Shape mismatch using the CPU at the ImageNet Pretrained Network (VGG_S).ipynb about recipes

[ebenolson]

Thanks for the report. I have a feeling this was caused by recent changes to the Pool2DLayer API. Unfortunately I won't be able to investigate further for a few days, but if you have a chance try adding ignore_border=False to the arguments for each PoolLayer call.

[pfeodrippe]

It worked, man \o/, thanks! Just adding ignore_border=False to the pool layers as you said just made the code work perfectly

And OMG, what a awesome library, its concept is beautiful... just beautiful =)

net = {}
net['input'] = InputLayer((None, 3, 224, 224))
net['conv1'] = ConvLayer(net['input'], num_filters=96, filter_size=7, stride=2)
net['norm1'] = NormLayer(net['conv1'], alpha=0.0001) # caffe has alpha = alpha * pool_size
net['pool1'] = PoolLayer(net['norm1'], pool_size=3, stride=3, ignore_border=False)
net['conv2'] = ConvLayer(net['pool1'], num_filters=256, filter_size=5)
net['pool2'] = PoolLayer(net['conv2'], pool_size=2, stride=2, ignore_border=False)
net['conv3'] = ConvLayer(net['pool2'], num_filters=512, filter_size=3, pad=1)
net['conv4'] = ConvLayer(net['conv3'], num_filters=512, filter_size=3, pad=1)
net['conv5'] = ConvLayer(net['conv4'], num_filters=512, filter_size=3, pad=1)
net['pool5'] = PoolLayer(net['conv5'], pool_size=3, stride=3, ignore_border=False)
net['fc6'] = DenseLayer(net['pool5'], num_units=4096)
net['drop6'] = DropoutLayer(net['fc6'], p=0.5)
net['fc7'] = DenseLayer(net['drop6'], num_units=4096)
net['drop7'] = DropoutLayer(net['fc7'], p=0.5)
net['fc8'] = DenseLayer(net['drop7'], num_units=1000, nonlinearity=lasagne.nonlinearities.softmax)
output_layer = net['fc8']

[f0k]

I have a feeling this was caused by recent changes to the Pool2DLayer API.

Right... we should update it!

Just adding ignore_border=False to the pool layers as you said just made the code work perfectly

Note that you will still get wrong predictions when replacing a Conv2DDNNLayer with a ConvLayer! The Conv2DDNNLayer has a flip_filters keyword argument which defaults to False: http://lasagne.readthedocs.org/en/latest/modules/layers/dnn.html#lasagne.layers.dnn.Conv2DDNNLayer.
So by default, it will compute a correlation, not a convolution. The Conv2DLayer does not have this keyword argument, it will always compute a convolution. To fix this, you will need to explicitly flip the convolution filters after loading everything:

for k, layer in net:
    if k.startswith('conv'):
        layer.W.set_value(np.asarray(layer.W.get_value())[:,:,::-1,::-1])

[f0k]

Right... we should update it!

While we're at it, the following should change as well:

# We need a recent theano version for this to work
assert theano.__version__ >= '0.7.0.dev-512c2c16ac1c7b91d2db3849d8e7f384b524d23b'

class AveragePool2DLayer(lasagne.layers.MaxPool2DLayer):

For one, the version check doesn't work this way, because SHA1 sums are not following any particular order. The next commit's version string may be (lexicographically) smaller.

And secondly, the subclass isn't needed any more since you added the Pool2DLayer.

[ebenolson]

For one, the version check doesn't work this way, because SHA1 sums are not following any particular order. The next commit's version string may be (lexicographically) smaller.

Yeah, I realized this during the Theano version debate but hadn't got around to fixing it yet. As far as I know, there's no way to actually version test Theano reliably, right?

[f0k]

As far as I know, there's no way to actually version test Theano reliably, right?

Correct. We just have to rely on people following Lasagne's installation instructions. I guess we shouldn't do additional version checks in Recipes then, unless a particular Recipe requires a more recent version of Theano than Lasagne. (In that case, if possible, I'd suggest to check for the presence of the required feature, rather than the version string.)