Transition from Legacy Conversion API

In the 2023.1 OpenVINO release OpenVINO Model Converter was introduced with the corresponding Python API: openvino.convert_model method. ovc and openvino.convert_model represent a lightweight alternative of mo and openvino.tools.mo.convert_model which are considered legacy API now. In this article, all the differences between mo and ovc are summarized and the transition guide from the legacy API to the new API is provided.

Parameters Comparison

The comparison of parameters between ov.convert_model() / OVC and mo.convert_model() / MO.

mo.convert_model() / MO

ov.convert_model() / OVC

Differences description

input_model

input_model

Along with model object or path to input model ov.convert_model() accepts list of model parts, for example, the path to TensorFlow weights plus the path to TensorFlow checkpoint. OVC tool accepts an unnamed input model.

output_dir

output_model

output_model in OVC tool sets both output model name and output directory.

model_name

output_model

output_model in OVC tool sets both output model name and output directory.

input

input

ov.convert_model() accepts tuples for setting multiple parameters. OVC tool ‘input’ does not have type setting and freezing functionality. ov.convert_model() does not allow input cut.

output

output

ov.convert_model() does not allow output cut.

input_shape

N/A

Not available in ov.convert_model() / OVC. Can be replaced by input parameter.

example_input

example_input

No differences.

batch

N/A

Not available in ov.convert_model() / OVC. Can be replaced by model reshape functionality. See details below.

mean_values

N/A

Not available in ov.convert_model() / OVC. Can be replaced by functionality from PrePostProcessor. See details below.

scale_values

N/A

Not available in ov.convert_model() / OVC. Can be replaced by functionality from PrePostProcessor. See details below.

scale

N/A

Not available in ov.convert_model() / OVC. Can be replaced by functionality from PrePostProcessor. See details below.

reverse_input_channels

N/A

Not available in ov.convert_model() / OVC. Can be replaced by functionality from PrePostProcessor. See details below.

source_layout

N/A

Not available in ov.convert_model() / OVC. Can be replaced by functionality from PrePostProcessor. See details below.

target_layout

N/A

Not available in ov.convert_model() / OVC. Can be replaced by functionality from PrePostProcessor. See details below.

layout

N/A

Not available in ov.convert_model() / OVC. Can be replaced by functionality from PrePostProcessor. See details below.

compress_to_fp16

compress_to_fp16

OVC provides ‘compress_to_fp16’ for command line tool only, as compression is performed during saving a model to IR (Intermediate Representation).

extensions

extension

No differences.

transform

N/A

Not available in ov.convert_model() / OVC. Can be replaced by functionality from PrePostProcessor. See details below.

transformations_config

N/A

Not available in ov.convert_model() / OVC.

static_shape

N/A

Not available in ov.convert_model() / OVC.

freeze_placeholder_with_value

N/A

Not available in ov.convert_model() / OVC.

use_legacy_frontend

N/A

Not available in ov.convert_model() / OVC.

use_legacy_frontend

N/A

Not available in ov.convert_model() / OVC.

silent

verbose

OVC / ov.convert_model provides ‘verbose’ parameter instead of ‘silent’ for printing of detailed conversion information if ‘verbose’ is set to True.

log_level

N/A

Not available in ov.convert_model() / OVC.

version

version

N/A

progress

N/A

Not available in ov.convert_model() / OVC.

stream_output

N/A

Not available in ov.convert_model() / OVC.

share_weights

share_weights

No differences.

framework

N/A

Not available in ov.convert_model() / OVC.

help / -h

help / -h

OVC provides help parameter only in command line tool.

example_output

output

OVC / ov.convert_model ‘output’ parameter includes capabilities of MO ‘example_output’ parameter.

input_model_is_text

N/A

Not available in ov.convert_model() / OVC.

input_checkpoint

input_model

All supported model formats can be passed to ‘input_model’.

input_meta_graph

input_model

All supported model formats can be passed to ‘input_model’.

saved_model_dir

input_model

All supported model formats can be passed to ‘input_model’.

saved_model_tags

N/A

Not available in ov.convert_model() / OVC.

tensorflow_custom_operations_config_update

N/A

Not available in ov.convert_model() / OVC.

tensorflow_object_detection_api_pipeline_config

N/A

Not available in ov.convert_model() / OVC.

tensorboard_logdir

N/A

Not available in ov.convert_model() / OVC.

tensorflow_custom_layer_libraries

N/A

Not available in ov.convert_model() / OVC.

input_symbol

N/A

Not available in ov.convert_model() / OVC.

nd_prefix_name

N/A

Not available in ov.convert_model() / OVC.

pretrained_model_name

N/A

Not available in ov.convert_model() / OVC.

save_params_from_nd

N/A

Not available in ov.convert_model() / OVC.

legacy_mxnet_model

N/A

Not available in ov.convert_model() / OVC.

enable_ssd_gluoncv

N/A

Not available in ov.convert_model() / OVC.

input_proto

N/A

Not available in ov.convert_model() / OVC.

caffe_parser_path

N/A

Not available in ov.convert_model() / OVC.

k

N/A

Not available in ov.convert_model() / OVC.

disable_omitting_optional

N/A

Not available in ov.convert_model() / OVC.

enable_flattening_nested_params

N/A

Not available in ov.convert_model() / OVC.

counts

N/A

Not available in ov.convert_model() / OVC.

remove_output_softmax

N/A

Not available in ov.convert_model() / OVC.

remove_memory

N/A

Not available in ov.convert_model() / OVC.

Transition from Legacy API to New API

mo.convert_model() provides a wide range of preprocessing parameters. Most of these parameters have analogs in OVC or can be replaced with functionality from ov.PrePostProcessor class. Here is the guide to transition from legacy model preprocessing to new API preprocessing.

input_shape

Legacy API

New API

from openvino.tools import mo

ov_model = mo.convert_model(model, input_shape=[[1, 3, 100, 100],[1]])
import openvino as ov

ov_model = ov.convert_model(model, input=[[1, 3, 100, 100],[1]])

Legacy API

New API

mo --input_model MODEL_NAME --input_shape [1,3,100,100],[1] --output_dir OUTPUT_DIR
ovc MODEL_NAME --input [1,3,100,100],[1] --output_model OUTPUT_MODEL

batch

Legacy API

New API

from openvino.tools import mo

ov_model = mo.convert_model(model, batch=2)
import openvino as ov

ov_model = ov.convert_model(model)
input_shape = ov_model.inputs[0].partial_shape
input_shape[0] = 2 # batch size
ov_model.reshape(input_shape)

Legacy API

New API

mo --input_model MODEL_NAME --batch 2 --output_dir OUTPUT_DIR

Not available in OVC tool. Please check Python API.

mean_values

Legacy API

New API

from openvino.tools import mo

ov_model = mo.convert_model(model, mean_values=[0.5, 0.5, 0.5])
import openvino as ov

ov_model = ov.convert_model(model)

prep = ov.preprocess.PrePostProcessor(ov_model)
prep.input(input_name).tensor().set_layout(ov.Layout("NHWC"))
prep.input(input_name).preprocess().mean([0.5, 0.5, 0.5])
ov_model = prep.build()

There is currently no heuristic for automatic detection of the channel to which mean, scale or reverse channels should be applied. Layout needs to be explicitly specified with “C” channel. For example “NHWC”, “NCHW”, “?C??”. See also Layout API overview.

Legacy API

New API

mo --input_model MODEL_NAME --mean_values [0.5,0.5,0.5] --output_dir OUTPUT_DIR

Not available in OVC tool. Please check Python API.

scale_values

Legacy API

New API

from openvino.tools import mo

ov_model = mo.convert_model(model, scale_values=[255., 255., 255.])
import openvino as ov

ov_model = ov.convert_model(model)

prep = ov.preprocess.PrePostProcessor(ov_model)
prep.input(input_name).tensor().set_layout(ov.Layout("NHWC"))
prep.input(input_name).preprocess().scale([255., 255., 255.])
ov_model = prep.build()

There is currently no heuristic for automatic detection of the channel to which mean, scale or reverse channels should be applied. Layout needs to be explicitly specified with “C” channel. For example “NHWC”, “NCHW”, “?C??”. See also Layout API overview.

Legacy API

New API

mo --input_model MODEL_NAME --scale_values [255,255,255] --output_dir OUTPUT_DIR

Not available in OVC tool. Please check Python API.

reverse_input_channels

Legacy API

New API

from openvino.tools import mo

ov_model = mo.convert_model(model, reverse_input_channels=True)
import openvino as ov

ov_model = ov.convert_model(model)

prep = ov.preprocess.PrePostProcessor(ov_model)
prep.input(input_name).tensor().set_layout(ov.Layout("NHWC"))
prep.input(input_name).preprocess().reverse_channels()
ov_model = prep.build()

There is currently no heuristic for automatic detection of the channel to which mean, scale or reverse channels should be applied. Layout needs to be explicitly specified with “C” channel. For example “NHWC”, “NCHW”, “?C??”. See also Layout API overview.

Legacy API

New API

mo --input_model MODEL_NAME --reverse_input_channels --output_dir OUTPUT_DIR

Not available in OVC tool. Please check Python API.

source_layout

Legacy API

New API

import openvino as ov
from openvino.tools import mo

ov_model = mo.convert_model(model, source_layout={input_name: ov.Layout("NHWC")})
import openvino as ov

ov_model = ov.convert_model(model)

prep = ov.preprocess.PrePostProcessor(ov_model)
prep.input(input_name).model().set_layout(ov.Layout("NHWC"))
ov_model = prep.build()

Legacy API

New API

mo --input_model MODEL_NAME --source_layout input_name(NHWC) --output_dir OUTPUT_DIR

Not available in OVC tool. Please check Python API.

target_layout

Legacy API

New API

import openvino as ov
from openvino.tools import mo

ov_model = mo.convert_model(model, target_layout={input_name: ov.Layout("NHWC")})
import openvino as ov

ov_model = ov.convert_model(model)

prep = ov.preprocess.PrePostProcessor(ov_model)
prep.input(input_name).tensor().set_layout(ov.Layout("NHWC"))
ov_model = prep.build()

Legacy API

New API

mo --input_model MODEL_NAME --target_layout input_name(NHWC) --output_dir OUTPUT_DIR

Not available in OVC tool. Please check Python API.

layout

Legacy API

New API

from openvino.tools import mo

ov_model = mo.convert_model(model, layout={input_name: mo.LayoutMap("NCHW", "NHWC")})
import openvino as ov

ov_model = ov.convert_model(model)

prep = ov.preprocess.PrePostProcessor(ov_model)
prep.input(input_name).model().set_layout(ov.Layout("NCHW"))
prep.input(input_name).tensor().set_layout(ov.Layout("NHWC"))
ov_model = prep.build()

Legacy API

New API

mo --input_model MODEL_NAME --layout "input_name(NCHW->NHWC)" --output_dir OUTPUT_DIR

Not available in OVC tool. Please check Python API.

transform

Legacy API

New API

from openvino.tools import mo

ov_model = mo.convert_model(model, transform=[('LowLatency2', {'use_const_initializer': False}), 'Pruning', ('MakeStateful', {'param_res_names': {'input_name': 'output_name'}})])
import openvino as ov
from openvino._offline_transformations import apply_low_latency_transformation, apply_pruning_transformation, apply_make_stateful_transformation

ov_model = ov.convert_model(model)
apply_low_latency_transformation(model, use_const_initializer=False)
apply_pruning_transformation(model)
apply_make_stateful_transformation(model, param_res_names={'input_name': 'output_name'})

Legacy API

New API

mo --input_model MODEL_NAME --transform LowLatency2[use_const_initializer=False],Pruning,MakeStateful[param_res_names={'input_name':'output_name'}] --output_dir OUTPUT_DIR

Not available in OVC tool. Please check Python API.

Cutting Off Parts of a Model

Performing surgery by cutting model inputs and outputs from a model is no longer available in the new conversion API. Instead, we recommend performing the cut in the original framework. Below are examples of model cutting of TensorFlow protobuf, TensorFlow SavedModel, and ONNX formats with the legacy conversion API, compared to achieving the same cut with tools provided by the Tensorflow and ONNX frameworks. For PyTorch, TensorFlow 2 Keras, and PaddlePaddle, we recommend changing the original model code to perform the model cut.

Note: This guide does not cover the cutting a model by input port of an operation that MO tool provides using input and output options, for example, –input 1:name_op.

PyTorch

Model cut for PyTorch is not available in legacy API.

When it is needed to remove a whole module from the model it is possible to replace such modules with Identity. Below is the example of removing conv1 and bn1 modules at the input and fc module at the output of the resnet50 model.

import openvino as ov
import torch
import torchvision
from torch.nn import Identity

# Load pretrained model
model = torchvision.models.resnet50(weights='DEFAULT')

# input cut
model.conv1 = Identity()
model.bn1 = Identity()

# output cut
model.fc = Identity()

# convert and compile the model
ov_model = ov.convert_model(model, input=([-1,64,-1,-1], torch.float32))
compiled_model = ov.compile_model(ov_model)

When it is needed to remove one or more outputs from the model it is possible to create a wrapper for the model and only output the needed output. Below is the example of removing second output from the model.

import openvino as ov
import torch

# Example of model with multiple outputs
class Model(torch.nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.linear1 = torch.nn.Linear(100, 200)
        self.activation1 = torch.nn.ReLU()
        self.linear2 = torch.nn.Linear(200, 10)
        self.activation2 = torch.nn.Sigmoid()

    def forward(self, x):
        x = self.linear1(x)
        x = self.activation1(x)
        y = self.linear2(x)
        y = self.activation2(y)
        return x, y

# New model, where some outputs are cut
class CutModel(torch.nn.Module):
    def __init__(self):
        super(CutModel, self).__init__()
        self.model = Model()

    def forward(self, x):

        # get first output
        x, _ = self.model(x)

        return x

# Model with output cut
cut_model = CutModel()

# convert and compile the model
ov_model = ov.convert_model(cut_model, input=([-1,-1,-1], torch.float32))
compiled_model = ov.compile_model(ov_model)

TensorFlow protobuf format / tf.Graph / tf.GraphDef

Legacy API.

import openvino as ov
import openvino.tools.mo as mo

import tensorflow as tf

def load_graph(model_path):
    graph_def = tf.compat.v1.GraphDef()
    with open(model_path, "rb") as f:
        graph_def.ParseFromString(f.read())
    with tf.compat.v1.Graph().as_default() as graph:
        tf.graph_util.import_graph_def(graph_def, name="")
        return graph

# Load TF model
graph = load_graph("/path_to_model/HugeCTR.pb")

# Convert the model with input and output cut
input_name = "concat"
output_name = "MatVec_3/Squeeze"
ov_model = mo.convert_model(graph, input=(input_name, [-1, -1]), output=output_name)

# Compile the model
compiled_model = ov.compile_model(ov_model)

Model cut in original FW.

import openvino as ov
import tensorflow as tf

from tensorflow.python.tools.strip_unused_lib import strip_unused

def load_graph(model_path):
    graph_def = tf.compat.v1.GraphDef()
    with open(model_path, "rb") as f:
        graph_def.ParseFromString(f.read())
    with tf.compat.v1.Graph().as_default() as graph:
        tf.graph_util.import_graph_def(graph_def, name="")
        return graph

# Load TF model
graph = load_graph("/path_to_model/HugeCTR.pb")

# Cut the model
input_name = "concat"
output_name = "MatVec_3/Squeeze"
graph_def = graph.as_graph_def()
new_graph_def = strip_unused(graph_def, [input_name], [output_name], tf.float32.as_datatype_enum)

# Convert and compile model
ov_model = ov.convert_model(new_graph_def, input=[-1, -1])
cmp_model = ov.compile_model(ov_model)

TensorFlow SavedModel format

Model cut for SavedModel format is not available in legacy API.

Example of model cut in original FW.

import openvino as ov
import tensorflow_hub as hub

import tensorflow as tf
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
from tensorflow.python.tools.strip_unused_lib import strip_unused

# Load TF model
model = hub.load("https://tfhub.dev/svampeatlas/vision/embedder/fungi_V2/1?tf-hub-format=compressed")

# Convert model to GraphDef
model_func = model.signatures["default"]
frozen_func = convert_variables_to_constants_v2(model_func)
graph_def = frozen_func.graph.as_graph_def()

# Cut the model
input_name = 'InceptionV4/InceptionV4/Conv2d_2b_3x3/Relu'
output_name = 'InceptionV4/InceptionV4/Mixed_7c/concat'
new_graph_def = strip_unused(graph_def, [input_name], [output_name], tf.float32.as_datatype_enum)

# Convert and compile the model
ov_model = ov.convert_model(new_graph_def)
compiled_model = ov.compile_model(ov_model)

ONNX

Legacy API.

import openvino as ov
import openvino.tools.mo as mo

input_path = "/path_to_model/yolov8x.onnx"

# Convert model and perform input and output cut
input_name = "/model.2/Concat_output_0"
output_name = "/model.22/Concat_3_output_0"
ov_model = mo.convert_model(input_path, input=input_name, output=output_name)

# Compile model
ov.compile_model(ov_model)

Model cut in original FW.

import onnx
import openvino as ov

input_path = "/path_to_model/yolov8x.onnx"

# Cut the model
input_name = "/model.2/Concat_output_0"
output_name = "/model.22/Concat_3_output_0"
cut_model_path = "/path_to_model/yolov8x_cut.onnx"
onnx.utils.extract_model(input_path, cut_model_path, [input_name], [output_name])

# Convert model
ov_model = ov.convert_model(cut_model_path)

# Compile model
ov.compile_model(ov_model)

Supported Frameworks in MO vs OVC

ov.convert_model() and OVC tool support conversion from PyTorch, TF, TF Lite, ONNX, PaddlePaddle. The following frameworks are supported only in MO and mo.convert_model(): Caffe, MxNet, Kaldi.