XingangPan
the source of weights in vgg.t7
@XingangPan Could you tell me what's the source of weights in vgg.t7? The net of part (44) as follows is not the same as the basic vgg16, so I wonder if these weights are obtained from pretraining or from some tricky initialization?
(44): nn.ConcatTable {
input
|`-> (1): nn.SpatialUpSamplingBilinear(8)
`-> (2): nn.Sequential {
[input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> output]
(1): cudnn.SpatialSoftMax
(2): nn.SpatialAveragePooling(2x2, 2,2)
(3): nn.View(4500)
(4): nn.Linear(4500 -> 128)
(5): nn.ReLU
(6): nn.Linear(128 -> 4)
(7): nn.Sigmoid
}
... -> output
}
@jcdubron , VGG-16 is loaded from the LargeFOV model (released by deeplab) according to the description of their paper.
@cardwing According to the paper, “The initial weights of the first 13 convolution layers are copied from VGG16 (Simonyan and Zisserman 2015) trained on ImageNet (Deng et al. 2009).” But I'm not sure the weights of layers after conv5 as they are not the same in detail.
@jcdubron The weights of layers after conv5 are random initialized as they are different from original vgg.
