{"id":10794,"date":"2022-12-12T10:11:01","date_gmt":"2022-12-12T15:11:01","guid":{"rendered":"https:\/\/blogs.mathworks.com\/deep-learning\/?p=10794"},"modified":"2024-02-12T20:37:29","modified_gmt":"2024-02-13T01:37:29","slug":"quickly-investigate-pytorch-models-from-matlab","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/deep-learning\/2022\/12\/12\/quickly-investigate-pytorch-models-from-matlab\/","title":{"rendered":"Quickly Investigate PyTorch Models from MATLAB"},"content":{"rendered":"The following post is from Sivylla Paraskevopoulou<\/a>, Product Marketing Manager at MathWorks, and Yann Debray<\/a>, Product Manager at MathWorks.<\/em>\r\n
<\/h6>\r\nThis blog post talks about how MATLAB, PyTorch\u00ae, and TensorFlow\u2122 can be used together.\r\n
<\/h6>\r\nDeep learning models commonly exist within a complete AI system, which can involve preparing the data, building the model, designing the system on which the model will run, and deploying to hardware or production. MATLAB provides tools to help you at each step of the AI system design.\r\n
<\/h6>\r\nYou can get a pretrained deep learning model from the MATLAB Deep Learning Model Hub<\/a>, or from TensorFlow, PyTorch, or ONNX\u2122 repositories. We will show here you how to quickly compare PyTorch image classification models without leaving the MATLAB environment.\r\n
<\/h6>\r\n \r\n
<\/h6>\r\n\"\"\r\n
<\/h6>\r\nFigure:<\/strong> Adding an extra step to my existing workflow in order to find the right PyTorch model\r\n
<\/h6>\r\n \r\n
<\/h6>\r\nOur image classification workflow includes loading and preprocessing an image, importing an image classification model from PyTorch, and using the imported network to predict the image label. This workflow is presented in the documentation example Import Network from PyTorch and Classify Image<\/a>.\r\n
<\/h6>\r\nLet\u2019s assume the most important selection factor for the classifier model is the model\u2019s prediction speed. The torchvision.models<\/a> library alone has 12 models to choose from. It would be quite cumbersome to import each of these 12 models for our comparison test. We will show you here how you can call PyTorch from MATLAB to run an inference speed test quickly on multiple PyTorch models.\r\n
<\/h6>\r\nIn this blog post we won\u2019t show you all the details for each step. You can find the detailed example at\u00a0Call Python from MATLAB to Compare PyTorch Models for Image Classification<\/a>. Instead, here we are focusing on key takeaways on exploring PyTorch models with co-execution from MATLAB.\r\n
<\/h6>\r\n \r\n

Preprocess Image<\/strong><\/p>\r\nRead the image you want to classify. Resize the image to the input size of the network.\r\n

imgOriginal = imread(\"banana.png\");\r\nInputSize = [224 224 3];\r\nimg = imresize(imgOriginal,InputSize(1:2));\r\n<\/pre>\r\nYou must preprocess the image in the same way as the training data. For more information, see Input Data Preprocessing<\/a>.\r\n
<\/h6>\r\nRescale the image. Then, normalize the image by subtracting the training images mean and dividing by the training images standard deviation.\r\n
imgProcessed = rescale(img,0,1);\r\n\r\nmeanIm = [0.485 0.456 0.406];\r\nstdIm = [0.229 0.224 0.225];\r\nimgProcessed = (imgProcessed - reshape(meanIm,[1 1 3])).\/reshape(stdIm,[1 1 3]);\r\n<\/pre>\r\n
<\/h6>\r\nPermute the image data from the Deep Learning Toolbox dimension ordering (HWCN<\/em>) to the PyTorch dimension ordering (NCHW<\/em>), where H<\/em> is the height of the images, W<\/em> is the width of the images, C<\/em> is the number of channels, and N<\/em> is the number of observations. This is a necessary step to use the image for prediction with a PyTorch model (before importing it into MATLAB).\r\n
<\/h6>\r\n
imgForTorch = permute(img,[4 3 1 2]);\r\n<\/pre>\r\n
<\/h6>\r\nFor more information on input dimension data ordering for different deep learning platforms, see Input Dimension Ordering<\/a>.\r\n
<\/h6>\r\n\r\n\r\n\r\n
Takeaways:<\/strong>\r\n
    \r\n \t
  1. Data preprocessing for prediction must match the training data preprocessing.<\/li>\r\n \t
  2. Permute the data from MATLAB to PyTorch dimension ordering to use the input data with a PyTorch model.<\/li>\r\n<\/ol>\r\n<\/td>\r\n<\/tr>\r\n<\/tbody>\r\n<\/table>\r\n
    <\/h6>\r\n \r\n
    Install Python and Libraries<\/strong><\/p>\r\nYou might have multiple versions of Python installed on your desktop. For example, a MacBook has a pre-installed Python version 2.7, which is likely not the version you want to use. So, it is good practice to create a virtual environment for your project to be in control of the Python version and libraries that you are using.\r\n
    <\/h6>\r\nThe following commands show how you can setup a virtual environment on a MacBook. If you are using a Windows machine, the commands might be slightly different.\r\n
    <\/h6>\r\nGo to your working folder. Create and activate the Python virtual environment env<\/span> in your working folder.\r\n
    <\/h6>\r\n
    python3.10 -m venv env\r\nsource env\/bin\/activate\r\n<\/pre>\r\n
    <\/h6>\r\nInstall the necessary Python libraries for this example. Check the installed versions of the libraries.\r\n
    <\/h6>\r\n
    pip3 install numpy torch torchvision\r\npython3 -m pip show numpy torch torchvision\r\n<\/pre>\r\n
    <\/h6>\r\nFor reference, we used:\r\n