Float64 dtype variable is created by nolearn

I have this warning when trying to run a simple network:

lib/python2.7/site-packages/nolearn-0.6a0.dev0-py2.7.egg/nolearn/lasagne/base.py:472: UserWarning: You are creating a TensorVariable with float64 dtype. You requested an action via the Theano flag warn_float64={ignore,warn,raise,pdb}. accuracy = T.mean(T.eq(predict, y_batch))

The input is numpy.float32, the target variable is uint8 and an identical network created only in Lasagne does not raise this error.

This is the nolearn network definition:

    def perform_nn_nolearn(data_train, label_train, data_test, label_test):
    def regularization_objective(layers,*args, **kwargs):

        target = kwargs['target'][:,np.newaxis]
        kwargs['target'] = target
        losses = nolearn.lasagne.objective(layers,*args, **kwargs)
        return losses

    class Custom_TrainSplit(object):
        def __init__(self,test_data, test_labels):
            self.test_data = test_data
            self.test_labels = test_labels

        def __call__(self, X, y, net):
            X_train, y_train = X, y
            X_valid, y_valid = self.test_data, self.test_labels
            return X_train, X_valid, y_train, y_valid

    no_feats=data_train.shape[1]
    batch_size = 32


    nn_layers = [
        (lasagne.layers.InputLayer, {'shape': (batch_size, no_feats)}),
        (lasagne.layers.DenseLayer,{'num_units': 1, 'nonlinearity':lasagne.nonlinearities.sigmoid})
        ]

    net = nolearn.lasagne.NeuralNet(
        layers=nn_layers,
        max_epochs=1,
        objective_loss_function = lasagne.objectives.binary_crossentropy,
        update=lasagne.updates.nesterov_momentum,
        update_learning_rate=0.01,
        update_momentum=0.9,
        objective=regularization_objective,
        regression=False,
        use_label_encoder = False,
        batch_iterator_train=nolearn.lasagne.BatchIterator(batch_size=32),
        train_split=Custom_TrainSplit(data_test,label_test),
        verbose=2
    )

    rez = net.fit(data_train, label_train)
    nn_pred_proba = net.predict_proba(data_test)
    print "Network output shape: {}".format(nn_pred_proba.shape)
    return nn_pred_proba

The warning is raised when starting fit() but before the network statistics. Also, there is a long list of warnings, from theano/compile to lasagne/objectives. I isolated the one regarding nolearn because, again, an identical network+data runs without warnings in “pure” lasagne.

[BenjaminBossan]

Multiplying or dividing float32 and int32 results in a float64 in theano, which is why you see these results. Still, as long as it does not cause any trouble, I would just ignore it.

[cristi-zz]

Ok, thanks!