Converting a PyTorch Model#

You can download a PyTorch model from PyTorch Hub or Hugging Face. To convert the model, use the openvino.convert_model function.

Here is the simplest example of PyTorch model conversion using a model from torchvision:

import torchvision
import torch
import openvino as ov

model = torchvision.models.resnet50(weights='DEFAULT')
ov_model = ov.convert_model(model)

openvino.convert_model function supports the following PyTorch model object types:

  • torch.nn.Module derived classes

  • torch.jit.ScriptModule

  • torch.jit.ScriptFunction

  • torch.export.ExportedProgram

When using torch.nn.Module as an input model, openvino.convert_model often requires the example_input parameter to be specified. Internally, it triggers the model tracing during the model conversion process, using the capabilities of the torch.jit.trace function.

The use of example_input can lead to a better quality OpenVINO model in terms of correctness and performance compared to converting the same original model without specifying example_input. While the necessity of example_input depends on the implementation details of a specific PyTorch model, it is recommended to always set the example_input parameter when it is available.

The value for the example_input parameter can be easily derived from knowing the input tensor’s element type and shape. While it may not be suitable for all cases, random numbers can frequently serve this purpose effectively:

import torchvision
import torch
import openvino as ov

model = torchvision.models.resnet50(weights='DEFAULT')
ov_model = ov.convert_model(model, example_input=torch.rand(1, 3, 224, 224))

In practice, the code to evaluate or test the PyTorch model is usually provided with the model itself and can be used to generate a proper example_input value. A modified example of using resnet50 model from torchvision is presented below. It demonstrates how to switch inference in the existing PyTorch application to OpenVINO and how to get value for example_input:

from torchvision.io import read_image
from torchvision.models import resnet50, ResNet50_Weights
import requests, PIL, io, torch

# Get a picture of a cat from the web:
img = PIL.Image.open(io.BytesIO(requests.get("https://placekitten.com/200/300").content))

# Torchvision model and input data preparation from https://pytorch.org/vision/stable/models.html

weights = ResNet50_Weights.DEFAULT
model = resnet50(weights=weights)
model.eval()
preprocess = weights.transforms()
batch = preprocess(img).unsqueeze(0)

# PyTorch model inference and post-processing

prediction = model(batch).squeeze(0).softmax(0)
class_id = prediction.argmax().item()
score = prediction[class_id].item()
category_name = weights.meta["categories"][class_id]
print(f"{category_name}: {100 * score:.1f}% (with PyTorch)")

# OpenVINO model preparation and inference with the same post-processing

import openvino as ov
compiled_model = ov.compile_model(ov.convert_model(model, example_input=batch))

prediction = torch.tensor(compiled_model(batch)[0]).squeeze(0).softmax(0)
class_id = prediction.argmax().item()
score = prediction[class_id].item()
category_name = weights.meta["categories"][class_id]
print(f"{category_name}: {100 * score:.1f}% (with OpenVINO)")

Check out more examples in interactive Python tutorials.

Note

In the examples above the openvino.save_model function is not used because there are no PyTorch-specific details regarding the usage of this function. In all examples, the converted OpenVINO model can be saved to IR by calling ov.save_model(ov_model, 'model.xml') as usual.

Supported Input Parameter Types#

If the model has a single input, the following input types are supported in example_input:

  • openvino.Tensor

  • torch.Tensor

  • tuple or any nested combination of tuples

If a model has multiple inputs, the input values are combined in a list, a tuple, or a dict:

  • values in a list or tuple should be passed in the same order as the original model specifies,

  • dict has keys from the names of the original model argument names.

Enclosing in list, tuple or dict can be used for a single input as well as for multiple inputs.

If a model has a single input parameter and the type of this input is a tuple, it should be always passed enclosed into an extra list, tuple or dict as in the case of multiple inputs. It is required to eliminate ambiguity between model((a, b)) and model(a, b) in this case.

Non-tensor Data Types#

When a non-tensor data type, such as a tuple or dict, appears in a model input or output, it is flattened. The flattening means that each element within the tuple will be represented as a separate input or output. The same is true for dict values, where the keys of the dict are used to form a model input/output name. The original non-tensor input or output is replaced by one or multiple new inputs or outputs resulting from this flattening process. This flattening procedure is applied recursively in the case of nested tuples, lists, and dicts until it reaches the assumption that the most nested data type is a tensor.

For example, if the original model is called with example_input=(a, (b, c, (d, e))), where a, b, … e are tensors, it means that the original model has two inputs. The first is a tensor a, and the second is a tuple (b, c, (d, e)), containing two tensors b and c and a nested tuple (d, e). Then the resulting OpenVINO model will have signature (a, b, c, d, e), which means it will have five inputs, all of type tensor, instead of two in the original model.

If your model has a dict input, such as, {"x": a, "y": b, "z": c}, it will be decomposed into multiple inputs of the OpenVINO model signature: (a, b, c), where inputs assume the names of x, y, and z respectively.

Note

An important consequence of flattening is that only tuple and dict with a fixed number of elements and key values are supported. The structure of such inputs should be fully described in the example_input parameter of convert_model. The flattening on outputs should be reproduced with the given example_input and cannot be changed once the conversion is done.

Check out more examples of model conversion with non-tensor data types in the following tutorials:

Input and output names of the model#

PyTorch doesn’t produce relevant names for model inputs and outputs in the TorchScript representation. OpenVINO will assign input names based on the signature of models’s forward method or dict keys provided in the example_input. Output names will be assigned if there is a dict at the output or when there is some internal name available in the TorchScript model representation. In general, the output name is not assigned and stays empty. It is recommended to address model outputs by the index rather then the name.

Support for torch.export#

torch.export is the current way to get a graph representation of a model (since PyTorch 2.1). It produces ExportedProgram which includes the graph representation in the FX format. To see why it has an advantage over the TorchScript representation, refer to PyTorch documentation.

Here is an example of how to convert a model obtained with torch.export:

from torchvision.models import resnet50, ResNet50_Weights
from torch.export import export
from openvino import convert_model

model = resnet50(weights=ResNet50_Weights.DEFAULT)
model.eval()
exported_model = export(model, (torch.randn(1, 3, 224, 224),))
ov_model = convert_model(exported_model)

Note

This is an experimental feature. Use it only if you know that you need to. PyTorch version 2.2 is recommended. Dynamic shapes are not supported yet.

Exporting a PyTorch Model to ONNX Format#

An alternative method of converting PyTorch models is exporting a PyTorch model to ONNX with torch.onnx.export first and then converting the resulting .onnx file to OpenVINO Model with openvino.convert_model. It can be considered as a backup solution if a model cannot be converted directly from PyTorch to OpenVINO as described in the above chapters. Converting through ONNX can be more expensive in terms of code, conversion time, and allocated memory.

  1. Refer to the Exporting PyTorch models to ONNX format guide to learn how to export models from PyTorch to ONNX.

  2. Follow Convert an ONNX model chapter to produce OpenVINO model.

Here is an illustration of using these two steps together:

import torchvision
import torch
import openvino as ov

model = torchvision.models.resnet50(weights='DEFAULT')
# 1. Export to ONNX
torch.onnx.export(model, (torch.rand(1, 3, 224, 224), ), 'model.onnx')
# 2. Convert to OpenVINO
ov_model = ov.convert_model('model.onnx')

Note

As of version 1.8.1, not all PyTorch operations can be exported to ONNX opset 9 which is used by default. It is recommended to export models to opset 11 or higher when export to default opset 9 is not working. In that case, use opset_version option of the torch.onnx.export. For more information about ONNX opset, refer to the Operator Schemas page.