View the runnable example on GitHub

Save and Load Optimized IPEX Model#

This example illustrates how to save and load a model accelerated by IPEX. In this example, we use a ResNet18 model pretrained. Then, by calling InferenceOptimizer.trace(..., use_ipex=True), we can obtain a model accelerated by IPEX method. By calling, path) , we could save the model to a folder. By calling InferenceOptimizer.load(path), we could load the model from a folder.

To inference using Bigdl-nano InferenceOptimizer, the following packages need to be installed first. We recommend you to use Miniconda to prepare the environment and install the following packages in a conda environment.

You can create a conda environment by executing:

# "nano" is conda environment name, you can use any name you like.
conda create -n nano python=3.7 setuptools=58.0.4
conda activate nano

📝 Note

During your installation, there may be some warnings or errors about version, just ignore them.

[ ]:
# Necessary packages for inference accelaration
!pip install --pre --upgrade bigdl-nano[pytorch]

First, prepare model. We need load the pretrained ResNet18 model.

[ ]:
import torch
from torchvision.models import resnet18

model_ft = resnet18(pretrained=True)

Accelerate Inference Using IPEX

[ ]:
from bigdl.nano.pytorch import InferenceOptimizer
ipex_model = InferenceOptimizer.trace(model_ft,

Save Optimized IPEX Model The saved model files will be saved at “./optimized_model_ipex” directory There are 2 files in optimized_model_ipex, users only need to take “ckpt.pth” file for further usage:

  • nano_model_meta.yml: meta information of the saved model checkpoint

  • ckpt.pth: pytorch state dict checkpoint for general use, describes model structure

[ ]:, "./optimized_model_ipex")

Load the Optimized Model

📝 Note

  • For a model accelerated by JIT, OpenVINO or ONNXRuntime, we saved the structure of its network, so we don’t need its unaccelerated model when we load the optimized model.

  • For a model accelerated by IPEX, we only store the state_dict which is simply a python dictionary object that maps each layer to its parameter tensor when saving the model, so when we load the optimized model, we need to pass in the orginal model.

[ ]:
loaded_model = InferenceOptimizer.load("./optimized_model_ipex", model=model_ft)

Inference with the Loaded Model

[ ]:
with InferenceOptimizer.get_context(loaded_model):
    x = torch.rand(2, 3, 224, 224)
    y_hat = loaded_model(x)
    predictions = y_hat.argmax(dim=1)