142 lines
5.8 KiB
Python
142 lines
5.8 KiB
Python
# -*- coding: utf-8 -*-
|
|
# Copyright (C) 2018-2022 Intel Corporation
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
"""Provide a layer of abstraction for an OpenVINO runtime environment."""
|
|
|
|
import logging
|
|
from typing import Dict, List, Union
|
|
|
|
import numpy as np
|
|
|
|
from openvino.runtime import Core
|
|
|
|
from openvino.runtime.exceptions import UserInputError
|
|
from openvino.runtime import Model, Node, Tensor, Type
|
|
from openvino.runtime.utils.types import NumericData, get_shape, get_dtype
|
|
|
|
import tests
|
|
|
|
log = logging.getLogger(__name__)
|
|
|
|
|
|
def runtime(backend_name: str = "CPU") -> "Runtime":
|
|
"""Create a Runtime object (helper factory)."""
|
|
return Runtime(backend_name)
|
|
|
|
|
|
def get_runtime():
|
|
"""Return runtime object."""
|
|
if tests.BACKEND_NAME is not None:
|
|
return runtime(backend_name=tests.BACKEND_NAME)
|
|
else:
|
|
return runtime()
|
|
|
|
|
|
class Runtime(object):
|
|
"""Represents a graph runtime environment."""
|
|
|
|
def __init__(self, backend_name: str) -> None:
|
|
self.backend_name = backend_name
|
|
log.debug(f"Creating runtime for {backend_name}")
|
|
self.backend = Core()
|
|
assert backend_name in self.backend.available_devices, 'The requested device "' + backend_name + '" is not supported!'
|
|
|
|
def set_config(self, config: Dict[str, str]) -> None:
|
|
"""Set the runtime configuration."""
|
|
self.backend.set_property(device_name=self.backend_name, properties=config)
|
|
|
|
def computation(self, node_or_model: Union[Node, Model], *inputs: Node) -> "Computation":
|
|
"""Return a callable Computation object."""
|
|
if isinstance(node_or_model, Node):
|
|
model = Model(node_or_model, inputs, node_or_model.name)
|
|
return Computation(self, model)
|
|
elif isinstance(node_or_model, Model):
|
|
return Computation(self, node_or_model)
|
|
else:
|
|
raise TypeError(
|
|
"Runtime.computation must be called with an OpenVINO Model object "
|
|
"or an OpenVINO node object an optionally Parameter node objects. "
|
|
"Called with: %s",
|
|
node_or_model,
|
|
)
|
|
|
|
def __repr__(self) -> str:
|
|
return f"<Runtime: Backend='{self.backend_name}'>"
|
|
|
|
|
|
class Computation(object):
|
|
"""Graph callable computation object."""
|
|
|
|
def __init__(self, runtime: Runtime, model: Model) -> None:
|
|
self.runtime = runtime
|
|
self.model = model
|
|
self.parameters = model.get_parameters()
|
|
self.results = model.get_results()
|
|
self.network_cache = {}
|
|
|
|
def convert_buffers(self, source_buffers, target_dtypes):
|
|
converted_buffers = []
|
|
for i in range(len(source_buffers)):
|
|
key = list(source_buffers)[i]
|
|
target_dtype = target_dtypes[i]
|
|
# custom conversion for bf16
|
|
if self.results[i].get_output_element_type(0) == Type.bf16:
|
|
converted_buffers.append((source_buffers[key].view(target_dtype)).astype(target_dtype))
|
|
else:
|
|
converted_buffers.append(source_buffers[key].astype(target_dtype))
|
|
return converted_buffers
|
|
|
|
def convert_to_tensors(self, input_values):
|
|
input_tensors = []
|
|
for parameter, input_val in zip(self.parameters, input_values):
|
|
if not isinstance(input_val, (np.ndarray)):
|
|
input_val = np.ndarray([], type(input_val), np.array(input_val))
|
|
if parameter.get_output_element_type(0) == Type.bf16:
|
|
input_tensors.append(Tensor(Type.bf16, input_val.shape))
|
|
input_tensors[-1].data[:] = input_val.view(np.float16)
|
|
else:
|
|
input_tensors.append(Tensor(input_val))
|
|
return input_tensors
|
|
|
|
def __repr__(self) -> str:
|
|
params_string = ", ".join([param.name for param in self.parameters])
|
|
return f"<Computation: {self.model.get_name()}({params_string})>"
|
|
|
|
def __call__(self, *input_values: NumericData) -> List[NumericData]:
|
|
"""Run computation on input values and return result."""
|
|
# Input validation
|
|
if len(input_values) < len(self.parameters):
|
|
raise UserInputError(
|
|
"Expected %s params, received not enough %s values.",
|
|
len(self.parameters),
|
|
len(input_values),
|
|
)
|
|
|
|
param_names = [param.friendly_name for param in self.parameters]
|
|
input_shapes = [get_shape(input_value) for input_value in input_values]
|
|
if self.network_cache.get(str(input_shapes)) is None:
|
|
model = self.model
|
|
self.network_cache[str(input_shapes)] = model
|
|
else:
|
|
model = self.network_cache[str(input_shapes)]
|
|
|
|
compiled_model = self.runtime.backend.compile_model(model, self.runtime.backend_name)
|
|
is_bfloat16 = any(parameter.get_output_element_type(0) == Type.bf16 for parameter in self.parameters)
|
|
if is_bfloat16:
|
|
input_values = self.convert_to_tensors(input_values)
|
|
request = compiled_model.create_infer_request()
|
|
result_buffers = request.infer(dict(zip(param_names, input_values)))
|
|
"""Note: other methods to get result_buffers from request
|
|
First call infer with no return value:
|
|
request.infer(dict(zip(param_names, input_values)))
|
|
Now use any of following options:
|
|
result_buffers = [request.get_tensor(n).data for n in request.outputs]
|
|
result_buffers = [request.get_output_tensor(i).data for i in range(len(request.outputs))]
|
|
result_buffers = [t.data for t in request.output_tensors]
|
|
"""
|
|
# # Since OV overwrite result data type we have to convert results to the original one.
|
|
original_dtypes = [get_dtype(result.get_output_element_type(0)) for result in self.results]
|
|
converted_buffers = self.convert_buffers(result_buffers, original_dtypes)
|
|
return converted_buffers
|