Hi Vincenzo, could you please help me find the NORB dataset in Theano/Lasagne format?
Thank you very much, Best, Christos On Thursday, 11 June 2015 00:38:04 UTC+3, Vincenzo Lomonaco wrote: > > Hello everyone, > > I am trying to reproduce with Theano the results obtained on the small > NORB dataset and reported in the paper "Learning Methods for Generic Object > Recognition with Invariance to Pose and Lighting" [ Huang, LeCun - > http://yann.lecun.com/exdb/publis/pdf/lecun-04.pdf ] using CNNs. > > Starting from the LeNet tutorial [ > http://deeplearning.net/tutorial/lenet.html ] I changed the model to fit > what described in the paper but I can't get the error rate below *8,7%* > while in the paper is reported as *6,8%*. > > Has anyone tried this before? > > Here the model details: > > nkerns=[8, 24] > > layer0_input = x.reshape((batch_size, 2, 96, 96)) > > layer0 = LeNetConvPoolLayer( > rng, > input=layer0_input, > image_shape=(batch_size, in_dim, img_dim, img_dim), > filter_shape=(nkerns[0], in_dim, 5, 5), > poolsize=(4,4), > pool_type='max' > ) > > layer1 = LeNetConvPoolLayer( > rng, > input=layer0.output, > image_shape=(batch_size, nkerns[0], 23, 23), > filter_shape=(nkerns[1], nkerns[0], 6, 6), > poolsize=(3,3), > pool_type='max' > ) > > layer2_input = layer1.output.flatten(2) > > layer2 = HiddenLayer( > rng, > input=layer2_input, > n_in=nkerns[1] * 6 * 6, > n_out=batch_size, > activation=T.tanh > ) > > layer3 = LogisticRegression(input=layer2.output, n_in=batch_size, > n_out=5) > > > > I've also tried sum and average pooling other than max, and implemented > dropout for the hidden layer but without great improvements. > > Do you think that the problem is the full-connected convolution operation? > Does anyone has an example code to select input feature maps in a > convolutional layer? > Any suggestion? > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
