caffe version of the TFeat Liberty

[ducha-aiki]

Weights are manually moved and normalization is hardcoded in .prototxt. It expects as input [0 255] non-normalized grayscale 32x32 patches

[edgarriba]

awesome!

[edgarriba]

could you provide any python or cpp script for running it?

[ducha-aiki]

@edgarriba done

[edgarriba]

@ducha-aiki great!

[edgarriba]

@ducha-aiki I'd really really love to see how did you convert from Pytorch/Tensorflow? to Caffe

[ducha-aiki]

@edgarriba very simple - manually. From lua (lutorpy) :

i = 0
for k in net.net.modules:
    print str(net.net.modules[k])
    if 'cudnn.SpatialConvolution' in str(net.net.modules[k]):
        print str(net.net.modules[k])
        if i == 1:
            w1 = net.net.modules[k].weight.asNumpyArray()
            print w1.shape
            caffenet.params['conv2'][0].data[:] = w1;
            caffenet.params['conv2'][1].data[:] =  net.net.modules[k].bias.asNumpyArray()
            i+=1
        if i == 0:
            w1 = net.net.modules[k].weight.asNumpyArray()
            print w1.shape
            caffenet.params['conv1'][0].data[:] = w1;
            caffenet.params['conv1'][1].data[:] = net.net.modules[k].bias.asNumpyArray()
            i+=1
    if 'Linear' in str(net.net.modules[k]):
        print str(net.net.modules[k])
        w1 = net.net.modules[k].weight.asNumpyArray()
        caffenet.params['ip1'][0].data[:] = w1;
        caffenet.params['ip1'][1].data[:] = net.net.modules[k].bias.asNumpyArray()

For pytorch (and other arch)

gg = {}
gg['counter'] = 1
def copy_weights(m,):
    if isinstance(m, nn.Conv2d):
        counter = gg['counter']
        l_name = 'conv' + str(counter)
        print l_name,  m.weight.data.cpu().numpy().shape;
        net.params[l_name][0].data[:] = np.rollaxis( m.weight.data.cpu().numpy(),3,2);
        net.params[l_name][1].data[:] = m.bias.data.cpu().numpy();
    if isinstance(m, nn.BatchNorm2d):
        counter = gg['counter']
        l_name = 'conv' + str(counter) + '_BN'
        print l_name
        net.params[l_name][0].data[:] = m.running_mean.cpu().numpy();
        net.params[l_name][1].data[:] = m.running_var.cpu().numpy();
        net.params[l_name][2].data[:] = 1.
        gg['counter'] += 1
def copy_last(m):
    if isinstance(m, nn.Conv2d):
        l_name = 'conv7'
        net.params[l_name][0].data[:] =np.rollaxis( m.weight.data.cpu().numpy(),3,2);
        net.params[l_name][1].data[:] = m.bias.data.cpu().numpy();
        print l_name
    if isinstance(m, nn.BatchNorm2d):
        counter = gg['counter']
        l_name = 'conv7_BN'
        print l_name
        net.params[l_name][0].data[:] = m.running_mean.cpu().numpy();
        net.params[l_name][1].data[:] = m.running_var.cpu().numpy();
        net.params[l_name][2].data[:] = 1.
        gg['counter'] += 1

model.features.apply(copy_weights)
model.classifier.apply(copy_last)

[ducha-aiki]

@edgarriba could you please help me with https://github.com/pytorch/pytorch/issues/1661 ?

[edgarriba]

I think you need register_forward_hook http://pytorch.org/docs/nn.html#torch.nn.Module.register_forward_hook

[edgarriba]

BTW, thanks for the script. I'm not very fluent with low level caffe features

[ducha-aiki]

@edgarriba thanks, I will try it.

[ducha-aiki]

@edgarriba here is full version of script https://github.com/DagnyT/hardnet/blob/master/examples/caffe/convert_weights_to_caffe.py It is different architecture, though

[edgarriba]

@ducha-aiki thanks dude! very interesting paper and results.