1417 lines
44 KiB
Plaintext
1417 lines
44 KiB
Plaintext
|
{
|
|||
|
|
"cells": [
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "slide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Uczenie maszynowe\n",
|
|||
|
|
"# 5. Rekurencyjne sieci neuronowe (RNN), splotowe sieci neuronowe (CNN)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "slide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"## 5.1. Rekurencyjne sieci neuronowe (*Recurrent Neural Networks* – RNN)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"https://www.youtube.com/watch?v=WCUNPb-5EYI"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Rekurencyjna sieć neuronowa – schemat\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img style=\"margin: auto\" width=\"20%\" src=\"http://colah.github.io/posts/2015-08-Understanding-LSTMs/img/RNN-rolled.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Rekurencyjna sieć neuronowa – schemat\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img style=\"margin: auto\" width=\"80%\" src=\"http://colah.github.io/posts/2015-08-Understanding-LSTMs/img/RNN-unrolled.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Zależności długodystansowe (*long-distance dependencies*) w sieciach rekurencyjnych\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img style=\"margin: auto\" width=\"60%\" src=\"http://colah.github.io/posts/2015-08-Understanding-LSTMs/img/RNN-longtermdependencies.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### RNN – typy sekwencji\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img style=\"margin: auto\" width=\"80%\" src=\"http://karpathy.github.io/assets/rnn/diags.jpeg\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Prosta sieć RNN – schemat\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img src=\"rnn.png\" style=\"margin: auto;\" width=\"80%\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### LSTM – *Long Short Term Memory*\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img src=\"lstm.jpg\" style=\"margin: auto;\" width=\"80%\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"* Rekurencyjne sieci neuronowe znajduja zastosowanie w przetwarzaniu sekwencji, np. szeregów czasowych i tekstów.\n",
|
|||
|
|
"* LSTM są rozwinięciem RNN, umożliwiają „zapamiętywanie” i „zapominanie”."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Co potrafią generować rekurencyjne sieci neuronowe?\n",
|
|||
|
|
"\n",
|
|||
|
|
"http://karpathy.github.io/2015/05/21/rnn-effectiveness/"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Przewidywanie ciągów czasowych za pomocą LSTM – przykład\n",
|
|||
|
|
"\n",
|
|||
|
|
"https://machinelearningmastery.com/time-series-forecasting-long-short-term-memory-network-python/"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### GRU – *Gated Recurrent Unit*\n",
|
|||
|
|
"\n",
|
|||
|
|
"* Rodzaj rekurencyjnej sieci neuronowej wprwadzony w 2014 roku\n",
|
|||
|
|
"* Ma prostszą budowę niż LSTM (2 bramki zamiast 3).\n"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### GRU – schemat\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img src=\"gru.png\" style=\"margin: auto;\" width=\"50%\" />\n"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### GRU vs LSTM\n",
|
|||
|
|
"LSTM – 3 bramki: wejścia (*input*), wyjścia (*output*) i zapomnienia (*forget*); GRU – 2 bramki: resetu (*reset*) i aktualizacji (*update*). Bramka resetu pełni podwójną funkcję: zastępuje bramki wyjścia i zapomnienia."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### GRU vs LSTM\n",
|
|||
|
|
"* GRU i LSTM mają podobną skuteczność, ale GRU dzięki prostszej budowie bywa bardziej wydajna.\n",
|
|||
|
|
"* LSTM sprawdza się lepiej w przetwarzaniu tekstu, ponieważ lepiej zapamiętuje zależności długosystansowe."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "slide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"## 5.2. Splotowe sieci neuronowe (*Convolutional Neural Networks* – CNN)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Splotowe (konwolucyjne) sieci neuronowe wykorzystuje się do:\n",
|
|||
|
|
"\n",
|
|||
|
|
"* rozpoznawania obrazu\n",
|
|||
|
|
"* analizy wideo\n",
|
|||
|
|
"* innych zagadnień o podobnej strukturze"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Innymi słowy, CNN przydają się, gdy mamy bardzo dużo danych wejściowych, w których istotne jest ich sąsiedztwo."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Przykład\n",
|
|||
|
|
"\n",
|
|||
|
|
"Dla uproszczenia przyjmijmy, że mamy dane w postaci jendowymiarowej – np. chcemy stwierdzić, czy na danym nagraniu obecny jest głos człowieka."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Tak wygląda nasze nagranie:"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"80%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv-9-xs.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"(ciąg próbek dźwiękowych – możemy traktować je jak jednowymiarowe „piksele”)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Najprostsza metoda – „zwykła” jednowarstwowa sieć neuronowa (każdy z każdym):"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"80%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv-9-F.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Wady:\n",
|
|||
|
|
"\n",
|
|||
|
|
"* dużo danych wejściowych\n",
|
|||
|
|
"* nie wykrywa własności „lokalnych” wejścia"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Chcielibyśmy wykrywać pewne lokalne „wzory” w danych wejściowych."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"W tym celu tworzymy mniejszą sieć neuronową (mniej neuronów wejściowych) i _kopiujemy_ ją tak, żeby każda jej kopia działała na pewnym fragmencie wejścia (fragmenty mogą nachodzić na siebie):"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"60%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv-9-Conv2.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Każda z sieci A ma 2 neurony wejściowe (mało realistycznie). "
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"80%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv-9-Conv3.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Każda z sieci A ma 3 neurony wejściowe (wciąż mało realistycznie, ale już trochę bardziej). "
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Warstwę sieci A nazywamy **warstwą splotową (konwolucyjną)**.\n",
|
|||
|
|
"\n",
|
|||
|
|
"Warstw splotowych może być więcej niż jedna:"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"50%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv-9-Conv2Conv2.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"W dwóch wymiarach wygląda to tak:"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"50%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv2-9x5-Conv2.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"50%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv2-9x5-Conv2Conv2.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Zblizenie na pojedynczą jednostkę A:"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" height=\"80%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv2-unit.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Tak definiujemy formalnie funckję splotu dla 2 wymiarów:\n",
|
|||
|
|
"\n",
|
|||
|
|
"$$\n",
|
|||
|
|
"\\left[\\begin{array}{ccc}\n",
|
|||
|
|
"a & b & c\\\\\n",
|
|||
|
|
"d & e & f\\\\\n",
|
|||
|
|
"g & h & i\\\\\n",
|
|||
|
|
"\\end{array}\\right]\n",
|
|||
|
|
"*\n",
|
|||
|
|
"\\left[\\begin{array}{ccc}\n",
|
|||
|
|
"1 & 2 & 3\\\\\n",
|
|||
|
|
"4 & 5 & 6\\\\\n",
|
|||
|
|
"7 & 8 & 9\\\\\n",
|
|||
|
|
"\\end{array}\\right] \n",
|
|||
|
|
"=\\\\\n",
|
|||
|
|
"(1 \\cdot a)+(2 \\cdot b)+(3 \\cdot c)+(4 \\cdot d)+(5 \\cdot e)\\\\+(6 \\cdot f)+(7 \\cdot g)+(8 \\cdot h)+(9 \\cdot i)\n",
|
|||
|
|
"$$\n",
|
|||
|
|
"\n",
|
|||
|
|
"Więcej: https://en.wikipedia.org/wiki/Kernel_(image_processing)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"A tak to mniej więcej działa:"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" height=\"70%\" src=\"https://devblogs.nvidia.com/wp-content/uploads/2015/11/Convolution_schematic.gif\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Jednostka warstwy konwolucyjnej może się składać z jednej lub kilku warstw neuronów:"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"40%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv-A.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"60%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv-A-NIN.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Jeden neuron może odpowiadać np. za wykrywanie pionowych krawędzi, drugi poziomych, a jeszcze inny np. krzyżujących się linii."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Przykładowe filtry, których może nauczyć się pierwsza warstwa konwolucyjna:\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img style=\"margin: auto\" width=\"80%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/KSH-filters.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"70%\" src=\"cnn-features.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### _Pooling_"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Obrazy składają się na ogół z milionów pikseli. Oznacza to, że nawet po zastosowaniu kilku warstw konwolucyjnych mielibyśmy sporo parametrów do wytrenowania.\n",
|
|||
|
|
"\n",
|
|||
|
|
"Żeby zredukować liczbę parametrów, a dzięki temu uprościć obliczenia, stosuje się warstwy ***pooling***."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"*Pooling* to rodzaj próbkowania. Najpopularniejszą jego odmianą jest *max-pooling*, czyli wybieranie najwyższej wartości spośród kilku sąsiadujących pikseli."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"60%\" src=\"https://upload.wikimedia.org/wikipedia/commons/e/e9/Max_pooling.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Warstwy _pooling_ i konwolucyjne można przeplatać ze sobą:"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"40%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/Conv2-9x5-Conv2Max2Conv2.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"_Pooling_ – idea: nie jest istotne, w którym *dokładnie* miejscu na obrazku dana cecha (krawędź, oko, itp.) się znajduje, wystarczy przybliżona lokalizacja."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Do sieci konwolucujnych możemy dokładać też warstwy ReLU."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"https://www.youtube.com/watch?v=FmpDIaiMIeA"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Możliwości konwolucyjnych sieci neuronowych"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"40%\" src=\"http://colah.github.io/posts/2014-07-Conv-Nets-Modular/img/KSH-results.png\"/>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Przykład: MNIST"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 23,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "notes"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [],
|
|||
|
|
"source": [
|
|||
|
|
"%matplotlib inline\n",
|
|||
|
|
"\n",
|
|||
|
|
"import math\n",
|
|||
|
|
"import matplotlib.pyplot as plt\n",
|
|||
|
|
"import numpy as np\n",
|
|||
|
|
"import random\n",
|
|||
|
|
"\n",
|
|||
|
|
"from IPython.display import YouTubeVideo"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 24,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "notes"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [],
|
|||
|
|
"source": [
|
|||
|
|
"# źródło: https://github.com/keras-team/keras/examples/minst_mlp.py\n",
|
|||
|
|
"\n",
|
|||
|
|
"import keras\n",
|
|||
|
|
"from keras.datasets import mnist\n",
|
|||
|
|
"\n",
|
|||
|
|
"from keras.models import Sequential\n",
|
|||
|
|
"from keras.layers import Dense, Dropout, Flatten\n",
|
|||
|
|
"from keras.layers import Conv2D, MaxPooling2D\n",
|
|||
|
|
"\n",
|
|||
|
|
"# załaduj dane i podziel je na zbiory uczący i testowy\n",
|
|||
|
|
"(x_train, y_train), (x_test, y_test) = mnist.load_data()"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 25,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "notes"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [],
|
|||
|
|
"source": [
|
|||
|
|
"def draw_examples(examples, captions=None):\n",
|
|||
|
|
" plt.figure(figsize=(16, 4))\n",
|
|||
|
|
" m = len(examples)\n",
|
|||
|
|
" for i, example in enumerate(examples):\n",
|
|||
|
|
" plt.subplot(100 + m * 10 + i + 1)\n",
|
|||
|
|
" plt.imshow(example, cmap=plt.get_cmap('gray'))\n",
|
|||
|
|
" plt.show()\n",
|
|||
|
|
" if captions is not None:\n",
|
|||
|
|
" print(6 * ' ' + (10 * ' ').join(str(captions[i]) for i in range(m)))"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 26,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "fragment"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [
|
|||
|
|
{
|
|||
|
|
"data": {
|
|||
|
|
"image/png": "iVBORw0KGgoAAAANSUhEUgAAA6IAAACPCAYAAADgImbyAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nt2zZUmX/GpEdO3ZU++x
|
|||
|
|
"text/plain": [
|
|||
|
|
"<matplotlib.figure.Figure at 0x7f70ba2e9090>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
"metadata": {},
|
|||
|
|
"output_type": "display_data"
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"name": "stdout",
|
|||
|
|
"output_type": "stream",
|
|||
|
|
"text": [
|
|||
|
|
" 5 0 4 1 9 2 1\n"
|
|||
|
|
]
|
|||
|
|
}
|
|||
|
|
],
|
|||
|
|
"source": [
|
|||
|
|
"draw_examples(x_train[:7], captions=y_train)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 27,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [],
|
|||
|
|
"source": [
|
|||
|
|
"batch_size = 128\n",
|
|||
|
|
"num_classes = 10\n",
|
|||
|
|
"epochs = 12\n",
|
|||
|
|
"\n",
|
|||
|
|
"# input image dimensions\n",
|
|||
|
|
"img_rows, img_cols = 28, 28"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 28,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "notes"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [],
|
|||
|
|
"source": [
|
|||
|
|
"if keras.backend.image_data_format() == 'channels_first':\n",
|
|||
|
|
" x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)\n",
|
|||
|
|
" x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)\n",
|
|||
|
|
" input_shape = (1, img_rows, img_cols)\n",
|
|||
|
|
"else:\n",
|
|||
|
|
" x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)\n",
|
|||
|
|
" x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)\n",
|
|||
|
|
" input_shape = (img_rows, img_cols, 1)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 29,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [
|
|||
|
|
{
|
|||
|
|
"name": "stdout",
|
|||
|
|
"output_type": "stream",
|
|||
|
|
"text": [
|
|||
|
|
"x_train shape: (60000, 28, 28, 1)\n",
|
|||
|
|
"60000 train samples\n",
|
|||
|
|
"10000 test samples\n"
|
|||
|
|
]
|
|||
|
|
}
|
|||
|
|
],
|
|||
|
|
"source": [
|
|||
|
|
"x_train = x_train.astype('float32')\n",
|
|||
|
|
"x_test = x_test.astype('float32')\n",
|
|||
|
|
"x_train /= 255\n",
|
|||
|
|
"x_test /= 255\n",
|
|||
|
|
"print('x_train shape: {}'.format(x_train.shape))\n",
|
|||
|
|
"print('{} train samples'.format(x_train.shape[0]))\n",
|
|||
|
|
"print('{} test samples'.format(x_test.shape[0]))\n",
|
|||
|
|
"\n",
|
|||
|
|
"# convert class vectors to binary class matrices\n",
|
|||
|
|
"y_train = keras.utils.to_categorical(y_train, num_classes)\n",
|
|||
|
|
"y_test = keras.utils.to_categorical(y_test, num_classes)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 30,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [],
|
|||
|
|
"source": [
|
|||
|
|
"model = Sequential()\n",
|
|||
|
|
"model.add(Conv2D(32, kernel_size=(3, 3),\n",
|
|||
|
|
" activation='relu',\n",
|
|||
|
|
" input_shape=input_shape))\n",
|
|||
|
|
"model.add(Conv2D(64, (3, 3), activation='relu'))\n",
|
|||
|
|
"model.add(MaxPooling2D(pool_size=(2, 2)))\n",
|
|||
|
|
"model.add(Dropout(0.25))\n",
|
|||
|
|
"model.add(Flatten())\n",
|
|||
|
|
"model.add(Dense(128, activation='relu'))\n",
|
|||
|
|
"model.add(Dropout(0.5))\n",
|
|||
|
|
"model.add(Dense(num_classes, activation='softmax'))"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 31,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [],
|
|||
|
|
"source": [
|
|||
|
|
"model.compile(loss=keras.losses.categorical_crossentropy,\n",
|
|||
|
|
" optimizer=keras.optimizers.Adadelta(),\n",
|
|||
|
|
" metrics=['accuracy'])"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 32,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [
|
|||
|
|
{
|
|||
|
|
"name": "stdout",
|
|||
|
|
"output_type": "stream",
|
|||
|
|
"text": [
|
|||
|
|
"Train on 60000 samples, validate on 10000 samples\n",
|
|||
|
|
"Epoch 1/12\n",
|
|||
|
|
"60000/60000 [==============================] - 333s - loss: 0.3256 - acc: 0.9037 - val_loss: 0.0721 - val_acc: 0.9780\n",
|
|||
|
|
"Epoch 2/12\n",
|
|||
|
|
"60000/60000 [==============================] - 342s - loss: 0.1088 - acc: 0.9683 - val_loss: 0.0501 - val_acc: 0.9835\n",
|
|||
|
|
"Epoch 3/12\n",
|
|||
|
|
"60000/60000 [==============================] - 366s - loss: 0.0837 - acc: 0.9748 - val_loss: 0.0429 - val_acc: 0.9860\n",
|
|||
|
|
"Epoch 4/12\n",
|
|||
|
|
"60000/60000 [==============================] - 311s - loss: 0.0694 - acc: 0.9788 - val_loss: 0.0380 - val_acc: 0.9878\n",
|
|||
|
|
"Epoch 5/12\n",
|
|||
|
|
"60000/60000 [==============================] - 325s - loss: 0.0626 - acc: 0.9815 - val_loss: 0.0334 - val_acc: 0.9886\n",
|
|||
|
|
"Epoch 6/12\n",
|
|||
|
|
"60000/60000 [==============================] - 262s - loss: 0.0552 - acc: 0.9835 - val_loss: 0.0331 - val_acc: 0.9890\n",
|
|||
|
|
"Epoch 7/12\n",
|
|||
|
|
"60000/60000 [==============================] - 218s - loss: 0.0494 - acc: 0.9852 - val_loss: 0.0291 - val_acc: 0.9903\n",
|
|||
|
|
"Epoch 8/12\n",
|
|||
|
|
"60000/60000 [==============================] - 218s - loss: 0.0461 - acc: 0.9859 - val_loss: 0.0294 - val_acc: 0.9902\n",
|
|||
|
|
"Epoch 9/12\n",
|
|||
|
|
"60000/60000 [==============================] - 219s - loss: 0.0423 - acc: 0.9869 - val_loss: 0.0287 - val_acc: 0.9907\n",
|
|||
|
|
"Epoch 10/12\n",
|
|||
|
|
"60000/60000 [==============================] - 218s - loss: 0.0418 - acc: 0.9875 - val_loss: 0.0299 - val_acc: 0.9906\n",
|
|||
|
|
"Epoch 11/12\n",
|
|||
|
|
"60000/60000 [==============================] - 218s - loss: 0.0388 - acc: 0.9879 - val_loss: 0.0304 - val_acc: 0.9905\n",
|
|||
|
|
"Epoch 12/12\n",
|
|||
|
|
"60000/60000 [==============================] - 218s - loss: 0.0366 - acc: 0.9889 - val_loss: 0.0275 - val_acc: 0.9910\n"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"data": {
|
|||
|
|
"text/plain": [
|
|||
|
|
"<keras.callbacks.History at 0x7f70b80b1a10>"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
"execution_count": 32,
|
|||
|
|
"metadata": {},
|
|||
|
|
"output_type": "execute_result"
|
|||
|
|
}
|
|||
|
|
],
|
|||
|
|
"source": [
|
|||
|
|
"model.fit(x_train, y_train,\n",
|
|||
|
|
" batch_size=batch_size,\n",
|
|||
|
|
" epochs=epochs,\n",
|
|||
|
|
" verbose=1,\n",
|
|||
|
|
" validation_data=(x_test, y_test))"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "code",
|
|||
|
|
"execution_count": 33,
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"outputs": [
|
|||
|
|
{
|
|||
|
|
"name": "stdout",
|
|||
|
|
"output_type": "stream",
|
|||
|
|
"text": [
|
|||
|
|
"('Test loss:', 0.027530849870144449)\n",
|
|||
|
|
"('Test accuracy:', 0.99099999999999999)\n"
|
|||
|
|
]
|
|||
|
|
}
|
|||
|
|
],
|
|||
|
|
"source": [
|
|||
|
|
"score = model.evaluate(x_test, y_test, verbose=0)\n",
|
|||
|
|
"print('Test loss:', score[0])\n",
|
|||
|
|
"print('Test accuracy:', score[1])"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "slide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"## 5.3. Autoencoder"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"* Uczenie nienadzorowane\n",
|
|||
|
|
"* Dane: zbiór nieanotowanych przykładów uczących $\\{ x^{(1)}, x^{(2)}, x^{(3)}, \\ldots \\}$, $x^{(i)} \\in \\mathbb{R}^{n}$"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Autoencoder (encoder-decoder)\n",
|
|||
|
|
"\n",
|
|||
|
|
"Sieć neuronowa taka, że:\n",
|
|||
|
|
"* warstwa wejściowa ma $n$ neuronów\n",
|
|||
|
|
"* warstwa wyjściowa ma $n$ neuronów\n",
|
|||
|
|
"* warstwa środkowa ma $k < n$ neuronów\n",
|
|||
|
|
"* $y^{(i)} = x^{(i)}$ dla każdego $i$"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"40%\" src=\"http://ufldl.stanford.edu/tutorial/images/Autoencoder636.png\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Co otrzymujemy dzięki takiej sieci?\n",
|
|||
|
|
"\n",
|
|||
|
|
"* $y^{(i)} = x^{(i)} \\; \\Longrightarrow \\;$ Autoencoder próbuje nauczyć się funkcji $h(x) \\approx x$, czyli funkcji identycznościowej.\n",
|
|||
|
|
"* Warstwy środkowe mają mniej neuronów niż warstwy zewnętrzne, więc żeby to osiągnąć, sieć musi znaleźć bardziej kompaktową (tu: $k$-wymiarową) reprezentację informacji zawartej w wektorach $x_{(i)}$.\n",
|
|||
|
|
"* Otrzymujemy metodę kompresji danych."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Innymi słowy:\n",
|
|||
|
|
"* Ograniczenia nałożone na reprezentację danych w warstwie ukrytej pozwala na „odkrycie” pewnej **struktury** w danych.\n",
|
|||
|
|
"* _Decoder_ musi odtworzyć do pierwotnej postaci reprezentację danych skompresowaną przez _encoder_."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"60%\" src=\"https://upload.wikimedia.org/wikipedia/commons/2/28/Autoencoder_structure.png\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"70%\" src=\"autoencoder_schema.jpg\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"* Całkowita liczba warstw w sieci autoencodera może być większa niż 3.\n",
|
|||
|
|
"* Jako funkcji kosztu na ogół używa się błędu średniokwadratowego (*mean squared error*, MSE) lub entropii krzyżowej (*binary crossentropy*).\n",
|
|||
|
|
"* Autoencoder może wykryć ciekawe struktury w danych nawet jeżeli $k \\geq n$, jeżeli na sieć nałoży się inne ograniczenia.\n",
|
|||
|
|
"* W wyniku działania autoencodera uzyskujemy na ogół kompresję **stratną**."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Autoencoder a PCA\n",
|
|||
|
|
"\n",
|
|||
|
|
"Widzimy, że autoencoder można wykorzystać do redukcji liczby wymiarów. Podobną rolę pełni poznany na jednym z poprzednich wykładów algorytm PCA (analiza głównych składowych, *principal component analysis*).\n",
|
|||
|
|
"\n",
|
|||
|
|
"Faktycznie, jeżeli zastosujemy autoencoder z liniowymi funkcjami aktywacji i pojedynczą sigmoidalną warstwą ukrytą, to na podstawie uzyskanych wag można odtworzyć główne składowe używając rozkładu według wartości osobliwych (*singular value decomposition*, SVD)."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Autoencoder odszumiający\n",
|
|||
|
|
"\n",
|
|||
|
|
"Jeżeli na wejściu zamiast „czystych” danych użyjemy danych zaszumionych, to otrzymamy sieć, która może usuwać szum z danych:\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img style=\"margin: auto\" width=\"70%\" src=\"denoising_autoencoder.png\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"60%\" src=\"denoising.png\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Autoencoder – zastosowania\n",
|
|||
|
|
"\n",
|
|||
|
|
"Autoencoder sprawdza się gorzej niż inne algorytmy kompresji, więc nie stosuje się go raczej jako metody kompresji danych, ale ma inne zastosowania:\n",
|
|||
|
|
"* odszumianie danych\n",
|
|||
|
|
"* redukcja wymiarowości\n",
|
|||
|
|
"* VAE (*variational autoencoders*) – http://kvfrans.com/variational-autoencoders-explained/"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "slide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"## 5.4. Word embeddings"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"_Word embeddings_ – sposoby reprezentacji słów jako wektorów liczbowych"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Znaczenie wyrazu jest reprezentowane przez sąsiednie wyrazy:\n",
|
|||
|
|
"\n",
|
|||
|
|
"“A word is characterized by the company it keeps.” (John R. Firth, 1957)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"* Pomysł pojawił sie jeszcze w latach 60. XX w.\n",
|
|||
|
|
"* _Word embeddings_ można uzyskiwać na różne sposoby, ale dopiero w ostatnim dziesięcioleciu stało się opłacalne użycie w tym celu sieci neuronowych."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Przykład – 2 zdania: \n",
|
|||
|
|
"* \"have a good day\"\n",
|
|||
|
|
"* \"have a great day\"\n",
|
|||
|
|
"\n",
|
|||
|
|
"Słownik:\n",
|
|||
|
|
"* {\"a\", \"day\", \"good\", \"great\", \"have\"}"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"* Aby wykorzystać metody uczenia maszynowego do analizy danych tekstowych, musimy je jakoś reprezentować jako liczby."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"* Najprostsza metoda to wektory jednostkowe:\n",
|
|||
|
|
" * \"a\" = $(1, 0, 0, 0, 0)$\n",
|
|||
|
|
" * \"day\" = $(0, 1, 0, 0, 0)$\n",
|
|||
|
|
" * \"good\" = $(0, 0, 1, 0, 0)$\n",
|
|||
|
|
" * \"great\" = $(0, 0, 0, 1, 0)$\n",
|
|||
|
|
" * \"have\" = $(0, 0, 0, 0, 1)$\n",
|
|||
|
|
"* Taka metoda nie uwzględnia jednak podobieństw i różnic między znaczeniami wyrazów."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Metody uzyskiwania *word embeddings*:\n",
|
|||
|
|
"* Common Bag of Words (CBOW)\n",
|
|||
|
|
"* Skip Gram\n",
|
|||
|
|
"\n",
|
|||
|
|
"Obie opierają się na odpowiednim użyciu autoencodera."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"80%\" src=\"we_autoencoder.png\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"*Common Bag of Words*\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img style=\"margin: auto\" width=\"40%\" src=\"cbow.png\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"*Skip Gram*\n",
|
|||
|
|
"\n",
|
|||
|
|
"<img style=\"margin: auto\" width=\"30%\" src=\"skipgram.png\" />"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Skip Gram a CBOW\n",
|
|||
|
|
"\n",
|
|||
|
|
"* Skip Gram lepiej reprezentuje rzadkie wyrazy i lepiej działa, jeżeli mamy mało danych.\n",
|
|||
|
|
"* CBOW jest szybszy i lepiej reprezentuje częste wyrazy."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"### Popularne modele _word embeddings_\n",
|
|||
|
|
"* Word2Vec (Google)\n",
|
|||
|
|
"* GloVe (Stanford)\n",
|
|||
|
|
"* FastText (Facebook)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "slide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"## 5.5. Tłumaczenie neuronowe\n",
|
|||
|
|
"\n",
|
|||
|
|
"*Neural Machine Translation* (NMT)"
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"Neuronowe tłumaczenie maszynowe również opiera się na modelu *encoder-decoder*:\n",
|
|||
|
|
"* *Encoder* koduje z języka źródłowego na abstrakcyjną reprezentację.\n",
|
|||
|
|
"* *Decoder* odkodowuje z abstrakcyjnej reprezentacji na język docelowy."
|
|||
|
|
]
|
|||
|
|
},
|
|||
|
|
{
|
|||
|
|
"cell_type": "markdown",
|
|||
|
|
"metadata": {
|
|||
|
|
"slideshow": {
|
|||
|
|
"slide_type": "subslide"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"source": [
|
|||
|
|
"<img style=\"margin: auto\" width=\"40%\" src=\"http://devblogs.nvidia.com/parallelforall/wp-content/uploads/2015/06/Figure2_NMT_system.png\"/>"
|
|||
|
|
]
|
|||
|
|
}
|
|||
|
|
],
|
|||
|
|
"metadata": {
|
|||
|
|
"celltoolbar": "Slideshow",
|
|||
|
|
"kernelspec": {
|
|||
|
|
"display_name": "Python 3",
|
|||
|
|
"language": "python",
|
|||
|
|
"name": "python3"
|
|||
|
|
},
|
|||
|
|
"language_info": {
|
|||
|
|
"codemirror_mode": {
|
|||
|
|
"name": "ipython",
|
|||
|
|
"version": 3
|
|||
|
|
},
|
|||
|
|
"file_extension": ".py",
|
|||
|
|
"mimetype": "text/x-python",
|
|||
|
|
"name": "python",
|
|||
|
|
"nbconvert_exporter": "python",
|
|||
|
|
"pygments_lexer": "ipython3",
|
|||
|
|
"version": "3.8.3"
|
|||
|
|
},
|
|||
|
|
"livereveal": {
|
|||
|
|
"start_slideshow_at": "selected",
|
|||
|
|
"theme": "white"
|
|||
|
|
}
|
|||
|
|
},
|
|||
|
|
"nbformat": 4,
|
|||
|
|
"nbformat_minor": 4
|
|||
|
|
}
|