DeepLearningTutorials: Convolutional MLP: How to use a trained network to do a classification

The script is indeed a demonstration piece, but everything you need to do classification on new data is in there. Focus on lines 183 - 193 in the convolutional_mlp.py example.

The logistic regression object layer3 performs the classification:

layer3 = LogisticRegression(input=layer2.output, n_in=500, n_out=10)

and the theano.function test_model (see below) computes the 0-1 loss on a mini-batch of data. If you want to compute some predicted labels for new data, you will need to construct another theano.function object whose output is the predicted labels (i.e layer3.y_pred instead of layer3.errors(y)).

# from the lenet example: create a function to compute the mistakes that are made by the model
test_model = theano.function([index], layer3.errors(y),
         givens={
            x: test_set_x[index * batch_size: (index + 1) * batch_size],
            y: test_set_y[index * batch_size: (index + 1) * batch_size]})

# what you want: create a function to predict labels that are made by the model
model_predict = theano.function([index], layer3.y_pred,
         givens={
            x: my_new_data[index * batch_size: (index + 1) * batch_size]})

You will have to package up your new data, denoted above as my_new_data, in the same format as the theano.shared objects returned by load_data (see logistic_sgd.py).

Hope this helps,