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 main():
parser = argparse.ArgumentParser(
description="load a network and run inference on random data",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument("net")
parser.add_argument(
"--device", "-d", help="set defult device, like 'gpux' or 'cpux'"
)
parser.add_argument(
"--calc-output-rms",
action="store_true",
help="compute RMS of outputs; useful for comparing computing results",
)
parser.add_argument(
"--output-name",
nargs="*",
help="Specify output name. This option can be"
" specified multiple times. We will look for opr/var"
" in the graph",
)
parser.add_argument(
"--load-input-data",
help="load input data from pickle file; it should be"
" a numpy array or a dict of numpy array",
)
parser.add_argument("--profile", help="profiler output file")
parser.add_argument(
"--fast-run",
action="store_true",
help="enable fast running by profiling conv algorithms during compiling.",
)
parser.add_argument(
"--reproducible", action="store_true", help="use reproducible kernels"
)
parser.add_argument(
"--input-desc",
help="specifiy input names and shapes manually in"
" format: <name>:<shape>[;<name>:<shape>, ...], where"
" name is a string and shape is a comma separated"
' string. e.g., "data:128,1,28,28,label:128".'
" different input tensor are separated by semicolon.",
)
parser.add_argument(
"--batchsize",
type=int,
help="change batchsize; the first dimension of each"
" input is assumed to be batch size",
)
parser.add_argument(
"--warm-up",
action="store_true",
help="warm up model before do timing " " for better estimation",
)
parser.add_argument(
"--verbose",
"-v",
action="store_true",
help="verbose output, logging in debug mode",
)
parser.add_argument(
"--iter", type=int, default=1, help="number of iters to run the model"
)
parser.add_argument("--log", help="give a file path to duplicate log to")
parser.add_argument(
"--seed",
type=int,
default=0,
help="seed for random number generator for input data",
)
parser.add_argument(
"--rng",
help="special RNG options to generate input data in"
" format: <name>:func[;<name>:func, ...] where name is"
" a string and func is a python expression containing"
' "{}" for the size param, e.g. '
' "label:randint(low=0,high=1000,size={})"',
)
parser.add_argument(
"--focused-nvprof",
action="store_true",
help="only profile last iter for `nvprof --profile-from-start off`",
)
parser.add_argument(
"--optimize-for-inference",
action="store_true",
help="optimize model for inference",
)
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-hwcd4",
action="store_true",
help="transform the model format from NCHW to NHWCD4 for inference",
)
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-chwn4",
action="store_true",
help="transform the model format to CHWN4 "
"for inference, mainly used for nvidia tensorcore",
)
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-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-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)",
)
parser.add_argument(
"--dump-cpp-model",
help="write a C++ model that can be loaded by "
"megbrain/sdk/load-and-run; "
"this implies --embed-input",
)
parser.add_argument(
"--embed-input",
action="store_true",
help="embed input data as SharedDeviceTensor in model, "
"to remove memory copy for inputs",
)
args = parser.parse_args()
if args.verbose:
enable_debug_log()
if args.log:
set_log_file(args.log)
if args.device:
set_default_device(args.device)
if args.dump_cpp_model:
args.embed_input = True
logger.info("loading model ...")
graph, _, output_vars = G.load_graph(args.net)
input_vars = tools.get_dep_vars(output_vars, "Host2DeviceCopy")
if args.output_name is not None:
output_vars = tools.find_vars_by_name(output_vars, args.output_name)
data = make_data_given_desc(args, input_vars)
run_model(args, graph, input_vars, output_vars, data)
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_false",
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 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)
def make_feeds(args):
cg_rt, _, outputs = G.load_graph(args.input)
inputs = cgtools.get_dep_vars(outputs, "Host2DeviceCopy")
inputs = {i.name: i for i in inputs}
if not args.no_assert:
replace_varmap = {}
inp_map = {}
# replace var use InputNode
for name, var in inputs.items():
inp = G.InputNode(device="xpux",
dtype=var.dtype,
shape=var.shape,
graph=cg_rt)
replace_varmap[var] = inp.outputs[0]
inp_map[name] = inp
new = cgtools.replace_vars(outputs, replace_varmap)
if isinstance(new, rt.VarNode):
new = list(new)
output_nodes = [G.OutputNode(var) for var in new]
func = cg_rt.compile([node.outputs[0] for node in output_nodes])
def make_dev_tensor(value, dtype=None, device=None):
return tensor(value, dtype=dtype, device=device)._dev_tensor()
def calculate(*args, **kwargs):
output_val = []
# set inputs value
for name, var in inputs.items():
val = kwargs.pop(name, None)
assert val is not None, "miss input name{}".format(name)
dev_tensor = make_dev_tensor(val,
dtype=var.dtype,
device="xpux")
inp_map[name].set_value(dev_tensor)
func.execute()
for res in output_nodes:
output_val.append(res.get_value().numpy())
return output_val
def expect_name(var):
return "{}:expect".format(var.name)
testcases = []
np.set_printoptions(precision=2, threshold=4, suppress=True)
data_list = []
for item in args.data:
if item.startswith("@"):
with open(item[1:], "r") as f:
data_list.extend(
[line.rstrip() for line in f if line.rstrip() != ""])
else:
data_list.append(item)
for inp_spec in data_list:
cur_testcase = gen_one_testcase(args, inputs, inp_spec)
assert len(cur_testcase) == len(
inputs), "required inputs: {}; given data: {}".format(
inputs.keys(), cur_testcase.keys())
if not args.no_assert:
outputs_get = calculate(**cur_testcase)
for var, val in zip(outputs, outputs_get):
cur_testcase[expect_name(var)] = val
logger.info(
"generate test groundtruth: var={} shape={} range=({}, {})"
" mean={} var={}".format(var, val.shape, val.min(),
val.max(), np.mean(val),
np.var(val)))
testcases.append(cur_testcase)
logger.info("add testcase: \n {}".format("\n ".join(
"{}: shape={} dtype={} range=({:.2f},{:.2f}) "
"mean={:.2f} sd={:.2f}".format(k, v.shape, v.dtype, v.min(),
v.max(), np.mean(v), np.std(v))
for k, v in sorted(cur_testcase.items()))))
if not args.no_assert:
def expect_shp(var):
ret = var.shape
if ret:
return ret
return testcases[0][expect_name(var)].shape
def assert_equal(expect, real, **kwargs):
op = builtin.AssertEqual(**kwargs)
(res, ) = G.apply_normal_varnode(op, expect, real)
return G.VarNode(res)
verbose = not args.silent
outputs_new = []
for i in outputs:
device = rt.CompNode("xpux")
dtype = i.dtype
name = expect_name(i)
shape = expect_shp(i)
# make expect output as one input of model.
expect_get = rt.make_h2d(cg_rt, device, dtype, shape, name)
# insert assert opr to check expect and real.
outputs_new.append(
assert_equal(
G.VarNode(expect_get),
G.VarNode(i),
verbose=verbose,
maxerr=args.maxerr,
))
inputs[expect_name(i)] = expect_get
outputs = outputs_new
return {"outputs": outputs, "testcases": testcases}