#!/usr/bin/env luajit
+--[[
+
+ Copyright (c) 2016 Idiap Research Institute, http://www.idiap.ch/
+ Written by Francois Fleuret <francois.fleuret@idiap.ch>
+
+ This file is free software: you can redistribute it and/or modify
+ it under the terms of the GNU General Public License version 3 as
+ published by the Free Software Foundation.
+
+ It is distributed in the hope that it will be useful, but WITHOUT
+ ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+ or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
+ License for more details.
+
+ You should have received a copy of the GNU General Public License
+ along with this file. If not, see <http://www.gnu.org/licenses/>.
+
+]]--
+
require 'torch'
require 'nn'
+-- require 'cunn'
+
require 'dagnn'
-a = nn.Linear(10, 10)
+torch.setdefaulttensortype('torch.DoubleTensor')
+torch.manualSeed(1)
+
+function checkGrad(model, criterion, input, target, epsilon)
+ local params, gradParams = model:getParameters()
+
+ local epsilon = epsilon or 1e-5
+
+ local output = model:forward(input)
+ local loss = criterion:forward(output, target)
+ local gradOutput = criterion:backward(output, target)
+ gradParams:zero()
+ model:backward(input, gradOutput)
+ local analyticalGradParam = gradParams:clone()
+
+ local err = 0
+
+ for i = 1, params:size(1) do
+ local x = params[i]
+
+ params[i] = x - epsilon
+ local output0 = model:forward(input)
+ local loss0 = criterion:forward(output0, target)
+
+ params[i] = x + epsilon
+ local output1 = model:forward(input)
+ local loss1 = criterion:forward(output1, target)
+
+ params[i] = x
+
+ local ana = analyticalGradParam[i]
+ local num = (loss1 - loss0) / (2 * epsilon)
+
+ if num ~= ana then
+ err = math.max(err, math.abs(num - ana) / math.max(epsilon, math.abs(num)))
+ end
+ end
+
+ return err
+end
+
+function printTensorTable(t)
+ if torch.type(t) == 'table' then
+ for i, t in pairs(t) do
+ print('-- ELEMENT [' .. i .. '] --')
+ printTensorTable(t)
+ end
+ else
+ print(tostring(t))
+ end
+end
+
+-- +-- Linear(10, 10) --> ReLU --> d -->
+-- / /
+-- / /
+-- --> a --> b -----------> c ---------------+
+-- \
+-- \
+-- +--------------- e -->
+
+dag = nn.DAG()
+
+a = nn.Linear(50, 10)
b = nn.ReLU()
-c = nn.Linear(10, 3)
-d = nn.Linear(10, 3)
-e = nn.CMulTable()
-f = nn.Linear(3, 2)
+c = nn.Linear(10, 15)
+d = nn.CMulTable()
+e = nn.Mul(-1)
---[[
+dag:connect(a, b, c)
+dag:connect(b, nn.Linear(10, 15), nn.ReLU(), d)
+dag:connect(c, d)
+dag:connect(c, e)
- a -----> b ---> c ----> e ---
- \ /
- \--> d ---/
- \
- \---> f ---
-]]--
+dag:setLabel(a, 'first module')
+
+dag:setInput(a)
+dag:setOutput({ d, e })
+
+-- Check the output of the dot file. Generate a pdf with:
+--
+-- dot ./graph.dot -Lg -T pdf -o ./graph.pdf
+--
+print('Writing ./graph.dot')
+dag:saveDot('./graph.dot')
-g = nn.DAG:new()
+-- Let's make a model where the dag is inside another nn.Container.
+model = nn.Sequential()
+ :add(nn.Linear(50, 50))
+ :add(dag)
+ :add(nn.CAddTable())
-g:setInput(a)
-g:setOutput({ e, f })
+criterion = nn.MSECriterion()
-g:addEdge(c, e)
-g:addEdge(a, b)
-g:addEdge(d, e)
-g:addEdge(b, c)
-g:addEdge(b, d)
-g:addEdge(d, f)
+if cunn then
+ print("Using CUDA")
+ model:cuda()
+ criterion:cuda()
+ torch.setdefaulttensortype('torch.CudaTensor')
+ epsilon = 1e-3
+end
-g:print()
+local input = torch.Tensor(30, 50):uniform()
+local output = model:updateOutput(input):clone()
+output:uniform()
-input = torch.Tensor(3, 10):uniform()
+-- Check that DAG:accGradParameters and friends work okay
+print('Gradient estimate error ' .. checkGrad(model, criterion, input, output, epsilon))
-output = g:updateOutput(input)
+-- Check that we can save and reload the model
+model:clearState()
+torch.save('./test.t7', model)
+local otherModel = torch.load('./test.t7')
+print('Gradient estimate error ' .. checkGrad(otherModel, criterion, input, output, epsilon))
-print(output[1])
-print(output[2])
+dag:print()
projects / dagnn.git / blobdiff