the Upsample op

[catsglasses]

Very great work, just find you code support the 0.3 version, and your implementation are very elegent, will you prepare to implement the Upsample operator?

Thanks a lot

[Yolan1013]

Hi, have you known how to implement the Upsample operator? I met some problems when converting pytorch model containing the upsample layer. Thanks a lot

[hahnyuan]

Hi, The upsampling operation is difficult to implement here. I am try to figure it out. I have wrote a transfer function at pytorch_to_caffe.py, but the output of the transferred layer in Caffe cannot reproduce the result. Could you help me to find the problem?

def _interpolate(raw,input, size=None, scale_factor=None, mode='nearest', align_corners=None):
    if mode=='bilinear':
        x=raw(input, size, scale_factor, mode, align_corners)
    else:
        raise NotImplementedError()
    name=log.add_layer(name='interpolate')
    log.add_blobs([x],name='interpolate_blob')
    layer=caffe_net.Layer_param(name=name, type='Deconvolution',
                                bottom=[log.blobs(input)], top=[log.blobs(x)])
    def bilinear_weight(shape):
        weight = np.zeros(np.prod(shape), dtype='float32')
        f = np.ceil(shape[3] / 2.)
        c = (2 * f - 1 - f % 2) / (2. * f)
        for i in range(np.prod(shape)):
            x = i % shape[3]
            y = (i / shape[3]) % shape[2]
            weight[i] = (1 - abs(x / f - c)) * (1 - abs(y / f - c))
        return weight.reshape(shape)
    kernel_size=2*scale_factor-scale_factor%2
    stride=scale_factor
    pad=int(np.ceil((scale_factor-1)/2))
    channels=x.size(1)
    weight=bilinear_weight([channels,1,kernel_size,kernel_size])
    layer.conv_param(channels,kernel_size,stride=stride,pad=pad,bias_term=False,groups=channels)
    layer.add_data(weight)
    log.cnet.add_layer(layer)
    return x

[deadmarston]

Hi, Have you ever found any solution for converting the upsample operation? I have also met this problem.