Inference Comparison Between TensorFlow and Imported Networks for Image Classification

This example shows how to compare the inference (prediction) results of a TensorFlow™ network and the imported network in MATLAB® for an image classification task. First, use the network for prediction in TensorFlow and save the prediction results. Then, import the network in MATLAB using the importTensorFlowNetwork function and predict the classification outputs for the same images used to predict in TensorFlow.

This example provides the supporting files digitsNet.zip and TFData.mat. To access these supporting files, open the example in Live Editor.

Image Data Set

Load the Digits data set. The data contains images of digits and the corresponding labels.

[XTest,YTest] = digitTest4DArrayData;

Create the test data that the TensorFlow network uses for prediction. Permute the 2-D image data from the Deep Learning Toolbox™ ordering (HWCN) to the TensorFlow ordering (NHWC), where H, W, and C are the height, width, and number of channels of the images, respectively, and N is the number of images.

x_test = permute(XTest,[4,1,2,3]);
y_test = double(string(YTest));

Save the data to a MAT file.

filename = "digitsMAT.mat";
save(filename,"x_test","y_test")

Inference with Pretrained Network in TensorFlow

Load a pretrained TensorFlow network for image classification in Python® and classify new images.

Import libraries.

import tensorflow as tf
import scipy.io as sio

Load the test data set from digitsMAT.mat.

data = sio.loadmat("digitsMAT.mat")
x_test = data["x_test"]
y_test = data["y_test"]

Load the digitsNet pretrained TensorFlow model, which is in the saved model format. If the folder is archived in digitsNet.zip, extract the archived contents of digitsNet.zip into the current folder.

from tensorflow import keras
model = keras.models.load_model("digitsNet")

Display a summary of the model.

model.summary()

Classify new digit images.

scores = model.predict(tf.expand_dims(x_test,-1))

Save the classification scores in the MAT file TFData.mat.

sio.savemat("TFData.mat", 
            {"scores_tf":scores})

Inference with Imported Network in MATLAB

Import the pretrained TensorFlow network into MATLAB using importTensorFlowNetwork and classify the same images as in TensorFlow.

Specify the model folder, which contains the digitsNet TensorFlow model in the saved model format.

if ~exist("digitsNet","dir")
    unzip("digitsNet.zip")
end
modelFolder = "./digitsNet";

Specify the class names.

classNames = string(0:9);

Import the TensorFlow network in the saved model format. By default, importTensorFlowNetwork imports the network as a DAGNetwork object.

net = importTensorFlowNetwork(modelFolder,Classes=classNames);

Importing the saved model...
Translating the model, this may take a few minutes...
Finished translation. Assembling network...
Import finished.

Display the network layers.

net.Layers

ans = 
  13x1 Layer array with layers:

     1   'conv2d_input'                  Image Input             28x28x1 images
     2   'conv2d'                        2-D Convolution         8 3x3x1 convolutions with stride [1  1] and padding [0  0  0  0]
     3   'conv2d_relu'                   ReLU                    ReLU
     4   'max_pooling2d'                 2-D Max Pooling         2x2 max pooling with stride [2  2] and padding [0  0  0  0]
     5   'conv2d_1'                      2-D Convolution         16 3x3x8 convolutions with stride [1  1] and padding [0  0  0  0]
     6   'conv2d_1_relu'                 ReLU                    ReLU
     7   'max_pooling2d_1'               2-D Max Pooling         2x2 max pooling with stride [2  2] and padding [0  0  0  0]
     8   'flatten'                       Flatten                 Flatten activations into 1-D assuming C-style (row-major) order
     9   'dense'                         Fully Connected         100 fully connected layer
    10   'dense_relu'                    ReLU                    ReLU
    11   'dense_1'                       Fully Connected         10 fully connected layer
    12   'dense_1_softmax'               Softmax                 softmax
    13   'ClassificationLayer_dense_1'   Classification Output   crossentropyex with '0' and 9 other classes

Predict class labels and classification scores using the imported network.

[labels_dlt,scores_dlt] = classify(net,XTest);

For this example, the data XTest is in the correct ordering. Note that if the image data XTest is in TensorFlow dimension ordering, you must convert XTest to the Deep Learning Toolbox ordering by entering Xtest = permute(Xtest,[2 3 4 1]).

Compare Accuracy

Load the TensorFlow network scores from TFData.mat.

load("TFData.mat")

Compare the inference results (classification scores) of the TensorFlow network and the imported network.

diff = max(abs(scores_dlt-scores_tf),[],"all")

diff = single
    5.1260e-06

The difference between inference results is negligible, which strongly indicates that the TensorFlow network and the imported network are the same.

As a secondary check, you can compare the classification labels. First, compute the class labels predicted by the TensorFlow network. Then, compare the labels predicted by the TensorFlow network and the imported network.

[~,ind] = max(scores_tf,[],2);
labels_tf = categorical(classNames(ind))';
isequal(labels_dlt,labels_tf)

ans = logical
   1

The labels are the same, which indicates that the two networks are the same.

Plot confusion matrix charts for the labels predicted by the TensorFlow network and the imported network.

tiledlayout(2,1)
nexttile
confusionchart(YTest,labels_tf)
title("TensorFlow Predictions")
nexttile
confusionchart(YTest,labels_dlt)
title("Deep Learning Toolbox Predictions")