{
"cells": [
{
"cell_type": "code",
"execution_count": 11,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "nSttyGQTLp2u"
},
"outputs": [],
"source": [
"#do not change\n",
"import sys\n",
"import os\n",
"from time import time\n",
"%matplotlib inline\n",
"from urllib.request import urlopen\n",
"import matplotlib.pyplot as plt\n",
"from sklearn.decomposition import PCA\n",
"import numpy as np\n",
"#from scipy.misc import imsave\n",
"import math\n",
" \n",
"def load_mnist():\n",
" images_url = 'https://github.com/guptashvm/Data/blob/master/data/train-images-idx3-ubyte?raw=true'\n",
" with urlopen(images_url) as urlopened:\n",
"    fd = urlopened.read()\n",
"    loaded = np.frombuffer(fd,dtype=np.uint8)\n",
"    trX = loaded[16:].reshape((60000,28*28)).astype(float)\n",
" \n",
" labels_url = 'https://github.com/guptashvm/Data/blob/master/data/train-labels-idx1-ubyte?raw=true'\n",
" with urlopen(labels_url) as urlopened:\n",
"    fd = urlopened.read()\n",
"    loaded = np.frombuffer(fd,dtype=np.uint8)\n",
"    trY = loaded[8:].reshape((60000))\n",
" \n",
" trY = np.asarray(trY)\n",
" \n",
" X = trX / 255.\n",
" y = trY\n",
" \n",
" subset = [i for i, t in enumerate(y) if t in [1, 0, 2, 3]]\n",
" X, y = X.astype('float32')[subset], y[subset]\n",
" return X[:1000], y[:1000]"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "4n1EPKtdLp20"
},
"source": [
"Run the following code to load the data we need. Here, **X** is the array of images, **y** is the array of labels, and **X2d** is the array of projections of the images into two-dimensional space using PCA. The two-dimensional scatterplot of the top two principal components of the images is displayed, where the color of each point represents its label. "
]
},
{
"cell_type": "code",
"execution_count": 12,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 269
},
"colab_type": "code",
"id": "2yULFRTMLp23",
"outputId": "c5701b02-984a-45e4-b2dd-3c0d52be12dc"
},

2. Consider a neural network with a single intermediate layer and a...