Convert Deep Studying Fashions between PyTorch, TensorFlow, and MATLAB » Synthetic Intelligence

On this weblog put up we’re going to present you find out how to use the latest MATLAB features to:

1. Import fashions from TensorFlow and PyTorch into MATLAB
2. Export fashions from MATLAB to TensorFlow and PyTorch

This can be a temporary weblog put up that factors you to the suitable features and different sources for changing deep studying fashions between MATLAB, PyTorch®, and TensorFlow™. Two good sources to get began with are the documentation subjects Interoperability Between Deep Studying Toolbox, TensorFlow, PyTorch, and ONNX and Tips about Importing Fashions from TensorFlow, PyTorch, and ONNX.

In case you have any questions in regards to the performance introduced on this weblog put up or wish to share the thrilling initiatives for which you might be utilizing mannequin conversion, remark beneath.

Import Fashions into MATLAB

You possibly can import fashions from PyTorch or TensorFlow with only one line of code.

Fast Instance

This instance exhibits you find out how to import a picture classification mannequin from PyTorch. The PyTorch mannequin have to be pretrained and traced.

Run the next code in Python to get and save the PyTorch mannequin. Load the MnasNet pretrained picture classification mannequin from the TorchVision library.

import torch
from torchvision import fashions
mannequin = fashions.mnasnet1_0(pretrained=True)

Hint the PyTorch mannequin. For extra data on find out how to hint a PyTorch mannequin, go to Torch documentation: Tracing a perform. Then, save the PyTorch mannequin.

X = torch.rand(1,3,224,224)
traced_model = torch.jit.hint(mannequin.ahead,X)
traced_model.save("traced_mnasnet1_0.pt")

Now, go to MATLAB and import the mannequin by utilizing the importNetworkFromPyTorch perform. Specify the name-value argument PyTorchInputSizes in order that the import perform robotically creates and provides the enter layer for a batch of photos.

internet = importNetworkFromPyTorch("mnasnet1_0.pt",PyTorchInputSizes=[NaN,3,224,224])

internet = 

  dlnetwork with properties:

         Layers: [153×1 nnet.cnn.layer.Layer]
    Connections: [162×2 table]
     Learnables: [210×3 table]
          State: [104×3 table]
     InputNames: {'InputLayer1'}
    OutputNames: {'aten__linear12'}
    Initialized: 1

  View abstract with abstract.

Learn the picture you wish to classify. Resize the picture to the enter measurement of the community.

Im_og = imread("peacock.jpg");
InputSize = [224 224 3];
Im = imresize(Im_og,InputSize(1:2));

The inputs to MnasNet require additional preprocessing. Rescale the picture. Then, normalize the picture by subtracting the coaching photos imply and dividing by the coaching photos commonplace deviation. For extra data, see Enter Knowledge Preprocessing.

Im = rescale(Im,0,1);

meanIm = [0.485 0.456 0.406];
stdIm = [0.229 0.224 0.225];
Im = (Im - reshape(meanIm,[1 1 3]))./reshape(stdIm,[1 1 3]);

Convert the picture to a dlarray object. Format the picture with the size “SSCB” (spatial, spatial, channel, batch).

Im_dlarray = dlarray(single(Im),"SSCB");

Classify the picture and discover the anticipated label.

prob = predict(internet,Im_dlarray);
[~,label_ind] = max(prob);

Show the picture and classification consequence.

imshow(Im_og)
title(strcat("Classification Outcome: ",ClassNames(label_ind)),FontSize=16)

 

Importing a mannequin from TensorFlow is sort of just like importing a mannequin from PyTorch. After all it’s essential use the importNetworkFromTensorFlow perform as a substitute. Be aware that your TensorFlow mannequin have to be within the SavedModel format and saved by utilizing the next code in Python.

mannequin.save("myModelTF")

For extra examples, try the reference pages of the importNetworkFromPyTorch and importNetworkFromTensorFlow features, and the documentation web page on Pretrained Networks from Exterior Platforms.

 

Why Import Fashions into MATLAB?

In brief, import fashions into MATLAB as a result of an imported community is a MATLAB community. This implies, you may carry out all the next duties with built-in instruments. To be taught extra, see Deep Studying Toolbox.

 

 

Export Fashions from MATLAB

You possibly can export fashions to TensorFlow immediately. To export a mannequin to PyTorch, you need to first convert the mannequin to the ONNX mannequin format.

What?	Export fashions to TensorFlow.	Export fashions to PyTorch.
How?	Use the exportNetworkToTensorFlow perform.	Export through ONNX by utilizing the exportONNXNetwork perform.
When?	Launched in R2022b.	ONNX export launched in R2018a.

 

Fast Instance

Load the pretrained SqueezeNet convolutional community by utilizing imagePretrainedNetwork. This perform, which was launched in R2024a, is the simplest (and beneficial) solution to load pretrained picture classification networks.

[net,ClassNames] = imagePretrainedNetwork("squeezenet");

Export the community internet to TensorFlow. The exportNetworkToTensorFlow perform saves the TensorFlow mannequin within the Python package deal myModel.

exportNetworkToTensorFlow(internet,"myModel")

And you might be completed with exporting!

If you wish to take a look at mannequin in Python, you should use a picture obtainable in MATLAB. Learn the picture you wish to classify. Resize the picture to the enter measurement of the community.

Im = imread("peacock.jpg");
InputSize = internet.Layers(1).InputSize;
Im = imresize(Im,InputSize(1:2));

Permute the 2-D picture knowledge from the MATLAB ordering (HWCN) to the TensorFlow ordering (NHWC), the place H, W, and C are the peak, width, and variety of channels of the picture, respectively, and N is the variety of photos. For extra data, see Enter Dimension Ordering. Save the picture in a MAT file.

ImTF = permute(Im,[4,1,2,3]);
filename = "peacock.mat";
save(filename,"ImTF")

The code beneath exhibits you find out how to use the exported mannequin to foretell in Python or put it aside within the SavedModel format. First, load the exported TensorFlow mannequin from the package deal myModel.

import myModel
mannequin = myModel.load_model()

Save the exported mannequin within the TensorFlow SavedModel format. Saving the mannequin in SavedModel format is optionally available. You possibly can carry out deep studying workflows immediately with the mannequin.

mannequin.save("myModelTF")

Classify the picture with the exported mannequin.

import scipy.io as sio
x = sio.loadmat("peacock.mat")
x = x["ImTF"]

import numpy as np
x_np = np.asarray(x, dtype=np.float32)

scores = mannequin.predict(x_np)

Learn extra on exporting deep neural networks to TensorFlow in this weblog put up. For extra examples, see the exportNetworkToTensorFlow and exportONNXNetwork reference pages and the documentation web page on Exporting Deep Neural Networks.