Permalink
Cannot retrieve contributors at this time
Name already in use
A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?
StepSizeMatters/theorem1.ipynb
Go to fileThis commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
257 lines (257 sloc)
6.94 KB
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
{ | |
"cells": [ | |
{ | |
"cell_type": "code", | |
"execution_count": null, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"import numpy as np\n", | |
"import tensorflow as tf\n", | |
"import matplotlib.pyplot as plt\n", | |
"import sklearn\n", | |
"import math\n", | |
"%matplotlib inline" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": null, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"# Setting random seed for numpy and tensorflow\n", | |
"tf.set_random_seed(963)\n", | |
"np.random.seed(963)" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 3, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"sess = tf.InteractiveSession()" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 4, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"def weight_variable(shape):\n", | |
" initial = tf.random_normal(shape, stddev=1.0)\n", | |
" return tf.Variable(initial)\n", | |
"\n", | |
"def bias_vector(size):\n", | |
" initial = np.random.normal(0, 1.0, size)\n", | |
" return initial" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 5, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"#Generate a series of input data points that match the condition in Theorem 1.\n", | |
"\n", | |
"def genPoints(size):\n", | |
" v = np.arange(1,size + 1,1)\n", | |
" vmag = np.linalg.norm(v)\n", | |
" vnorm = v / vmag\n", | |
" output = np.sqrt(size)*vnorm\n", | |
" return np.expand_dims(output,1)" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 6, | |
"metadata": {}, | |
"outputs": [ | |
{ | |
"name": "stdout", | |
"output_type": "stream", | |
"text": [ | |
"WARNING:tensorflow:From /usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.\n", | |
"Instructions for updating:\n", | |
"Colocations handled automatically by placer.\n" | |
] | |
} | |
], | |
"source": [ | |
"#Create deep linear neural network struture with following parameters:\n", | |
"\n", | |
"N = 1000 # size of dataset\n", | |
"FN = 1 # number of features in dataset\n", | |
"d = 10 # number of nodes per layer\n", | |
"numLayers = 1 # number of hidden layers\n", | |
"x_dat = genPoints(N) # generate data pts\n", | |
"\n", | |
"W = []\n", | |
"\n", | |
"W.append(weight_variable([FN,d]))\n", | |
"for i in range(0,numLayers):\n", | |
" W.append(weight_variable([d,d]))\n", | |
"W.append(weight_variable([d,1]))\n", | |
"\n", | |
"x = tf.placeholder(tf.float32, shape=[N,FN])\n", | |
"y_ = tf.placeholder(tf.float32, shape=[N,1])\n", | |
"\n", | |
"y = tf.identity(tf.matmul(x,W[0]))\n", | |
"\n", | |
"for i in range(1, numLayers+2):\n", | |
" y = tf.identity(tf.matmul(y, W[i]))" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 7, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"#\n", | |
"\n", | |
"train_loss = tf.reduce_sum(tf.math.scalar_mul((1/(2*N)),tf.square(y-y_)))" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 8, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"# generate a function that can be represented by the network\n", | |
"resolution = np.arange(0,2,0.002)\n", | |
"resolution = np.expand_dims(resolution,1)\n", | |
"sess.run(tf.global_variables_initializer())\n", | |
"y_dat = y.eval(feed_dict={x: x_dat})\n", | |
"resolution_y = y.eval(feed_dict={x: resolution})" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 9, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"# generate R (optimal weights) and find largest singular value in R\n", | |
"M = []\n", | |
"\n", | |
"for i in range(0, numLayers+2):\n", | |
" M.append(W[i].eval(session=sess))\n", | |
"\n", | |
"R = M[0]\n", | |
"for i in range(1, numLayers+2):\n", | |
" R = np.matmul(R,M[i])\n", | |
"\n", | |
"u, s, vh = np.linalg.svd(R, full_matrices=True)\n", | |
"\n", | |
"rLarge = s.max()" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 10, | |
"metadata": {}, | |
"outputs": [ | |
{ | |
"name": "stdout", | |
"output_type": "stream", | |
"text": [ | |
"0.046412347403755924\n" | |
] | |
} | |
], | |
"source": [ | |
"# calculate the upperbound on the step size shown in Theorem 1\n", | |
"L = numLayers+2\n", | |
"\n", | |
"deltaUp = 2.0 / (L * (rLarge ** (2*(L-1)/L)))\n", | |
"print(deltaUp)" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": 11, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"# Generate a series of decreasing step sizes from the calculated upperbound.\n", | |
"deltas = [deltaUp]\n", | |
"sum = deltaUp\n", | |
"for i in range(0, 19):\n", | |
" sum = sum - 0.005\n", | |
" if(sum > 0.005):\n", | |
" deltas.append(sum)" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": null, | |
"metadata": {}, | |
"outputs": [], | |
"source": [ | |
"# Run the Gradient Descent Algorithm for 1000 independent random initializations with a maximum threshold\n", | |
"# of 1000 iterations of GD. When the difference in gradient is less than the value 'tolerance', then the\n", | |
"# run has converged to a solution.\n", | |
"\n", | |
"tolerance = 0.001\n", | |
"numRuns = 1000\n", | |
"maxIters = 1000\n", | |
"for d in deltas:\n", | |
" numConverge = 0\n", | |
" for r in range(numRuns):\n", | |
" opt = tf.train.GradientDescentOptimizer(d)\n", | |
" train_step = opt.minimize(train_loss)\n", | |
" sess.run(tf.global_variables_initializer())\n", | |
" grads_and_vars = opt.compute_gradients(train_loss)\n", | |
" grad_norms = [tf.nn.l2_loss(g) for g, v in grads_and_vars]\n", | |
" grad_norm = tf.add_n(grad_norms)\n", | |
" loss,oldgr = sess.run([train_loss,grad_norm],feed_dict={x: x_dat, y_: y_dat})\n", | |
" for i in range(maxIters):\n", | |
" if math.isnan(loss) or math.isinf(loss):\n", | |
" break\n", | |
" if math.isnan(oldgr) or math.isinf(oldgr):\n", | |
" break\n", | |
" train_step.run(feed_dict={x: x_dat, y_: y_dat})\n", | |
" loss,newgr = sess.run([train_loss,grad_norm],feed_dict={x: x_dat, y_: y_dat})\n", | |
" if abs(newgr - oldgr) < tolerance:\n", | |
" numConverge = numConverge + 1\n", | |
" break\n", | |
" oldgr = newgr\n", | |
" print(\"delta[\"+ str(d) +\"]: percentage: \" + str(numConverge) + \"/\" + str(numRuns))" | |
] | |
}, | |
{ | |
"cell_type": "code", | |
"execution_count": null, | |
"metadata": {}, | |
"outputs": [], | |
"source": [] | |
} | |
], | |
"metadata": { | |
"anaconda-cloud": {}, | |
"kernelspec": { | |
"display_name": "Python 2", | |
"language": "python", | |
"name": "python2" | |
}, | |
"language_info": { | |
"codemirror_mode": { | |
"name": "ipython", | |
"version": 3 | |
}, | |
"file_extension": ".py", | |
"mimetype": "text/x-python", | |
"name": "python", | |
"nbconvert_exporter": "python", | |
"pygments_lexer": "ipython3", | |
"version": "3.5.3" | |
} | |
}, | |
"nbformat": 4, | |
"nbformat_minor": 1 | |
} |