Setting Input Shapes¶
With model conversion API you can increase your model’s efficiency by providing an additional shape definition using the
Specifying Shapes in the
openvino.convert_model supports conversion of models with dynamic input shapes that contain undefined dimensions.
However, if the shape of data is not going to change from one inference request to another,
it is recommended to set up static shapes (when all dimensions are fully defined) for the inputs.
Doing it at this stage, instead of during inference in runtime, can be beneficial in terms of performance and memory consumption.
To set up static shapes, model conversion API provides the
For more information on changing input shapes in runtime, refer to the Changing input shapes guide.
To learn more about dynamic shapes in runtime, refer to the Dynamic Shapes guide.
The OpenVINO Runtime API may present certain limitations in inferring models with undefined dimensions on some hardware. See the Features support matrix for reference.
In this case, the
input parameter and the reshape method can help to resolve undefined dimensions.
For example, run model conversion for the TensorFlow MobileNet model with the single input
and specify the input shape of
import openvino as ov ov_model = ov.convert_model("MobileNet.pb", input=[2, 300, 300, 3])
ovc MobileNet.pb --input [2,300,300,3]
If a model has multiple inputs, the input shape should be specified in
input parameter as a list. In
ovc, this is a command separate list, and in
openvino.convert_model this is a Python list or tuple with number of elements matching the number of inputs in the model. Use input names from the original model to define the mapping between inputs and shapes specified.
The following example demonstrates the conversion of the ONNX OCR model with a pair of inputs
and specifies shapes
 for them respectively:
import openvino as ov ov_model = ov.convert_model("ocr.onnx", input=[("data", [3,150,200,1]), ("seq_len", )])
ovc ocr.onnx --input data[3,150,200,1],seq_len
If the order of inputs is defined in the input model and the order is known for the user, names could be omitted. In this case, it is important to specify shapes in the same order of input model inputs:
import openvino as ov ov_model = ov.convert_model("ocr.onnx", input=([3,150,200,1], ))
ovc ocr.onnx --input [3,150,200,1],
Whether the model has a specified order of inputs depends on the original framework. Usually, it is convenient to set shapes without specifying the names of the parameters in the case of PyTorch model conversion because a PyTorch model is considered as a callable that usually accepts positional parameters. On the other hand, names of inputs are convenient when converting models from model files, because naming of inputs is a good practice for many frameworks that serialize models to files.
input parameter allows overriding original input shapes if it is supported by the model topology.
Shapes with dynamic dimensions in the original model can be replaced with static shapes for the converted model, and vice versa.
The dynamic dimension can be marked in model conversion API parameter as
? when using
For example, launch model conversion for the ONNX OCR model and specify dynamic batch dimension for inputs:
import openvino as ov ov_model = ov.convert_model("ocr.onnx", input=[("data", [-1, 150, 200, 1]), ("seq_len", [-1])])
ovc ocr.onnx --input "data[?,150,200,1],seq_len[?]"
To optimize memory consumption for models with undefined dimensions in run-time, model conversion API provides the capability to define boundaries of dimensions.
The boundaries of undefined dimension can be specified with ellipsis in the command line or with
openvino.Dimension class in Python.
For example, launch model conversion for the ONNX OCR model and specify a boundary for the batch dimension 1..3, which means that the input tensor will have batch dimension minimum 1 and maximum 3 in inference:
import openvino as ov batch_dim = ov.Dimension(1, 3) ov_model = ov.convert_model("ocr.onnx", input=[("data", [batch_dim, 150, 200, 1]), ("seq_len", [batch_dim])])
ovc ocr.onnx --input data[1..3,150,200,1],seq_len[1..3]
In practice, not every model is designed in a way that allows change of input shapes. An attempt to change the shape for such models may lead to an exception during model conversion, later in model inference, or even to wrong results of inference without explicit exception raised. A knowledge about model topology is required to set shapes appropriately. For more information about shape follow the inference troubleshooting and ways to relax shape inference flow guides.