Setting Input Shapes¶
Model Optimizer provides the option of making models more efficient by providing additional shape definition. It is achieved with two parameters:
--static_shape, used under certain conditions.
When to Specify input_shape Command-line Parameter¶
Model Optimizer 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 to another, it is recommended to set up static shapes (when all dimensions are fully defined) for the inputs. It can be beneficial from a performance perspective and memory consumption. To set up static shapes, Model Optimizer provides the
--input_shape parameter. The same functionality is also available in runtime via
reshape method, please refer to Changing input shapes. For more information about dynamic shapes in runtime, refer to Dynamic Shapes
OpenVINO Runtime API can have limitations to infer models with undefined dimensions on some hardware (see Features support matrix for reference). In this case, the
--input_shape parameter and the reshape method can help to resolve undefined dimensions.
Sometimes Model Optimizer is unable to convert models out-of-the-box (only the
--input_model parameter is specified). Such problem can relate to models with inputs of undefined ranks and a case of cutting off parts of a model. In this case, user has to specify input shapes explicitly using
For example, run the Model Optimizer for the TensorFlow* MobileNet model with the single input and specify input shape
mo --input_model MobileNet.pb --input_shape [2,300,300,3]
If a model has multiple inputs,
--input_shape must be used in conjunction with
--input parameter. The parameter
--input contains a list of input names for which shapes in the same order are defined via
--input_shape. For example, launch the Model Optimizer for the ONNX* OCR model with a pair of inputs
seq_len and specify shapes
 for them.
mo --input_model ocr.onnx --input data,seq_len --input_shape [3,150,200,1],
The alternative way to specify input shapes is to use the
--input parameter as follows:
mo --input_model ocr.onnx --input data[3 150 200 1],seq_len
--input_shape allows overriding original input shapes to the shapes compatible with a given model. Dynamic shapes, i.e. 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 Optimizer command-line as
?. For example, launch the Model Optimizer for the ONNX* OCR model and specify dynamic batch dimension for inputs.
mo --input_model ocr.onnx --input data,seq_len --input_shape [-1,150,200,1],[-1]
To optimize memory consumption for models with undefined dimensions in run-time, Model Optimizer provides the capability to define boundaries of dimensions. The boundaries of undefined dimension can be specified with ellipsis. For example, launch the Model Optimizer for the ONNX* OCR model and specify a boundary for the batch dimension.
mo --input_model ocr.onnx --input data,seq_len --input_shape [1..3,150,200,1],[1..3]
Practically, some models are not ready for input shapes change. In this case, a new input shape cannot be set via Model Optimizer. Learn more about shape inference troubleshooting and ways to relax shape inference flow.
When to Specify static_shape Command-line Parameter¶
Model Optimizer provides the
--static_shape parameter that allows evaluating shapes of all operations in the model for fixed input shapes and to fold shape computing sub-graphs into constants. The resulting IR can be more compact in size and the loading time for such IR can be decreased. However, the resulting IR will not be reshape-able with the help of the reshape method from OpenVINO Runtime API. It is worth noting that the
--input_shape parameter does not affect reshape-ability of the model.
For example, launch the Model Optimizer for the ONNX* OCR model using
mo --input_model ocr.onnx --input data[3 150 200 1],seq_len --static_shape