def validate_onnx_model(model_file, input_file, mace_out_file, input_names,
input_shapes, input_data_formats, output_names,
output_shapes, output_data_formats,
validation_threshold, input_data_types, backend,
log_file):
import onnx
if backend == "tensorflow":
from onnx_tf.backend import prepare
print("valivate on onnx tensorflow backend.")
elif backend == "caffe2" or backend == "pytorch":
from caffe2.python.onnx.backend import prepare
print("valivate on onnx caffe2 backend.")
else:
util.MaceLogger.error(
VALIDATION_MODULE,
"onnx backend framwork '" + backend + "' is invalid.")
if not os.path.isfile(model_file):
util.MaceLogger.error(
VALIDATION_MODULE,
"Input graph file '" + model_file + "' does not exist!")
model = onnx.load(model_file)
input_dict = {}
for i in range(len(input_names)):
input_value = load_data(
util.formatted_file_name(input_file, input_names[i]),
input_data_types[i])
input_value = input_value.reshape(input_shapes[i])
if input_data_formats[i] == DataFormat.NHWC and \
len(input_shapes[i]) == 4:
input_value = input_value.transpose((0, 3, 1, 2))
input_dict[input_names[i]] = input_value
onnx_outputs = []
for i in range(len(output_names)):
out_shape = output_shapes[i][:]
if output_data_formats[i] == DataFormat.NHWC and \
len(out_shape) == 4:
out_shape[1], out_shape[2], out_shape[3] = \
out_shape[3], out_shape[1], out_shape[2]
onnx_outputs.append(
onnx.helper.make_tensor_value_info(output_names[i],
onnx.TensorProto.FLOAT,
out_shape))
model.graph.output.extend(onnx_outputs)
rep = prepare(model)
output_values = rep.run(input_dict)
for i in range(len(output_names)):
out_name = output_names[i]
value = output_values[out_name].flatten()
output_file_name = util.formatted_file_name(mace_out_file,
output_names[i])
mace_out_value = load_data(output_file_name)
if output_data_formats[i] == DataFormat.NHWC and \
len(output_shapes[i]) == 4:
mace_out_value = mace_out_value.reshape(output_shapes[i]) \
.transpose((0, 3, 1, 2))
compare_output(output_names[i], mace_out_value, value,
validation_threshold, log_file)
def validate_tf_model(model_file, input_file, mace_out_file, input_names,
input_shapes, input_data_formats, output_names,
output_shapes, output_data_formats, validation_threshold,
input_data_types, log_file):
import tensorflow as tf
if not os.path.isfile(model_file):
util.MaceLogger.error(
VALIDATION_MODULE,
"Input graph file '" + model_file + "' does not exist!")
tf.reset_default_graph()
input_graph_def = tf.GraphDef()
with open(model_file, "rb") as f:
data = f.read()
input_graph_def.ParseFromString(data)
tf.import_graph_def(input_graph_def, name="")
with tf.Session() as session:
with session.graph.as_default() as graph:
tf.import_graph_def(input_graph_def, name="")
input_dict = {}
for i in range(len(input_names)):
input_value = load_data(
util.formatted_file_name(input_file, input_names[i]),
input_data_types[i])
input_value = input_value.reshape(input_shapes[i])
if input_data_formats[i] == DataFormat.NCHW and \
len(input_shapes[i]) == 4:
input_value = input_value.transpose((0, 2, 3, 1))
elif input_data_formats[i] == DataFormat.OIHW and \
len(input_shapes[i]) == 4:
# OIHW -> HWIO
input_value = input_value.transpose((2, 3, 1, 0))
input_node = graph.get_tensor_by_name(
normalize_tf_tensor_name(input_names[i]))
input_dict[input_node] = input_value
output_nodes = []
for name in output_names:
output_nodes.extend([
graph.get_tensor_by_name(
normalize_tf_tensor_name(name))
])
output_values = session.run(output_nodes, feed_dict=input_dict)
for i in range(len(output_names)):
output_file_name = util.formatted_file_name(
mace_out_file, output_names[i])
mace_out_value = load_data(output_file_name)
if output_data_formats[i] == DataFormat.NCHW and \
len(output_shapes[i]) == 4:
mace_out_value = mace_out_value. \
reshape(output_shapes[i]).transpose((0, 2, 3, 1))
compare_output(output_names[i], mace_out_value,
output_values[i], validation_threshold,
log_file)
def validate_caffe_model(model_file, input_file, mace_out_file, weight_file,
input_names, input_shapes, input_data_formats,
output_names, output_shapes, output_data_formats,
validation_threshold, log_file):
os.environ['GLOG_minloglevel'] = '1' # suprress Caffe verbose prints
import caffe
if not os.path.isfile(model_file):
util.MaceLogger.error(
VALIDATION_MODULE,
"Input graph file '" + model_file + "' does not exist!")
if not os.path.isfile(weight_file):
util.MaceLogger.error(
VALIDATION_MODULE,
"Input weight file '" + weight_file + "' does not exist!")
caffe.set_mode_cpu()
net = caffe.Net(model_file, caffe.TEST, weights=weight_file)
for i in range(len(input_names)):
input_value = load_data(
util.formatted_file_name(input_file, input_names[i]))
input_value = input_value.reshape(input_shapes[i])
if input_data_formats[i] == DataFormat.NHWC and \
len(input_shapes[i]) == 4:
input_value = input_value.transpose((0, 3, 1, 2))
input_blob_name = input_names[i]
try:
if input_names[i] in net.top_names:
input_blob_name = net.top_names[input_names[i]][0]
except ValueError:
pass
new_shape = input_value.shape
net.blobs[input_blob_name].reshape(*new_shape)
for index in range(input_value.shape[0]):
net.blobs[input_blob_name].data[index] = input_value[index]
net.forward()
for i in range(len(output_names)):
value = net.blobs[output_names[i]].data
output_file_name = util.formatted_file_name(mace_out_file,
output_names[i])
mace_out_value = load_data(output_file_name)
if output_data_formats[i] == DataFormat.NHWC and \
len(output_shapes[i]) == 4:
mace_out_value = mace_out_value.reshape(output_shapes[i]) \
.transpose((0, 3, 1, 2))
compare_output(output_names[i], mace_out_value, value,
validation_threshold, log_file)
def validate_with_file(output_names, output_shapes, mace_out_file,
validation_outputs_data, validation_threshold,
log_file):
for i in range(len(output_names)):
if validation_outputs_data[i].startswith("http://") or \
validation_outputs_data[i].startswith("https://"):
validation_file_name = util.formatted_file_name(
mace_out_file, output_names[i] + '_validation')
six.moves.urllib.request.urlretrieve(validation_outputs_data[i],
validation_file_name)
else:
validation_file_name = validation_outputs_data[i]
value = load_data(validation_file_name)
out_shape = output_shapes[i]
if len(out_shape) == 4:
out_shape[1], out_shape[2], out_shape[3] = \
out_shape[3], out_shape[1], out_shape[2]
value = value.reshape(out_shape).transpose((0, 2, 3, 1))
output_file_name = util.formatted_file_name(mace_out_file,
output_names[i])
mace_out_value = load_data(output_file_name)
compare_output(output_names[i], mace_out_value, value,
validation_threshold, log_file)
def generate_input_data(input_file, input_node, input_shape, input_ranges,
input_data_type):
np.random.seed()
for i in range(len(input_node)):
data = np.random.random(input_shape[i]) * (
input_ranges[i][1] - input_ranges[i][0]) + input_ranges[i][0]
input_file_name = util.formatted_file_name(input_file, input_node[i])
MaceLogger.info('Generate input file: %s' % input_file_name)
if input_data_type[i] == mace_pb2.DT_FLOAT:
np_data_type = np.float32
elif input_data_type[i] == mace_pb2.DT_INT32:
np_data_type = np.int32
data.astype(np_data_type).tofile(input_file_name)
def run_model_for_device(flags, args, dev, model_name, model_conf):
runtime = flags.runtime
target_abi = flags.target_abi
install_dir = run_target.default_install_dir(target_abi) + "/" + model_name
sysdir = install_dir + "/interior"
dev.mkdir(sysdir)
if not runtime:
runtime = model_conf[ModelKeys.runtime]
if runtime == DeviceType.CPU_GPU:
runtime = DeviceType.GPU
else:
runtime = config_parser.parse_device_type(runtime)
# install models to devices
workdir = flags.output + "/" + model_name
model_file = model_name + ".pb"
model_data_file = model_name + ".data"
model_path = workdir + "/model/" + model_file
model_data_path = workdir + "/model/" + model_data_file
if os.path.exists(model_path) and os.path.exists(model_data_path):
dev.install(Target(model_path), install_dir)
dev.install(Target(model_data_path), install_dir)
else:
MaceLogger.warning("No models exist in %s, use --model_file and"
" --model_data_file specified in args" % model_path)
if ModelKeys.check_tensors in model_conf:
model_conf[ModelKeys.output_tensors] = model_conf[
ModelKeys.check_tensors]
model_conf[ModelKeys.output_shapes] = model_conf[
ModelKeys.check_shapes]
model_file_path = ""
if not flags.gencode_model:
model_file_path = install_dir + "/" + model_file
model_data_file_path = ""
if not flags.gencode_param:
model_data_file_path = install_dir + "/" + model_data_file
model_args = {
"model_name":
model_name,
"model_file":
model_file_path,
"model_data_file":
model_data_file_path,
"input_node":
",".join(model_conf[ModelKeys.input_tensors]),
"input_shape":
join_2d_array(model_conf[ModelKeys.input_shapes]),
"output_node":
",".join(model_conf[ModelKeys.output_tensors]),
"output_shape":
join_2d_array(model_conf[ModelKeys.output_shapes]),
"input_data_format":
",".join([df.name for df in model_conf[ModelKeys.input_data_formats]]),
"output_data_format":
",".join([df.name
for df in model_conf[ModelKeys.output_data_formats]]),
"device":
runtime.name
}
opts = [
"--%s=%s" % (arg_key, arg_val)
for arg_key, arg_val in model_args.items()
] + args
should_generate_data = (flags.validate or flags.tune
or "--benchmark" in opts)
if should_generate_data:
tmpdirname = tempfile.mkdtemp()
input_file_prefix = tmpdirname + "/" + model_name
if ModelKeys.validation_inputs_data in model_conf:
input_tensor = model_conf[ModelKeys.input_tensors]
input_data = model_conf[ModelKeys.validation_inputs_data]
mace_check(
len(input_tensor) == len(input_data),
"len(input_tensor) != len(validate_data")
for i in range(len(input_tensor)):
util.download_or_get_file(
model_conf[ModelKeys.validation_inputs_data][i], "",
util.formatted_file_name(input_file_prefix,
input_tensor[i]))
else:
generate_input_data(input_file_prefix,
model_conf[ModelKeys.input_tensors],
model_conf[ModelKeys.input_shapes],
model_conf[ModelKeys.input_ranges],
model_conf[ModelKeys.input_data_types])
dev.install(Target(tmpdirname), install_dir + "/validate_in")
target_input_file = "%s/validate_in/%s" % (install_dir, model_name)
target_output_dir = "%s/validate_out" % install_dir
dev.mkdir(target_output_dir)
target_output_file = target_output_dir + "/" + model_name
opts += [
"--input_file=%s" % target_input_file,
"--output_file=%s" % target_output_file
]
# run
envs = flags.envs.split(" ") + ["MACE_INTERNAL_STORAGE_PATH=%s" % sysdir]
if flags.tune:
envs += [
"MACE_TUNING=1",
"MACE_RUN_PARAMETER_PATH=%s/interior/tune_params" % install_dir
]
opts += ["--round=0"]
if flags.vlog_level > 0:
envs += ["MACE_CPP_MIN_VLOG_LEVEL=%s" % flags.vlog_level]
build_dir = flags.build_dir + "/" + target_abi
libs = []
if model_conf[ModelKeys.runtime] == DeviceType.HEXAGON:
libs += ["third_party/nnlib/%s/libhexagon_controller.so" % target_abi]
elif model_conf[ModelKeys.runtime] == DeviceType.APU:
libs += ["third_party/apu/libapu-frontend.so"]
target = Target(build_dir + "/install/bin/mace_run",
libs,
opts=opts,
envs=envs)
run_target.run_target(target_abi,
install_dir,
target,
device_ids=flags.target_socs)
if runtime == DeviceType.GPU:
opencl_dir = workdir + "/opencl"
util.mkdir_p(opencl_dir)
dev.pull(
Target(install_dir + "/interior/mace_cl_compiled_program.bin"),
"%s/%s_compiled_opencl_kernel.%s.%s.bin" %
(opencl_dir, model_name, dev.info()["ro.product.model"].replace(
' ', ''), dev.info()["ro.board.platform"]))
if flags.tune:
dev.pull(
Target(install_dir + "/interior/tune_params"),
"%s/%s_tuned_opencl_parameter.%s.%s.bin" %
(opencl_dir, model_name,
dev.info()["ro.product.model"].replace(
' ', ''), dev.info()["ro.board.platform"]))
if flags.validate:
validate_model_file = util.download_or_get_model(
model_conf[ModelKeys.model_file_path],
model_conf[ModelKeys.model_sha256_checksum], tmpdirname)
validate_weight_file = ""
if ModelKeys.weight_file_path in model_conf:
validate_weight_file = util.download_or_get_model(
model_conf[ModelKeys.weight_file_path],
model_conf[ModelKeys.weight_sha256_checksum], tmpdirname)
dev.pull(Target(target_output_dir), tmpdirname + "/validate_out")
output_file_prefix = tmpdirname + "/validate_out/" + model_name
validate.validate(
model_conf[ModelKeys.platform], validate_model_file,
validate_weight_file, input_file_prefix, output_file_prefix,
model_conf[ModelKeys.input_shapes],
model_conf[ModelKeys.output_shapes],
model_conf[ModelKeys.input_data_formats],
model_conf[ModelKeys.output_data_formats],
model_conf[ModelKeys.input_tensors],
model_conf[ModelKeys.output_tensors], flags.validate_threshold,
model_conf[ModelKeys.input_data_types], flags.backend, "", "")
if should_generate_data:
shutil.rmtree(tmpdirname)