def get_net_runtime(
self,
model,
blobs,
input_dims,
input_dtype=core.DataType.FLOAT,
device="SM-G950U-7.0-24",
mode="ave",
verbose=True,
):
"""
Calculated the run-time latency for a given model from the loop-up table
Args:
model: caffe2 pb
input_dims: input dimensions
device: device used for benchmarking
wait: if not found, whether to wait for the result
"""
db_file = os.path.join(
self.db_dir, device + "-" + "op_lut_database.json"
)
if self.current_db != db_file:
self.current_db = db_file
self.op_lut.load(db_file)
ops, input_shapes, input_dtypes = get_ops_from_net(
model, blobs, input_dims
)
for i in range(len(input_shapes)):
input_shapes[i] = map(lambda x: list(x), input_shapes[i])
devices = [device for _ in range(len(input_shapes))]
while True:
op_time, net_time, all_found = [], 0, True
for i, op in enumerate(ops):
op_query = LUTSchema()
load_caffe2_op(
op_query,
op,
input_shapes=input_shapes[i],
input_dtypes=input_dtypes[i],
device=devices[i],
model_name=str(model),
)
latency = self.get_op_runtime(
op_query, mode=mode, verbose=verbose
)
op_time.append({"op": op_query, "time": latency})
if latency is not None:
net_time += latency
else:
all_found = False
return op_time, net_time, all_found
def load(self, dbfile):
"""
Load a database file and convert all records to LUTSchema
Args:
dbfile: path to the database file
"""
self.ops = []
assert os.path.exists(dbfile)
db_file = open(dbfile, "r")
db = db_file.readlines()
for record in db:
op = LUTSchema()
op.load_from_json(json.loads(record))
self.ops.append(op)
db_file.close()