[LEGACY] Cutting Off Parts of a Model#

Danger

The code described here has been deprecated! Do not use it to avoid working with a legacy solution. It will be kept for some time to ensure backwards compatibility, but you should not use it in contemporary applications.

Sometimes, it is necessary to remove parts of a model when converting it to OpenVINO IR. This chapter describes how to do it, using model conversion API parameters. Model cutting applies mostly to TensorFlow models, which is why TensorFlow will be used in this chapter’s examples, but it may be also useful for other frameworks.

Purpose of Model Cutting#

The following examples are the situations when model cutting is useful or even required:

A model has pre- or post-processing parts that cannot be translated to existing OpenVINO operations.
A model has a training part that is convenient to be kept in the model but not used during inference.
A model is too complex be converted at once, because it contains a lot of unsupported operations that cannot be easily implemented as custom layers.
A problem occurs with model conversion or inference in OpenVINO™ Runtime. To identify the issue, limit the conversion scope by iterative search for problematic areas in the model.
A single custom layer or a combination of custom layers is isolated for debugging purposes.

Note

Internally, when you run model conversion API, it loads the model, goes through the topology, and tries to find each layer type in a list of known layers. Custom layers are layers that are not included in the list. If your topology contains such kind of layers, model conversion API classifies them as custom.

Model conversion API parameters#

Model conversion API provides input and output command-line options to specify new entry and exit nodes, while ignoring the rest of the model:

input option accepts a list of layer names of the input model that should be treated as new entry points to the model. See the full list of accepted types for input on Model Conversion Python API page.
output option accepts a list of layer names of the input model that should be treated as new exit points from the model.

The input option is required for cases unrelated to model cutting. For example, when the model contains several inputs and input_shape or mean_values options are used, the input option specifies the order of input nodes for correct mapping between multiple items provided in input_shape and mean_values and the inputs in the model.

Model cutting is illustrated with the Inception V1 model, found in the models/research/slim repository. To proceed with this chapter, make sure you do the necessary steps to prepare the model for model conversion.

Default Behavior without input and output#

The input model is converted as a whole if neither input nor output command line options are used. All Placeholder operations in a TensorFlow graph are automatically identified as entry points. The Input layer type is generated for each of them. All nodes that have no consumers are automatically identified as exit points.

For Inception_V1, there is one Placeholder: input. If the model is viewed in TensorBoard, the input operation is easy to find:

Reshape is the only output operation, which is enclosed in a nested name scope of InceptionV1/Logits/Predictions, under the full name of InceptionV1/Logits/Predictions/Reshape_1.

In TensorBoard, along with some of its predecessors, it looks as follows:

Convert this model to ov.Model:

Python

from openvino.tools.mo import convert_model
ov_model = convert_model("inception_v1.pb", batch=1)

CLI

mo --input_model inception_v1.pb -b 1 --output_dir <OUTPUT_MODEL_DIR>

ov.Model can be serialized with the ov.serialize() method to Intermediate Representation which can be used for model structure exploring. In IR, the structure of a model has the following layers:

<layer id="286" name="input" precision="FP32" type="Input">
    <output>
        <port id="0">
            <dim>1</dim>
            <dim>3</dim>
            <dim>224</dim>
            <dim>224</dim>
        </port>
    </output>
</layer>

The input layer is converted from the TensorFlow graph Placeholder operation input and has the same name.

The -b option is used here for conversion to override a possible undefined batch size (coded as -1 in TensorFlow models). If a model was frozen with a defined batch size, you may omit this option in all the examples.

The last layer in the model is InceptionV1/Logits/Predictions/Reshape_1, which matches an output operation in the TensorFlow graph:

<layer id="389" name="InceptionV1/Logits/Predictions/Reshape_1" precision="FP32" type="Reshape">
    <data axis="0" dim="1,1001" num_axes="-1"/>
    <input>
        <port id="0">
            <dim>1</dim>
            <dim>1001</dim>
        </port>
    </input>
    <output>
        <port id="1">
            <dim>1</dim>
            <dim>1001</dim>
        </port>
    </output>
</layer>

Due to automatic identification of inputs and outputs, providing the input and output options to convert the whole model is not required. The following commands are equivalent for the Inception V1 model:

Python

from openvino.tools.mo import convert_model
ov_model = convert_model("inception_v1.pb", batch=1)

ov_model = convert_model("inception_v1.pb", batch=1, input="input", output="InceptionV1/Logits/Predictions/Reshape_1")

CLI

mo --input_model inception_v1.pb -b 1 --output_dir <OUTPUT_MODEL_DIR>

mo --input_model inception_v1.pb -b 1 --input input --output InceptionV1/Logits/Predictions/Reshape_1 --output_dir <OUTPUT_MODEL_DIR>

The Intermediate Representations are identical for both conversions. The same is true if the model has multiple inputs and/or outputs.

Inputs with Multiple Input Ports#

There are operations that contain more than one input port. In the example considered here, the convolution InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution is such operation. When input_shape is not provided, a new Input layer is created for each dynamic input port for the node. If a port is evaluated to a constant blob, this constant remains in the model and a corresponding input layer is not created. TensorFlow convolution used in this model contains two ports:

port 0: input tensor for convolution (dynamic)
port 1: convolution weights (constant)

Following this behavior, convert_model() creates an Input layer for port 0 only, leaving port 1 as a constant. Thus, the result of:

Python

from openvino.tools.mo import convert_model
ov_model = convert_model("inception_v1.pb", batch=1, input="InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution")

CLI

mo --input_model inception_v1.pb -b 1 --input InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution --output_dir <OUTPUT_MODEL_DIR>

is identical to the result of conversion of the model as a whole, because this convolution is the first executable operation in Inception V1.

Different behavior occurs when input_shape is also used as an attempt to override the input shape:

Python

from openvino.tools.mo import convert_model
ov_model = convert_model("inception_v1.pb", input="InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution", input_shape=[1,224,224,3])

CLI

mo --input_model inception_v1.pb--input=InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution --input_shape [1,224,224,3]  --output_dir <OUTPUT_MODEL_DIR>

An error occurs (for more information, see the Model Conversion FAQ):

[ ERROR ]  Node InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution has more than 1 input and input shapes were provided.
Try not to provide input shapes or specify input port with PORT:NODE notation, where PORT is an integer.
For more information, see FAQ #30

When input_shape is specified and the node contains multiple input ports, you need to provide an input port index together with an input node name. The input port index is specified in front of the node name with ‘:’ as a separator (PORT:NODE). In this case, the port index 0 of the node InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution should be specified as 0:InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution.

The correct command line is:

Python

from openvino.tools.mo import convert_model
ov_model = convert_model("inception_v1.pb", input="0:InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution", input_shape=[1,224,224,3])

CLI

mo --input_model inception_v1.pb --input 0:InceptionV1/InceptionV1/Conv2d_1a_7x7/convolution --input_shape=[1,224,224,3] --output_dir <OUTPUT_MODEL_DIR>

[LEGACY] Cutting Off Parts of a Model#

Purpose of Model Cutting#

Model conversion API parameters#

Default Behavior without input and output#

Model Cutting#

Cutting at the End#

Cutting from the Beginning#

Inputs with Multiple Input Ports#