Caffe shape mismatch error using pretrained VGG-16

Question

I am using PyCaffe to implement a neural network inspired by the VGG 16 layer network. I want to use the pre-trained model available from their GitHub page. Generally this works by matching layer names.

For my "fc6" layer I have the following definition in my train.prototxt file:

layer {
  name: "fc6"
  type: "InnerProduct"
  bottom: "pool5"
  top: "fc6"
  inner_product_param {
    num_output: 4096
  }
}

Here is the prototxt file for the VGG-16 deploy architecture. Note that the "fc6"in their prototxt is identical to mine (except for the learning rate, but that's irrelevant). It's also worth noting that the inputs are all the same size in my model too: 3-channel 224x224px images.

I have been following this tutorial pretty closely, and the block of code that's giving me an issue is the following:

solver = caffe.SGDSolver(osp.join(model_root, 'solver.prototxt'))
solver.net.copy_from(model_root + 'VGG_ILSVRC_16_layers.caffemodel')
solver.test_nets[0].share_with(solver.net)
solver.step(1)

The first line loads my solver prototxt and then the second line copies the weights from the pre-trained model (VGG_ILSVRC_16_layers.caffemodel). When the solver runs, I get this error:

Cannot copy param 0 weights from layer 'fc6'; shape mismatch.  Source param 
shape is 1 1 4096 25088 (102760448); target param shape is 4096 32768 (134217728). 
To learn this layer's parameters from scratch rather than copying from a saved 
net, rename the layer.

The gist of it is that their model expects the layer to be of size 1x1x4096 while mine is just 4096. But I don't get how I can change this?

I found this answer in the Users Google group instructing me to do net surgery to reshape the pre-trained model before copying, but in order to do that I need the lmdb files from the original architecture's data layers, which I don't have (it throws an error when I try to run the net surgery script).

Answer 1

This error comes up if you're cropping the images to 224, rather than the 227 that was done with the original dataset. Adjust that and you should be good to go.

Answer 2

The problem is not with 4096 but rather with 25088. You need to calculate the output feature maps for each layer of your network based on the input feature maps. Note that the fc layer takes an input of fixed size so the output of the previous conv layer must match the input size required by the fc layer. Calculate your fc6 input feature map size (this is the output feature map of the previous conv layer) using the input feature map size of the previous conv layer. Here's the formula:

H_out = ( H_in + 2 x Padding_Height - Kernel_Height ) / Stride_Height + 1
W_out = (W_in + 2 x Padding_Width - Kernel_Width) / Stride_Width + 1