Add resources res

README.md

@@ -6,35 +6,35 @@ This aims to be a growing collections of introductory video tutorials on the [*T
 *Torch* is one of the fastest and most flexible framework existing for Machine and Deep Learning.
 And yes, flexibility was used to come with an intimidatig learning curve... until now.
 
-Enjoy the view of these videos and code transcripts (which will be added soon).
+Enjoy the view of these videos, transcripts and quizes (you can find in the [`res`](res) folder together with some notes about how I made these videos).
 
 
 ## 1 - Get the basics straight
 
-### 1.0 - An overview on *Lua*
+### 1.0 - An overview on *Lua* ([slides](res/1.0/slides.pdf))
 
 [![Practical 1.0 - Lua](http://img.youtube.com/vi/QLYLOPeI92g/0.jpg)](http://www.youtube.com/watch?v=QLYLOPeI92g)
 
-### 1.1 - An overview on *Torch*’s `Tensor`s
+### 1.1 - An overview on *Torch*’s `Tensor`s ([slides](res/1.1/slides.pdf))
 
 [![Practical 1.1 - Torch](http://img.youtube.com/vi/o3aRgD1uzsc/0.jpg)](http://www.youtube.com/watch?v=o3aRgD1uzsc)
 
-### 1.2 - An overview on *Torch*’s `image` package
+### 1.2 - An overview on *Torch*’s `image` package ([slides](res/1.2/slides.pdf))
 
 [![Practical 1.2 - image package](http://img.youtube.com/vi/dEjvydjcwOE/0.jpg)](http://www.youtube.com/watch?v=dEjvydjcwOE)
 
 
 ## 2 - Artificial Neural Networks
 
-### 2.0 - Neural Networks – feed forward (inference)
+### 2.0 - Neural Networks – feed forward (inference) ([slides](res/2.0/slides.pdf), [quiz](res/2.0/quiz.tex))
 
 [![Practical 2.0 – NN forward](http://img.youtube.com/vi/hxA0wxibv8g/0.jpg)](http://www.youtube.com/watch?v=hxA0wxibv8g)
 
-### 2.1 - Neural Networks – back propagation (training)
+### 2.1 - Neural Networks – back propagation (training) ([slides](res/2.1/slides.pdf))
 
 [![Practical 2.1 - NN backward](http://img.youtube.com/vi/VaQUx7m3oR4/0.jpg)](http://www.youtube.com/watch?v=VaQUx7m3oR4)
 
-### 2.2 - Neural Networks – An overview on *Torch*’s `nn` package
+### 2.2 - Neural Networks – An overview on *Torch*’s `nn` package ([slides](res/2.2/slides.pdf), [script](res/2.2/script.lua))
 
 [![Practical 2.2 - nn package](http://img.youtube.com/vi/atZYdZ8hVCw/0.jpg)](http://www.youtube.com/watch?v=atZYdZ8hVCw)
 

res/2.0/compile.sh

@@ -0,0 +1,5 @@
+# Builds 35 different quizes naming them testXX, XX = 01 -> 35
+
+for i in 0{1..9} {10..35}; do
+    rubber --pdf --jobname test$i quiz.tex
+done

res/2.0/quiz.tex

@@ -0,0 +1,37 @@
+\documentclass{article}
+\usepackage{pgf}
+\usepackage{amsmath,amssymb,bm}
+% Random int
+\pgfmathsetseed{\number\pdfrandomseed}
+\newcommand\rint{\pgfmathparse{random(10)}\pgfmathresult}
+\begin{document}
+
+\title{Quiz: forward propagation dimensionality}
+\author{BME595 DeepLearning}
+\date{\today}
+\maketitle
+
+Given a network with the size vector $\bm{s}=(\rint, \rint, \rint, \rint)^\top$, determine $n$, $K$, $L$ and write the dimensionality of:
+%
+\begin{align*}
+  \bm{x}            \\
+  \bm{\hat x}       \\
+  \bm{a}^{(1)}      \\
+  \bm{\hat a}^{(1)} \\
+  \bm{\Theta}^{(1)} \\
+  \bm{z}^{(2)}      \\
+  \bm{a}^{(2)}      \\
+  \bm{\hat a}^{(2)} \\
+  \bm{\Theta}^{(2)} \\
+  \bm{z}^{(3)}      \\
+  \bm{a}^{(3)}      \\
+  \bm{\hat a}^{(3)} \\
+  \bm{\Theta}^{(3)} \\
+  \bm{z}^{(4)}      \\
+  \bm{a}^{(4)}      \\
+  \bm{\hat a}^{(4)} \\
+  h_{\bm{\Theta}}(\bm{x}) \\
+  \bm{y}
+\end{align*}
+
+\end{document}

res/2.2/script.lua

@@ -0,0 +1,109 @@
+-- Recording script, not a runnable script
+
+require 'nn';
+lin = nn.Linear(5, 3)
+lin
+{lin}
+lin.weight
+lin.bias
+Theta_1 = torch.cat(lin.bias, lin.weight, 2) -- New Tensor
+Theta_1
+lin:zeroGradParameters()
+
+sig = nn.Sigmoid()
+{sig}
+require 'gnuplot';
+z = torch.linspace(-10, 10, 21)
+gnuplot.plot(z, sig:forward(z))
+-- Forward pass
+x = torch.randn(5)
+a1 = x
+h_Theta = sig:forward(lin:forward(x)):clone()
+z2 = Theta_1 * torch.cat(torch.ones(1), x, 1)
+a2 = z_1:clone():apply(function (z) return 1/(1 + math.exp(-z)) end)
+
+-- Backward pass
+loss = nn.MSECriterion()
+? nn.MSECriterion
+loss
+loss.sizeAverage = false
+y = torch.rand(3)
+-- forward(input, target)
+E = loss:forward(h_Theta, y)
+E
+(h_Theta - y):pow(2):sum()
+
+dE_dh = loss:updateGradInput(h_Theta, y):clone()
+dE_dh
+2 * (h_Theta - y)
+
+delta2 = sig:updateGradInput(z2, dE_dh)
+dE_dh:clone():cmul(a2):cmul(1 - a2)
+
+lin:accGradParameters(x, delta2)
+{lin}
+lin.gradWeight
+lin.gradBias
+delta2:view(-1, 1) * torch.cat(torch.ones(1), x, 1):view(1, -1)
+
+lin_gradInput = lin:updateGradInput(x, delta2)
+lin.weight:t() * delta2
+
+net = nn.Sequential()
+net:add(lin);
+net:add(sig);
+net
+
+-- While true
+pred = net:forward(x)
+pred
+h_Theta
+err = loss:forward(pred, y)
+err
+E
+gradCriterion = loss:backward(pred, y)
+gradCriterion
+dE_dh
+net:zeroGradParameters()
+net:get(1)
+torch.cat(net:get(1).gradBias, net:get(1).gradWeight, 2)
+
+oldWeight = net:get(1).weight:clone()
+oldBias = net:get(1).bias:clone()
+etha = 0.01
+net:updateParameters(etha)
+net:get(1).weight
+oldWeight - 0.01 * net:get(1).gradWeight
+net:get(1).bias
+oldBias - 0.01 * net:get(1).gradBias
+
+-- X: design matrix
+-- Y: labels / targets matrix / vector
+
+for i = 1, m do
+   local pred = net:forward(X[i])
+   local err = criterion:forward(pred, Y[i])
+   local gradCriterion = criterion:backward(pred, Y[i])
+   net:zeroGradParameters()
+   net:backward(X[i], gradCriterion)
+   net:updateParameters(learningRate)
+end
+
+for i = 1, m, batchSize do
+   net:zeroGradParameters()
+   for j = 0, batchSize - 2 do
+      if i+j > m then break end
+      local pred = net:forward(X[i+j])
+      local err = criterion:forward(pred, Y[i+j])
+      local gradCriterion = criterion:backward(pred, Y[i+j])
+      net:backward(X[i+j], gradCriterion)
+   end
+   net:updateParameters(learningRate)
+end
+
+dataset = {}
+function dataset:size() return m end
+for i = 1, m do dataset[i] = {X[i], Y[i]} end
+trainer = nn.StochasticGradient(net, loss)
+trainer:train(dataset)
+