plots.lua

require 'gnuplot'
require 'optim'
require 'nn'
require 'image'

function displayLossHist(criticalPts, model, criterion, dataset, grid, eps)
   local nIdx = 2*criticalPts[1]['eigenvals']:size(1)+1
   local indices = torch.Tensor(#criticalPts)
   local loss = torch.Tensor(#criticalPts)

   local param,gradParam=model:getParameters()
   local D = param:size(1)
   local input = dataset[1]
   local target = dataset[2]

   for i = 1,#criticalPts do
      local e = criticalPts[i]['eigenvals']
      local idx = 0
      for j = 1,e:size(1) do
	 if e[j] > eps then
	    idx = idx + 1
	 else
	    if e[j] < -eps then
	       idx = idx - 1
	    end
	 end
      end
      indices[i] = idx
      if criticalPts[i]['x'] ~= nil then
	 param:copy(criticalPts[i]['x'])
      else
	 local indx = criticalPts[i]['indx']
	 for j=1,D do
	    param[j] = grid[ indx[j] ]
	 end
      end
      loss[i] = criterion:forward(model:forward(input), target)
   end
   gnuplot.figure(1)
   gnuplot.hist(indices, nIdx)
   gnuplot.figure(2)
   gnuplot.hist(loss, 100)
   local ifig = 3

   for iIdx = -criticalPts[1]['eigenvals']:size(1),criticalPts[1]['eigenvals']:size(1) do
      lossIdx = loss[indices:eq(iIdx)]
      if lossIdx:nDimension() > 0 then
	 gnuplot.figure(ifig)
	 ifig = ifig + 1
	 gnuplot.title('Index = ' .. iIdx)
	 gnuplot.hist(lossIdx,100)
      end
   end
end

-- change name of this file
data = torch.load('criticalPtsHessian_2hidden_6.th')

criticalPts = data['criticalPts']
grid = data['grid']
model = data['model']
criterion = data['criterion']
dataset = data['dataset']
eps = 1e-2
displayLossHist(criticalPts, model, criterion, dataset, grid, eps)