def perf_tuning(self, temp_json, input_params):
"""
Define a celery job which runs OLive pipeline perf-tuning.
Args:
temp_json: Input specs for perf-tuning job.
input_params: JSON object storing the input arguments from frontend user.
Return:
JSON response with run status and results.
"""
response_object = {'status': 'success'}
pipeline = onnxpipeline.Pipeline()
# create result dir
result_dir = os.path.join(app_config.PERF_RES_DIR, get_timestamp())
if not os.path.exists(result_dir):
os.makedirs(result_dir)
result = pipeline.perf_tuning(input_json=temp_json, result=result_dir)
try:
r = pipeline.get_result(result)
response_object['result'] = json.dumps(r.latency)
response_object['profiling'] = r.profiling_ops
except RuntimeError:
pass
response_object['logs'] = pipeline.output
# clean up result dir
# if os.path.exists(result_dir):
# rmtree(result_dir)
clean(app.root_path)
return response_object
def main():
args = get_args()
pipeline = onnxpipeline.Pipeline()
model = pipeline.convert_model(
model_type=args.model_type,
model=pipeline.win_path_to_linux_relative(args.model),
model_input_shapes=args.model_input_shapes,
model_inputs_names=args.model_inputs_names,
model_outputs_names=args.model_outputs_names,
target_opset=args.target_opset,
input_json=args.input_json,
initial_types=args.initial_types,
caffe_model_prototxt=args.caffe_model_prototxt,
windows=not args.linux)
if 'Runtimes' in pipeline.client.info(
) and 'nvidia' in pipeline.client.info()['Runtimes']:
pipeline.perf_tuning(model=model,
result=pipeline.win_path_to_linux_relative(
args.result),
runtime=args.gpu,
windows=not args.linux)
else:
if args.gpu:
print('Not support Nvidia in local machine. Need to be installed.')
args.gpu = False
pipeline.perf_tuning(model=model,
result=args.result,
runtime=args.gpu,
windows=not args.linux)
def convert(self, model_name, temp_json, cur_ts, root_path, input_params):
"""
Define a celery job which runs OLive pipeline onnx-converter.
Args:
model_name: Model path to run onnx-convert.
temp_json: Input specs for onnx-convert job.
cur_ts: Unique timestamp for the job.
root_path: Project app absolute root path.
input_params: JSON object storing the input arguments from frontend user
Return:
JSON response with run status and results.
"""
response_object = {'status': 'success'}
# Initiate pipeline object with targeted directory
pipeline = onnxpipeline.Pipeline(
convert_directory=os.path.join(app_config.CONVERT_RES_DIR, cur_ts))
model = pipeline.convert_model(model=model_name, input_json=temp_json)
try:
with open(posixpath.join(pipeline.convert_directory,
'output.json')) as f:
json_data = json.load(f)
response_object['output_json'] = json_data
response_object['logs'] = pipeline.output
response_object['converted_model'] = model
except:
#fail to convert
pass
target_dir = app_config.DOWNLOAD_DIR
input_root = os.path.join(root_path, target_dir)
# compress input directory
compress_path = os.path.join(pipeline.convert_directory,
app_config.TEST_DATA_DIR)
input_path = os.path.join(input_root, cur_ts)
if not os.path.exists(input_path):
os.makedirs(input_path)
tar = tarfile.open(os.path.join(input_path, app_config.COMPRESS_NAME),
"w:gz")
try:
tar.add(compress_path, arcname=app_config.TEST_DATA_DIR)
except:
# fail to generate input
pass
tar.close()
# copy converted onnx model
if os.path.exists(pipeline.convert_path):
copyfile(pipeline.convert_path,
os.path.join(input_root, cur_ts, pipeline.convert_name))
response_object['input_path'] = posixpath.join(target_dir, cur_ts,
app_config.COMPRESS_NAME)
response_object['model_path'] = posixpath.join(target_dir, cur_ts,
pipeline.convert_name)
clean(root_path)
return response_object
def test_constructor_fail(self):
directory_name = 'not_exist_directory'
try:
pipeline = onnxpipeline.Pipeline(directory_name)
self.fail("Pipeline should raised RuntimeError")
except:
self.assertRaises(RuntimeError)
def test_cntk_pass(self):
directory_name = self.deep_dir['cntk']
pipeline = onnxpipeline.Pipeline(directory_name, convert_directory=self.convert_dir_pass, print_logs=self.print_logs)
def test_convert_pass():
model = pipeline.convert_model(model_type='cntk', model='ResNet50_ImageNet_Caffe.model')
return model
model = test_convert_pass()
output_json = osp.join(pipeline.path, pipeline.convert_directory, config.OUTPUT_JSON)
self.check_json_staus(['SUCCESS', 'SUCCESS'], self.check_converted_json(output_json))
def test_tensorflow_pass(self):
directory_name = self.deep_dir['tensorflow']
pipeline = onnxpipeline.Pipeline(directory_name, convert_directory=self.convert_dir_pass, print_logs=self.print_logs)
def test_convert_pass():
model = pipeline.convert_model(model_type='tensorflow')
return model
model = test_convert_pass()
output_json = osp.join(pipeline.path, pipeline.convert_directory, config.OUTPUT_JSON)
self.check_json_staus(['SUCCESS', 'SUCCESS'], self.check_converted_json(output_json))
def test_pytorch_pass(self):
directory_name = self.deep_dir['pytorch']
pipeline = onnxpipeline.Pipeline(directory_name, convert_directory=self.convert_dir_pass, print_logs=self.print_logs)
def test_convert_pass():
model = pipeline.convert_model(model_type='pytorch', model='saved_model.pb', model_input_shapes='(1,3,224,224)')
return model
model = test_convert_pass()
output_json = osp.join(pipeline.path, pipeline.convert_directory, config.OUTPUT_JSON)
self.check_json_staus(['SUCCESS', 'SUCCESS'], self.check_converted_json(output_json))
def test_convert_from_onnx(self):
directory_name = self.deep_dir['onnx']
pipeline = onnxpipeline.Pipeline(directory_name, convert_directory=self.convert_dir_pass)
def test_convert_pass():
model = pipeline.convert_model(model_type='onnx', model='model.onnx', model_input_shapes='(1,3,224,224)')
return model
model = test_convert_pass()
output_json = osp.join(pipeline.path, pipeline.convert_directory, config.OUTPUT_JSON)
self.check_json_staus(['SUCCESS', 'SKIPPED'], self.check_converted_json(output_json))
def test_cntk_fail(self):
directory_name = self.deep_dir['cntk']
pipeline = onnxpipeline.Pipeline(directory_name, convert_directory=self.convert_dir_fail, print_logs=self.print_logs)
def test_convert_fail_no_model():
model = pipeline.convert_model(model_type='cntk')
return model
model = test_convert_fail_no_model()
output_json = osp.join(pipeline.path, pipeline.convert_directory, config.OUTPUT_JSON)
self.check_json_staus(['FAILED', 'FAILED'], self.check_converted_json(output_json))
def test_tensorflow_fail(self):
directory_name = self.deep_dir['tensorflow']
pipeline = onnxpipeline.Pipeline(directory_name, convert_directory=self.convert_dir_fail)
def test_convert_fail_no_model():
model = pipeline.convert_model(model_type='tensorflow', model='not_exist_path')
return model
model = test_convert_fail_no_model()
output_json = osp.join(pipeline.path, pipeline.convert_directory, config.OUTPUT_JSON)
self.check_json_staus(['FAILED', 'FAILED'], self.check_converted_json(output_json))
def convert():
response_object = {'status': 'success'}
pipeline = onnxpipeline.Pipeline()
# may not work in Unix OS cuz the permission
if os.path.exists(pipeline.convert_directory):
rmtree(pipeline.convert_directory)
os.mkdir(pipeline.convert_directory)
model_name, temp_json = get_params(request, pipeline.convert_directory)
model = pipeline.convert_model(model=model_name, input_json=temp_json)
try:
with open(posixpath.join(pipeline.convert_directory,
'output.json')) as f:
json_data = json.load(f)
response_object['output_json'] = json_data
response_object['logs'] = pipeline.output
response_object['converted_model'] = model
except:
#fail to convert
pass
target_dir = app_config.DOWNLOAD_DIR
input_root = os.path.join(app.root_path, app_config.STATIC_DIR, target_dir)
# compress input directory
compress_path = os.path.join(pipeline.convert_directory,
app_config.INPUT_DIR)
input_path = os.path.join(input_root)
if not os.path.exists(input_path):
os.makedirs(input_path)
tar = tarfile.open(os.path.join(input_path, app_config.COMPRESS_NAME),
"w:gz")
try:
tar.add(compress_path, arcname=app_config.INPUT_DIR)
except:
# fail to generate input
pass
tar.close()
# copy converted onnx model
if os.path.exists(pipeline.convert_path):
copyfile(pipeline.convert_path,
os.path.join(input_root, pipeline.convert_name))
response_object['input_path'] = posixpath.join(target_dir,
app_config.COMPRESS_NAME)
response_object['model_path'] = posixpath.join(target_dir,
pipeline.convert_name)
return jsonify(response_object)
def convert(self, model_name, temp_json, cur_ts, root_path, input_params):
response_object = {'status': 'success'}
# Initiate pipeline object with targeted directory
pipeline = onnxpipeline.Pipeline(
convert_directory=os.path.join(app_config.CONVERT_RES_DIR, cur_ts))
model = pipeline.convert_model(model=model_name, input_json=temp_json)
try:
with open(posixpath.join(pipeline.convert_directory,
'output.json')) as f:
json_data = json.load(f)
response_object['output_json'] = json_data
response_object['logs'] = pipeline.output
response_object['converted_model'] = model
except:
#fail to convert
pass
target_dir = app_config.DOWNLOAD_DIR
input_root = os.path.join(root_path, target_dir)
# compress input directory
compress_path = os.path.join(pipeline.convert_directory,
app_config.TEST_DATA_DIR)
input_path = os.path.join(input_root, cur_ts)
if not os.path.exists(input_path):
os.makedirs(input_path)
tar = tarfile.open(os.path.join(input_path, app_config.COMPRESS_NAME),
"w:gz")
try:
tar.add(compress_path, arcname=app_config.TEST_DATA_DIR)
except:
# fail to generate input
pass
tar.close()
# copy converted onnx model
if os.path.exists(pipeline.convert_path):
copyfile(pipeline.convert_path,
os.path.join(input_root, cur_ts, pipeline.convert_name))
response_object['input_path'] = posixpath.join(target_dir, cur_ts,
app_config.COMPRESS_NAME)
response_object['model_path'] = posixpath.join(target_dir, cur_ts,
pipeline.convert_name)
clean(root_path)
return response_object
def test_pytorch_fail(self):
directory_name = self.deep_dir['pytorch']
pipeline = onnxpipeline.Pipeline(directory_name, convert_directory=self.convert_dir_fail)
def test_convert_fail_no_shapes():
model = pipeline.convert_model(model_type='pytorch', model='saved_model.pb') #model_input_shapes='(1,3,224,224)')
return model
def test_convert_fail_no_model():
model = pipeline.convert_model(model_type='pytorch', model_input_shapes='(1,3,224,224)')
return model
model = test_convert_fail_no_shapes()
output_json = osp.join(pipeline.path, pipeline.convert_directory, config.OUTPUT_JSON)
self.check_json_staus(['FAILED', 'FAILED'], self.check_converted_json(output_json))
model = test_convert_fail_no_model()
output_json = osp.join(pipeline.path, pipeline.convert_directory, config.OUTPUT_JSON)
self.check_json_staus(['FAILED', 'FAILED'], self.check_converted_json(output_json))
def perf_tuning():
response_object = {'status': 'success'}
pipeline = onnxpipeline.Pipeline()
if 'file' in request.files:
_, temp_json = get_params(request, RESERVED_INPUT_PATH)
else:
_, temp_json = get_params(request, './test')
result = pipeline.perf_tuning(input_json=temp_json)
response_object['logs'] = pipeline.output
try:
r = pipeline.get_result(result)
response_object['result'] = json.dumps(r.latency)
response_object['profiling'] = r.profiling_ops
except RuntimeError:
pass
return jsonify(response_object)
def perf_tuning(self, temp_json, input_params):
response_object = {'status': 'success'}
pipeline = onnxpipeline.Pipeline()
# create result dir
result_dir = os.path.join(app_config.PERF_RES_DIR, get_timestamp())
if not os.path.exists(result_dir):
os.makedirs(result_dir)
result = pipeline.perf_tuning(input_json=temp_json, result=result_dir)
try:
r = pipeline.get_result(result)
response_object['result'] = json.dumps(r.latency)
response_object['profiling'] = r.profiling_ops
except RuntimeError:
pass
response_object['logs'] = pipeline.output
# clean up result dir
if os.path.exists(result_dir):
rmtree(result_dir)
clean(app.root_path)
return response_object
def test_constructor_pass(self):
directory_name = self.deep_dir['pytorch']
pipeline = onnxpipeline.Pipeline(directory_name, print_logs=self.print_logs)
self.assertEquals(osp.join(os.getcwd(), directory_name), pipeline.path)
