How to train multiple datasets/faces_training-*.pkl files with train_test_mlp.py script ?

[G10DRAS]

Hello Michael,

Is it possible to train MLP (params/mlp.xml) using multiple datasets/faces_training-*.pkl files?
Basically I want to do continuous Training and Testing in multiple sessions.

Thanks

[Michael Beyeler]

Hi,

You should be able to simply run homebrew.load_data() on different pickle files from different sessions and feed the resulting data one-by-one to MLP.fit():

for idx in xrange(10):

   # load some file from session, put all samples in X_train

   (X_train, y_train) = homebrew.load_data("datasets/faces-training-" + `idx` + ".pkl", test_split=0)

   # convert to numpy

   X_train = np.squeeze(np.array(X_train)).astype(np.float32)

   y_train = np.array(y_train)

   # train for 10 epochs

   params = dict(term_crit=(cv2.TERM_CRITERIA_COUNT, 10, 0.01),

                 train_method=cv2.ANN_MLP_TRAIN_PARAMS_BACKPROP,

                 bp_dw_scale=0.001, bp_moment_scale=0.9)

   MLP.fit(X_train, y_train, params=params)

Does that answer your question?

[G10DRAS]

Thank for code Michael.

I quickly did other way. That is, in Main GUI, I have loaded previously trained sample and labels as follows:

            print "Loading data from previously trained data_file", training_file
            f = open(training_file, 'rb')
            self.samples = pickle.load(f)
            self.labels = pickle.load(f)
            print "Loaded", len(self.samples), " samples from previously trained data_file"
            f.close()

and then continue with adding more training data to samples and labels. Hence on exit it will save all the data including data loaded from previous training file.

[Paskl SUNNY]

 Hi Michael,

     Thank you very much for your code. Actually i have a problem
     on how i can get the datasets/faces-training.pkl file. on your
     page(https://github.com/mbeyeler/opencv-python-blueprints/tree/master/chapter7/datasets) I can just find in datasets directory the following files: homerbew.py
     and __init__.py which is empty and no faces-training.pkl or faces_preprocessed.pkl

Thank you in advance for your help

[Michael Beyeler]

Hi,

In short, you need to assemble a training set first using the app. The .pkl file will then be created from the face images you record. This is all explained in my book in Chapter 7 and in the source code of "chapter7/chapter7.py" (see the first step below):

	"""OpenCV with Python Blueprints Chapter 7: Learning to Recognize Emotion in Faces
	An app that combines both face detection and face recognition, with a
	focus on recognizing emotional expressions in the detected faces.
	The process flow is as follows:
	* Run the GUI in Training Mode to assemble a training set. Upon exiting
	the app will dump all assembled training samples to a pickle file
	"datasets/faces_training.pkl".
	* Run the script train_test_mlp.py to train a MLP classifier on the
	dataset. This file will store the parameters of the trained MLP in
	a file "params/mlp.xml" and dump the preprocessed dataset to a
	pickle file "datasets/faces_preprocessed.pkl".
	* Run the GUI in Testing Mode to apply the pre-trained MLP classifier
	to the live stream of the webcam.
	"""

Best,

Michael