ImportanceDataSet with BackpropTrainer results in IndexError

[kkleidal]

I have a dataset which I am clustering using a gaussian mixture model, and then I want to train a neural network for each of the clusters. I want to use all the points in my dataset weighted based on the probability they are in the cluster for which the net is being trained.

Originally, I was not weighting the training data and it worked fine:

'''
Create and train a neural net on the training data, given the actual labels
'''
def create_neural_net(training, labels, weights=None, T=10, silent=False):
    input_units = len(training[0])
    output_units = len(labels[0])
    n = len(training)

    net = FeedForwardNetwork()
    layer_in = SoftmaxLayer(input_units)
    layer_hidden = SigmoidLayer(1000)
    layer_hidden2 = SigmoidLayer(50)
    layer_out = LinearLayer(output_units)

    net.addInputModule(layer_in)
    net.addModule(layer_hidden)
    net.addModule(layer_hidden2)
    net.addOutputModule(layer_out)

    net.addConnection(FullConnection(layer_in, layer_hidden))
    net.addConnection(FullConnection(layer_hidden, layer_hidden2))
    net.addConnection(FullConnection(layer_hidden2, layer_out))

    net.sortModules()

    training_data = SupervisedDataSet(input_units, output_units)
    for i in xrange(n):
        # print len(training[i]) # prints 148
        # print len(labels[i]) # prints 13
        training_data.appendLinked(training[i], labels[i])
    trainer = BackpropTrainer(net, training_data)

    for i in xrange(T):
        if not silent: print "Training %d" % (i + 1)
        error = trainer.train()
        if not silent: print net.activate(training[0]), labels[0]
        if not silent: print "Training iteration %d.  Error: %f." % (i + 1, error)
    return net

But now when I try to weight the data points:

'''
Create and train a neural net on the training data, given the actual labels
'''
def create_neural_net(training, labels, weights=None, T=10, silent=False):
    input_units = len(training[0])
    output_units = len(labels[0])
    n = len(training)

    net = FeedForwardNetwork()
    layer_in = SoftmaxLayer(input_units)
    layer_hidden = SigmoidLayer(1000)
    layer_hidden2 = SigmoidLayer(50)
    layer_out = LinearLayer(output_units)

    net.addInputModule(layer_in)
    net.addModule(layer_hidden)
    net.addModule(layer_hidden2)
    net.addOutputModule(layer_out)

    net.addConnection(FullConnection(layer_in, layer_hidden))
    net.addConnection(FullConnection(layer_hidden, layer_hidden2))
    net.addConnection(FullConnection(layer_hidden2, layer_out))

    net.sortModules()

    training_data = ImportanceDataSet(input_units, output_units)
    for i in xrange(n):
        # print len(training[i]) # prints 148
        # print len(labels[i]) # prints 13
        training_data.addSample(training[i], labels[i], importance=(weights[i] if weights is not None else None))
    trainer = BackpropTrainer(net, training_data)

    for i in xrange(T):
        if not silent: print "Training %d" % (i + 1)
        error = trainer.train()
        if not silent: print net.activate(training[0]), labels[0]
        if not silent: print "Training iteration %d.  Error: %f." % (i + 1, error)
    return net

I get the following error:

Traceback (most recent call last):
  File "clustering_experiment.py", line 281, in <module>
    total_model = get_model(training, training_labels, num_clusters=NUM_CLUSTERS
, T=NUM_ITERS_NEURAL_NET)
  File "clustering_experiment.py", line 177, in get_model
    neural_nets.append(neural_net_plugin.create_neural_net(tra.tolist(), val.tol
ist(), T=T, silent=True))
  File "/home/neural_net_plugin.py", line 43, in create_neural_net
    error = trainer.train()
  File "/usr/local/lib/python2.7/dist-packages/PyBrain-0.3.1-py2.7.egg/pybrain/s
upervised/trainers/backprop.py", line 61, in train
    e, p = self._calcDerivs(seq)
  File "/usr/local/lib/python2.7/dist-packages/PyBrain-0.3.1-py2.7.egg/pybrain/s
upervised/trainers/backprop.py", line 92, in _calcDerivs
    outerr = target - self.module.outputbuffer[offset]
IndexError: index 162 is out of bounds for axis 0 with size 1

[kkleidal]

Changed _calcDerivs to

def _calcDerivs(self, seq):
        """Calculate error function and backpropagate output errors to yield
        the gradient."""
        self.module.reset()
        for sample in seq:
            self.module.activate(sample[0])
        error = 0
        ponderation = 0.
        print("seq shape: (%d x %d)" % (len(seq), len(seq[0])))
        print("\n\n")
        for offset, sample in reversed(list(enumerate(seq))):
            # need to make a distinction here between datasets containing
            # importance, and others
            target = sample[1]
            print("target: %s\n outputbuffer: %s\noffset: %s\n\n" % (target, self.module.outputbuffer, offset))
            outerr = target - self.module.outputbuffer[offset]
            if len(sample) > 2:
                importance = sample[2]
                error += 0.5 * dot(importance, outerr ** 2)
                ponderation += sum(importance)
                self.module.backActivate(outerr * importance)
            else:
                error += 0.5 * sum(outerr ** 2)
                ponderation += len(target)
                # FIXME: the next line keeps arac from producing NaNs. I don't
                # know why that is, but somehow the __str__ method of the
                # ndarray class fixes something,
                str(outerr)
                self.module.backActivate(outerr)

        return error, ponderation

and got the following output:

seq shape: (88 x 3)



target: [ 1.  1.  1.  1.  1.  1.  1.  1.  1.  1.  1.  1.  1.]
 outputbuffer: [[-12.05904822  -2.70622648   2.96380057  -0.35463178   3.3011338
8
   -6.42784335  -2.96413836   5.53089292   4.06147799  -1.6830401
    8.43510704  -2.41641803  -4.16226018]]
offset: 87


Traceback (most recent call last):
  File "clustering_experiment.py", line 281, in <module>
    total_model = get_model(training, training_labels, num_clusters=NUM_CLUSTERS
, T=NUM_ITERS_NEURAL_NET)
  File "clustering_experiment.py", line 177, in get_model
    neural_nets.append(neural_net_plugin.create_neural_net(tra.tolist(), val.tol
ist(), T=T, silent=True))
  File "/home/neural_net_plugin.py", line 43, in create_neural_net
    error = trainer.train()
  File "/usr/local/lib/python2.7/dist-packages/PyBrain-0.3.3-py2.7.egg/pybrain/s
upervised/trainers/backprop.py", line 61, in train
    e, p = self._calcDerivs(seq)
  File "/usr/local/lib/python2.7/dist-packages/PyBrain-0.3.3-py2.7.egg/pybrain/s
upervised/trainers/backprop.py", line 94, in _calcDerivs
    outerr = target - self.module.outputbuffer[offset]
IndexError: index 87 is out of bounds for axis 0 with size 1

[kkleidal]

Fixed it. I forgot to call newSequence. It would be nice to have had a more useful error message to make this more obvious (although based on the current datastructure in dataset.py, I imagine that might be hard to pick up on). Here is the final code:

'''
Create and train a neural net on the training data, given the actual labels
'''
def create_neural_net(training, labels, weights=None, T=10, silent=False):
    input_units = len(training[0])
    output_units = len(labels[0])
    n = len(training)

    net = FeedForwardNetwork()
    layer_in = SoftmaxLayer(input_units)
    layer_hidden = SigmoidLayer(1000)
    layer_hidden2 = SigmoidLayer(50)
    layer_out = LinearLayer(output_units)

    net.addInputModule(layer_in)
    net.addModule(layer_hidden)
    net.addModule(layer_hidden2)
    net.addOutputModule(layer_out)

    net.addConnection(FullConnection(layer_in, layer_hidden))
    net.addConnection(FullConnection(layer_hidden, layer_hidden2))
    net.addConnection(FullConnection(layer_hidden2, layer_out))

    net.sortModules()

    training_data = ImportanceDataSet(input_units, output_units) # SupervisedDataSet(input_units, output_units)
    for i in xrange(n):
        # print len(training[i]) # prints 148
        # print len(labels[i]) # prints 13
        training_data.newSequence()
        training_data.addSample(training[i], labels[i], importance=ones(output_units) * (weights[i] if weights is not None else 1)) # training_data.appendLinked(training[i], labels[i])
    trainer = BackpropTrainer(net, training_data)

    for i in xrange(T):
        if not silent: print "Training %d" % (i + 1)
        error = trainer.train()
        if not silent: print net.activate(training[0]), labels[0]
        if not silent: print "Training iteration %d.  Error: %f." % (i + 1, error)
    return net