def test_load_refcnt():
graph = mgb_graph.Graph()
varnode = graph.make_const(0)
buf, _ = mgb_graph.dump_graph([varnode])
graph, _, (varnode, ) = mgb_graph.load_graph(io.BytesIO(buf))
del graph
varnode.owner
def test_load_refcnt():
graph = mgb_graph.Graph()
varnode = graph.make_const(0)
buf, _ = mgb_graph.dump_graph([varnode])
ret = mgb_graph.load_graph(io.BytesIO(buf))
graph, (varnode, ) = ret.graph, ret.output_vars_list
del ret
del graph
varnode.owner
def change_batch_and_dump(inp_file, oup_file):
cg, _, outputs = G.load_graph(inp_file)
inputs = cgtools.get_dep_vars(outputs[0], "Host2DeviceCopy")
replace_dict = {}
for var in inputs:
n_shape = list(var.shape)
n_shape[0] = 1
new_input = make_h2d(cg, "xpux", var.dtype, n_shape, var.name)
replace_dict[var] = new_input
new_outputs = cgtools.replace_vars(outputs, replace_dict)
dump_content, _ = G.dump_graph(map(G.VarNode, new_outputs), keep_var_name=2)
with open(oup_file, "wb") as file:
file.write(dump_content)
def test_assert_equal():
g = G.Graph()
inp1 = g.make_h2d(dtype=np.float32, device="xpux")
inp2 = g.make_h2d(dtype=np.float32, device="xpux")
op = builtin.AssertEqual(maxerr=1e-5)
out = G.apply_normal_varnode(op, inp1._node, inp2._node)[0]
g.compile(out)
file = io.BytesIO()
out_model = G.dump_graph([out])
file.write(out_model[0])
file.seek(0)
net = Net.load(file)
dump_file = io.BytesIO()
net.dump(dump_file)
dump_file.seek(0)
g = GraphInference(dump_file)
g.run(np.array([1.0, 2.0]), np.array([1.0, 2.0]))
def run_model(args, graph, inputs, outputs, data):
# must use level0 to avoid unintended opr modification
graph.options.graph_opt_level = 0
logger.info("input tensors: ")
for k, v in data.items():
logger.info(" {}: {}".format(k, v.shape))
G.modify_opr_algo_strategy_inplace(outputs, get_execution_strategy(args))
if args.optimize_for_inference:
opt_kwargs = get_opt_kwargs(args)
outputs = G.optimize_for_inference(outputs, **opt_kwargs)
# embed inputs must be on the last, to avoid const fold
if args.embed_input:
outputs, inp_dict = tools.embed_inputs(outputs, data.values(), inputs=inputs)
else:
outputs, inp_dict = tools.convert_inputs(outputs, inputs=inputs)
if args.dump_cpp_model:
dump_content, _ = G.dump_graph(outputs, keep_var_name=2)
with open(args.dump_cpp_model, "wb") as file:
file.write(dump_content)
logger.info("C++ model written to {}".format(args.dump_cpp_model))
outputs, output_dict = tools.convert_outputs(outputs)
if args.profile:
profiler = tools.GraphProfiler(graph)
func = graph.compile(outputs)
def run():
if not args.embed_input:
for key in inp_dict:
inp_dict[key].set_value(mge.Tensor(data[key])._dev_tensor())
func.execute()
func.wait()
return [oup_node.get_value().numpy() for oup_node in output_dict.values()]
if args.warm_up:
logger.info("warming up")
run()
total_time = 0
for i in range(args.iter):
logger.info("iter {}".format(i))
start_time = time.time()
retval = run()
cur_time = time.time() - start_time
total_time += cur_time
avg_speed = (i + 1) / total_time
if "data" in data:
avg_speed *= data["data"].shape[0]
avg_speed_txt = "{:.3f}sample/s".format(avg_speed)
else:
avg_speed_txt = "{:.3f}batch/s".format(avg_speed)
msg = (
"iter {}: duration={:.4f}({:.4f})s average={:.4f}s "
"avg_speed={} time={:.4f}s"
).format(
i,
cur_time,
func.get_prev_exec_time(),
total_time / (i + 1),
avg_speed_txt,
total_time,
)
if args.calc_output_rms:
rms = []
for v in retval:
rms.append("{:.3g}".format(float(((v ** 2).mean()) ** 0.5)))
msg += " output_rms=[{}]".format(", ".join(rms))
if logger.level > logging.INFO:
print(msg)
else:
logger.info(msg)
if args.focused_nvprof:
if get_device_count("gpu") < 1:
logger.warning(
"No cuda device detected. ``focused_nvprof`` will be ignored."
)
else:
try:
import pycuda.driver as D
D.start_profiler()
func.execute()
func.wait()
D.stop_profiler()
except ImportError:
logger.error("`focused_nvprof need pycuda`", exc_info=True)
if args.profile:
with open(args.profile, "w") as fout:
fout.write(profiler.get())
return avg_speed
def main():
parser = argparse.ArgumentParser(
description="Pack computing graph, input values and expected output "
"values into one file for checking correctness. README.md gives more "
"details on the usage",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("input", help="MegEngine dumped model file")
parser.add_argument("-o", "--output", help="output file", required=True)
parser.add_argument(
"-d",
"--data",
default=[],
action="append",
required=True,
help="Given input test data when input file is a network, "
"and current network output would be used as groundtruth. "
"The format is var0:file0;var1:file1... to specify data files for "
"input vars. It can also be #rand(min,max,shape...) for generating "
"random input data, for example, #rand(0,255), "
"#rand(0,255,1,3,224,224) or #rand(0, 255, 1, ...) where `...` means "
"the remaining part of the original shape. "
"If the shape is not specified, the shape of "
"corresponding input tensors in the network will be used. "
"If there is only one input var, its name can be omitted. "
"Each data file can either be an image which can be loaded by opencv, "
"or a pickled numpy.ndarray. "
"This option can be given multiple times to add multiple testcases. "
" *NOTE* "
"If you start the data with the letter @, the rest should be a "
"filename, and each line in the file should be a single datum in "
"the format described above. ",
)
parser.add_argument(
"--repeat",
type=int,
default=1,
help="Specify how many times the input image is repeated. "
"Useful when running benchmark for batch size other than one. "
"Have no effect on randomly generated input data.",
)
parser.add_argument(
"--silent",
action="store_true",
help="set verbose to False in asserti_equal opr",
)
parser.add_argument(
"--optimize-for-inference",
action="store_true",
help="enbale optimization for inference",
)
parser.add_argument(
"--no-assert",
action="store_true",
help="do not insert assert_equal opr to check result; "
"this option is useful for benchmarking",
)
parser.add_argument(
"--maxerr",
type=float,
default=1e-4,
help="max error for assert_equal check during runtime",
)
parser.add_argument(
"--resize-input",
action="store_true",
help="resize input image to fit input var shape",
)
parser.add_argument(
"--input-transform",
help="a python expression to transform the input data. "
"Example: data / np.std(data)",
)
parser.add_argument(
"--discard-var-name",
action="store_true",
help="discard variable and param names in the "
"generated output",
)
parser.add_argument("--output-strip-info",
action="store_true",
help="output code strip information")
parser.add_argument(
"--enable-io16xc32",
action="store_true",
help="transform the mode to float16 io float32 compute",
)
parser.add_argument(
"--enable-ioc16",
action="store_true",
help="transform the dtype of the model to float16 io "
"and compute",
)
parser.add_argument(
"--enable-fuse-conv-bias-nonlinearity",
action="store_true",
help="fuse convolution bias and nonlinearity opr to a "
"conv_bias opr and compute",
)
parser.add_argument(
"--enable-hwcd4",
action="store_true",
help="transform the model format from NCHW to NHWCD4 "
"for inference; you may need to disable CUDA and set "
"MGB_USE_MEGDNN_DBG=2",
)
parser.add_argument(
"--enable-nchw4",
action="store_true",
help="transform the model format from NCHW to NCHW4 "
"for inference",
)
parser.add_argument(
"--enable-nchw88",
action="store_true",
help="transform the model format from NCHW to NCHW88 "
"for inference",
)
parser.add_argument(
"--enable-nchw44",
action="store_true",
help="transform the model format from NCHW to NCHW44 "
"for inference",
)
parser.add_argument(
"--enable-nchw44-dot",
action="store_true",
help="transform the model format from NCHW to NCHW44_DOT "
"for optimizing armv8.2 dot in inference",
)
parser.add_argument(
"--enable-nchw32",
action="store_true",
help="transform the model format from NCHW4 to NCHW32 "
"for inference on nvidia TensoCore",
)
parser.add_argument(
"--enable-chwn4",
action="store_true",
help="transform the model format to CHWN4 "
"for inference, mainly used for nvidia tensorcore",
)
parser.add_argument(
"--enable-fuse-conv-bias-with-z",
action="store_true",
help="fuse conv_bias with z input for inference on "
"nvidia GPU (this optimization pass will result in mismatch "
"of the precision of output of training and inference)",
)
args = parser.parse_args()
_, feeds = make_feeds(args)
assert isinstance(
feeds, dict) and feeds["testcases"], "testcases can not be empty"
output_mgbvars = feeds["outputs"]
output_mgbvars = optimize_for_inference(args, output_mgbvars)
inputs = cgtools.get_dep_vars(output_mgbvars, "Host2DeviceCopy")
inputs = sorted((i.name, i.dtype) for i in inputs)
if args.discard_var_name:
sereg_kwargs = dict(keep_var_name=0, keep_param_name=False)
else:
sereg_kwargs = dict(keep_var_name=2, keep_param_name=True)
strip_info_file = args.output + '.json' if args.output_strip_info else None
with open(args.output, "wb") as fout:
fout.write(b"mgbtest0")
fout.write(struct.pack("I", len(feeds["testcases"])))
if isinstance(output_mgbvars, dict):
wrap_output_vars = dict([(i, VarNode(j))
for i, j in output_mgbvars])
else:
wrap_output_vars = [VarNode(i) for i in output_mgbvars]
dump_content, stat = G.dump_graph(wrap_output_vars,
append_json=True,
strip_info_file=strip_info_file,
**sereg_kwargs)
fout.write(dump_content)
logger.info(
'graph dump sizes: tot_size={:.3f}KiB overhead={:.3f}KiB'.format(
stat.tot_bytes / 1024,
(stat.tot_bytes - stat.tensor_value_bytes) / 1024))
def make_dev_tensor(value, dtype=None, device=None):
return as_raw_tensor(value, dtype=dtype, device=device)._dev_tensor()
for testcase in feeds["testcases"]:
assert isinstance(testcase, dict)
cg = G.Graph()
output_mgbvars = []
for name, dtype in inputs:
output_mgbvars.append(
cg.make_const(
make_dev_tensor(testcase.pop(name),
dtype=dtype,
device="cpux")))
assert not testcase, "extra inputs provided in testcase: {}".format(
testcase.keys())
with open(args.output, "ab") as fout:
dump_content, _ = G.dump_graph(output_mgbvars,
strip_info_file=strip_info_file,
append_json=True)
fout.write(dump_content)
