puzzles about the network raw output visualization in your cvprw paper

[brisker]

the raw output of the network should be batch_size nDim H W, in your code it seems that nDim is 8, how did you get the figure on the left-bottom corner on the figure above, since it seems like a 3 H * W rgb image?(you refer to that figure to "the raw output of the network")

[davyneven]

Indeed, in order to visualize the raw output of the network, we set the number of dimensions to 2. The raw output is then a 2-dimensional image, or a RG-image, which is the bottom left image. The same representation is on the top right, but then displayed in a xy plot.

[brisker]

@DavyNeven two more questions:

1. what tools did you use to get the top right figure?
2. I use the following code to test your instance loss function, (just adding a line: print(loss_d) )

   
   in_margin = 0.5
   out_margin = 1.5
   Lnorm = 2
   function norm(inp, L)
   local n
   if (L == 1) then
       n = torch.sum(torch.abs(inp), 1)
   else
       n = torch.sqrt(torch.sum(torch.pow(inp, 2), 1) + 1e-8)
   end
   return n
   end

function instance_loss(prediction, labels)

local batchsize = prediction:size(1)
local c = prediction:size(2)
local height = prediction:size(3)
local width = prediction:size(4)
local nInstanceMaps = labels:size(2)
local loss = 0

for b = 1, batchsize do
    local pred = prediction[b] -- c x h x w
    local loss_var = 0
    local loss_dist = 0

    for h = 1, nInstanceMaps do
        local label = labels[b][h]:view(1, height, width) -- 1 x h x w
        local means = {}
        local loss_v = 0
        local loss_d = 0

        -- center pull force
        for j = 1, label:max() do
            local mask = label:eq(j)
            local mask_sum = mask:sum()
            if (mask_sum > 1) then
                local inst = pred[mask:expandAs(pred)]:view(c, -1, 1) -- c x -1 x 1

                -- Calculate mean of instance
                local mean = torch.mean(inst, 2) -- c x 1 x 1
                table.insert(means, mean)

                -- Calculate variance of instance
                local var = norm((inst - mean:expandAs(inst)), 2) -- 1 x -1 x 1
                var = torch.cmax(var - (in_margin), 0)
                local not_hinged = torch.sum(torch.gt(var, 0))

                var = torch.pow(var, 2)
                var = var:view(-1)

                var = torch.mean(var)
                loss_v = loss_v + var
            end
        end

        loss_var = loss_var + loss_v

        -- center push force
        if (#means > 1) then
            for j = 1, #means do
                local mean_A = means[j] -- c x 1 x 1
                for k = j + 1, #means do
                    local mean_B = means[k] -- c x 1 x 1
                    local d = norm(mean_A - mean_B, Lnorm) -- 1 x 1 x 1
                    d = torch.pow(torch.cmax(-(d - 2 * out_margin), 0), 2)
                    loss_d = loss_d + d[1][1][1]
                end
            end

            loss_dist = loss_dist + loss_d / ((#means - 1) + 1e-8)
        end
    end

    loss = loss + (loss_dist + loss_var)
   print(loss_d)
end

loss = loss / batchsize + torch.sum(prediction) * 0

return loss

end

prediction = torch.ones(2,8,50,50) labels=torch.ones(2,1,50,50) labels:sub(-1,-1,-1,-1,29,34,39,49):fill(2) loss=instance_loss(prediction,labels)

But why the loss_d is two **nil values**? Are you sure the instance_loss implementation has no bugs?
![image](https://user-images.githubusercontent.com/13804492/32220329-44ff9170-be6c-11e7-9d43-050b169eaadc.png)

[brisker]

Besides, if batch_size is different, why is the loss different??? @DavyNeven

[davyneven]

loss_d is nil because it is a local variable inside a for loop, you are printing its value outside this loop

[brisker]

@DavyNeven what about the different loss values between different batch_size? I think they should be the same.

[davyneven]

They will be the same if the batch samples are the same. If you want them to be the same, change your sub-method to labels:sub(1,-1,1,-1,29,34,39,49):fill(2).

[brisker]

@DavyNeven
sorry, new to torch