Skip to content

Commit

Permalink
day16
Browse files Browse the repository at this point in the history
  • Loading branch information
@hokekiyoo
hokekiyoo committed Dec 16, 2017
1 parent e346325 commit 3d1dbdd
Showing 1 changed file with 217 additions and 0 deletions.
217 changes: 217 additions & 0 deletions day16-neuralnet.ipynb
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,217 @@
{
"cells": [
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"%matplotlib inline\n",
"import numpy as np\n",
"from numpy.random import *\n",
"import matplotlib.pyplot as plt\n",
"from sklearn.datasets import make_blobs"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"class NN:\n",
" def __init__(self, num_input, num_hidden, num_output, learning_rate):\n",
" self.num_input = num_input\n",
" self.num_hidden = num_hidden\n",
" self.num_output = num_output\n",
" self.learning_rate = learning_rate\n",
"\n",
" self.w_input2hidden = np.random.random((self.num_hidden, self.num_input))\n",
" self.w_hidden2output = np.random.random((self.num_output, self.num_hidden))\n",
" self.b_input2hidden = np.ones((self.num_hidden))\n",
" self.b_hidden2output = np.ones((self.num_output))\n",
"\n",
" ##活性化関数（シグモイド関数）\n",
" def activate_func(self, x):\n",
" return 1/(1+np.exp(-x))\n",
" ##活性化関数の微分\n",
" def dactivate_func(self,x):\n",
" return self.activate_func(x)*(1-self.activate_func(x))\n",
" ##ソフトマックス関数\n",
" def softmax_func(self,x):\n",
" C = x.max()\n",
" f = np.exp(x-C)/np.exp(x-C).sum()\n",
" return f\n",
" ##順伝播計算_\n",
" def forward_propagation(self, x):\n",
" u_hidden = np.dot(self.w_input2hidden, x) + self.b_input2hidden\n",
" z_hidden = self.activate_func(u_hidden)\n",
" u_output = np.dot(self.w_hidden2output, z_hidden) + self.b_hidden2output\n",
" z_output = self.softmax_func(u_output)\n",
" return u_hidden, u_output, z_hidden, z_output\n",
" ##逆伝播でδを求める\n",
" def backward_propagation(self,t,u_hidden,z_output):\n",
" t_vec = np.zeros(len(z_output))\n",
" t_vec[t] = 1\n",
" delta_output = z_output - t_vec\n",
" delta_hidden = np.dot(delta_output, self.w_hidden2output * self.dactivate_func(u_hidden))\n",
" return delta_hidden, delta_output\n",
" ##損失関数のパラメータwに関する勾配\n",
" def calc_gradient(self,delta,z):\n",
" dW = np.zeros((len(delta), len(z)))\n",
" for i in range(len(delta)):\n",
" for j in range(len(z)):\n",
" dW[i][j] = delta[i] * z[j]\n",
" return dW\n",
" ##重みをアップデートする\n",
" def update_weight(self,w0,gradE):\n",
" return w0 - self.learning_rate*gradE"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false,
"scrolled": true
},
"outputs": [],
"source": [
"def train(nn, iteration,savefig=False):\n",
" epoch = 0\n",
" for epoch in range(iteration+1):\n",
" grad_i2h = 0\n",
" grad_h2o = 0\n",
" gradbias_i2h = 0\n",
" gradbias_h2o = 0\n",
" n = 0\n",
" err = 0\n",
" rand = randint(0,len(data),100)\n",
" for r in rand:\n",
" u_hidden, u_output, z_hidden, z_output = nn.forward_propagation(data[r])\n",
" delta_hidden, delta_output = nn.backward_propagation(target[r], u_hidden, z_output)\n",
" grad_i2h += nn.calc_gradient(delta_hidden, data[r])\n",
" grad_h2o += nn.calc_gradient(delta_output, z_hidden)\n",
" gradbias_i2h += delta_hidden\n",
" gradbias_h2o += delta_output\n",
" nn.w_input2hidden = nn.update_weight(nn.w_input2hidden, grad_i2h / len(rand))\n",
" nn.w_hidden2output = nn.update_weight(nn.w_hidden2output, grad_h2o / len(rand))\n",
" nn.b_input2hidden = nn.update_weight(nn.b_input2hidden, gradbias_i2h / len(rand))\n",
" nn.b_hidden2output = nn.update_weight(nn.b_hidden2output, gradbias_h2o / len(rand))\n",
" #print grad_h2o\n",
" if epoch%10 == 0:\n",
" print(epoch,\",\",end=\"\")\n",
" if savefig:\n",
" plt.figure(figsize=(5,5))\n",
" #色の用意\n",
" base_color = [\"red\",\"blue\",\"green\",\"yellow\",\"cyan\",\n",
" \"pink\",\"brown\",\"gray\",\"purple\",\"orange\"]\n",
" colors = [base_color[label] for label in target]\n",
" # 教師データのプロット\n",
" plt.scatter(data[:,0],data[:,1],color=colors,alpha=0.5)\n",
"\n",
" print(\"plotting...\")\n",
" xx = np.linspace(-15,15,80)\n",
" yy = np.linspace(-15,15,80)\n",
" for xi in xx:\n",
" for yi in yy:\n",
" _,_,_,z_output = nn.forward_propagation((xi, yi))\n",
" cls = np.argmax(z_output) #softmaxのスコアの最大のインデックス\n",
" score = np.max(z_output)\n",
" plt.plot(xi, yi, base_color[cls],marker=\"x\",alpha=score)\n",
" print(\".\",end=\"\")\n",
" plt.xlim(-15,15)\n",
" plt.ylim(-15,15)\n",
" plt.savefig(\"{}.png\".format(epoch))\n",
" plt.clf()\n",
" print(\"finish plotting\")\n",
" \n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": false
},
"outputs": [],
"source": [
"#データの用意\n",
"num_cls = 5\n",
"data,target = make_blobs(n_samples=1000, n_features=2, centers=num_cls)\n",
"#色の用意\n",
"base_color = [\"red\",\"blue\",\"green\",\"yellow\",\"cyan\",\n",
" \"pink\",\"brown\",\"gray\",\"purple\",\"orange\"]\n",
"colors = [base_color[label] for label in target]\n",
"# 教師データのプロット\n",
"plt.scatter(data[:,0],data[:,1],color=colors,alpha=0.5)\n",
"plt.savefig(\"original.png\")\n",
"# Neural Netの用意\n",
"nn = NN(num_input=2,num_hidden=20,num_output=num_cls,learning_rate=0.2)\n",
"# 学習\n",
"train(nn, iteration=300, savefig=False)\n",
"\n",
"plt.figure(figsize=(5,5))\n",
"\n",
"print(\"plotting...\")\n",
"xx = np.linspace(-15,15,70)\n",
"yy = np.linspace(-15,15,70)\n",
"for xi in xx:\n",
" for yi in yy:\n",
" _,_,_,z_output = nn.forward_propagation((xi, yi))\n",
" cls = np.argmax(z_output) #softmaxのスコアの最大のインデックス\n",
" score = np.max(z_output)\n",
" plt.plot(xi, yi, base_color[cls],marker=\"x\",alpha=score)\n",
" print(\".\",end=\"\")\n",
"plt.xlim(-15,15)\n",
"plt.ylim(-15,15)\n",
"plt.show()\n",
"print(\"finish plotting\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": []
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": []
}
],
"metadata": {
"anaconda-cloud": {},
"kernelspec": {
"display_name": "Python [Root]",
"language": "python",
"name": "Python [Root]"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.5.1"
}
},
"nbformat": 4,
"nbformat_minor": 0
}

0 comments on commit 3d1dbdd

Please sign in to comment.