
assert True
def test_ONNXGraph_to_annette(network="cf_resnet50", inputs=None):
network_file = get_database('graphs', 'onnx', network + '.onnx')
onnx_network = graph.ONNXGraph(network_file)
annette_graph = onnx_network.onnx_to_annette(network, inputs)
json_file = get_database('graphs', 'annette',
annette_graph.model_spec["name"] + '.json')
annette_graph.to_json(json_file)
assert True
return 0
def test_matcher(network='annette_bench3', shape=None):
gen = generator.Graph_generator(network)
gen.add_configfile("config_v6.csv")
gen.generate_graph_from_config(401)
test_net = get_database('graphs', 'tf', network + '.pb')
#ncs2.optimize_network(test_net, source_fw = "tf", network = network, image = shape , input_node = "data", save_folder = get_database('benchmarks','tmp'))
#test_net = get_database('benchmarks','tmp',network+'.xml')
#not running network
#ncs2.run_network(test_net, report_dir = get_database('benchmarks','tmp'))
test_report = get_database('benchmarks', 'tmp',
'benchmark_average_counters_report.csv')
ncs2.read_report(test_report)
print(gen.graph.model_spec)
def tf_export_to_pb(self, output_node, save_path=None):
# Collect default graph information
g = tf.get_default_graph()
with tf.Session() as sess:
# Initialize the variables
sess.run(tf.global_variables_initializer())
g = g.as_graph_def(add_shapes=True)
# Convert variables to constants until the "fully_conn_1/Softmax" node
frozen_graph_def = tf.graph_util.convert_variables_to_constants(
sess, g, output_node)
print("load graph")
graph_nodes = [n for n in frozen_graph_def.node]
names = []
for t in graph_nodes:
if not ("Variable" in t.name or "BiasAdd" in t.name):
names.append(t.name.replace("/", "_").replace("-", "_"))
print(names)
# Write the intermediate representation of the graph to .pb file
if save_path:
net_file = save_path
else:
net_file = get_database('graphs', 'tf',
self.graph.model_spec['name'] + ".pb")
print(net_file)
with open(os.path.join(net_file), 'wb') as f:
graph_string = (frozen_graph_def.SerializeToString())
f.write(graph_string)
def load_estimator(self, est_model=None, est_dict=None):
if est_model != None:
self.est_model = pickle.load(open(get_database(est_model), 'rb'))
self.desc['est_model'] = est_model
else:
return False
self.est_dict = est_dict
self.desc['est_dict'] = est_dict
return True
def load_model(self, est_model=None, conv_dict=None):
if est_model != None:
self.est_model = pickle.load(open(get_database(est_model), 'rb'))
self.desc['est_model'] = est_model
else:
print("No Model to be loaded")
pass
self.conv_dict = conv_dict
self.desc['conv_dict'] = conv_dict
print("Conversion Dict: ")
print(self.conv_dict)
def generate_tf_model(graph):
"""generates Tensorflow 2 graph out of ANNETTE graph description
and stores to benchmark/graphs/tf2/ directory
Args:
graph :obj:`annette.graph.AnnetteGraph`: annette graph description to generate the tf2 graph from
"""
# generate tensorflow model and export to out_file
# with __dict__ we can see the content of the class
logging.debug(graph.__dict__)
# model_spec contains some info about the model
for key, value in graph.model_spec.items():
logging.debug(key)
logging.debug(value)
network_name = graph.model_spec['name']
filename = get_database('benchmark', 'graphs', 'tf2', network_name + '.pb')
logging.debug("Stored to: %s" % filename)
def add_configfile(self, configfile):
self.config = pd.read_csv(
get_database('benchmarks', 'config', configfile))
print(self.config)
def __init__(self, network):
#load graphstruct
json_file = get_database('graphs', 'annette', network + '.json')
self.graph = AnnetteGraph(network, json_file)
print(self.graph)
