def raw_interpolation(one, two, count):
images=np.empty([784,count])
for ii in xrange(count):
current_patch=((patches[:,one]*ii)+(patches[:,two]*(count-ii)))/count
images[:,ii]=current_patch
display_network.display_network('interpolations/raw_interpolation'+str(one)+'_'+str(two)+'.png',images)
print "Figure written to 'raw_interpolation.png'"
def deep_interpolation(one,two,count,W1,W2,b1,b2):
deep_image1=np.dot(W1,patches[:,one])
deep_image2=np.dot(W1,patches[:,two])
images=np.empty([784,count])
for ii in xrange(count):
deep_interpolation_img=((deep_image1*ii)+deep_image2*(count-ii))/count
images[:,ii]=np.dot(W2,deep_interpolation_img)
display_network.display_network('interpolations/deep_interpolation'+str(one)+'_'+str(two)+'.png',images)
print "Figure Written to 'deep_interpolation.png'"
def complex_train_test():
np.random.seed(0)
do_images = True
train_images = load_MNIST.load_MNIST_images('data/train-images-idx3-ubyte')
dsize = 10000
patches = train_images[:, :dsize]
fs = [dsize, 28 * 28, 196, 28 * 28]
cost, train_op = cost_and_grad(fs=fs,
X0=patches,
lambda_=3e-3,
rho=0.1,
beta=3,
lr=0.1)
sess = tf.get_default_session()
u.reset_time()
old_cost = sess.run(cost)
old_i = 0
frame_count = 0
costs = []
for i in range(2000):
cost0, _ = sess.run([cost, train_op])
costs.append(cost0)
if i % 100 == 0:
print(cost0)
# filters are transposed in visualization
if ((old_cost - cost0) / old_cost > 0.05
or i - old_i > 50) and do_images:
Wf_ = sess.run("Wf_var/read:0")
W1_ = u.unflatten_np(Wf_, fs[1:])[0]
display_network.display_network(W1_.T,
filename="pics/weights-%03d.png" %
(frame_count, ))
frame_count += 1
old_cost = cost0
old_i = i
u.record_time()
# u.dump(costs, "costs_adam.csv")
u.dump(costs, "costs_adam_bn1.csv")
u.summarize_time()
"""
Step 0a: Load data
Here we provide the code to load natural image data into x.
x will be a 144 * 10000 matrix, where the kth column x(:, k) corresponds to
the raw image data from the kth 12x12 image patch sampled.
You do not need to change the code below.
"""
# Return patches from sample images
x = sample_images_raw('data/IMAGES_RAW.mat')
# n is the number of dimensions and m is the number of patches
n, m = x.shape
random_sel = np.random.randint(0, m, 200)
image_x = display_network(x[:, random_sel])
fig = plt.figure()
plt.imshow(image_x, cmap=plt.cm.gray)
plt.title('Raw patch images')
"""
Step 0b: Zero-mean the data (by row)
"""
x -= np.mean(x, axis=1).reshape(-1, 1)
"""
Step 1a: Implement PCA to obtain x_rot
Implement PCA to obtain x_rot, the matrix in which the data is expressed
with respect to the eigenbasis of sigma, which is the matrix U.
from scipy.optimize import minimize
from initial_params import initial_params
from sparse_autoencoder_cost import sparse_autoencoder_cost, sigmoid, der_sigmoid
from load_mnist import generate_patch, load_data
from display_network import display_network
visible_size = 28 * 28
hidden_size = 196
sparsity_param = 0.1
lamda = 0.003
beta = 3
images = np.transpose(load_data())[:, 0:10000]
patches = generate_patch()
theta = initial_params(visible_size, hidden_size)
J = lambda th: sparse_autoencoder_cost(
visible_size, hidden_size, th, lambda x: sigmoid(x), lambda x: der_sigmoid(x), lamda, beta, sparsity_param, images
)
options_ = {"maxiter": 800, "disp": True}
result = minimize(J, theta, method="L-BFGS-B", jac=True, options=options_)
opt_theta = result.x
print result
W1 = opt_theta[0 : hidden_size * visible_size].reshape(hidden_size, visible_size).transpose()
display_network(W1)
if slope_ratio < alpha and abs(target_delta) > 1e-6 and adaptive_step:
print("%.2f %.2f %.2f" % (cost0, cost1, slope_ratio))
print("Slope optimality %.2f, shrinking learning rate to %.2f" % (
slope_ratio,
lr0 * beta,
))
sess.run(lr_set, feed_dict={lr_p: lr0 * beta})
else:
# see if our learning rate got too conservative, and increase it
# 99 was ideal for gradient
# if i>0 and i%50 == 0 and slope_ratio>0.99:
if i > 0 and i % 50 == 0 and slope_ratio > 0.90 and adaptive_step:
print("%.2f %.2f %.2f" % (cost0, cost1, slope_ratio))
print("Growing learning rate to %.2f" % (lr0 * growth_rate))
sess.run(lr_set, feed_dict={lr_p: lr0 * growth_rate})
if do_images and i > 0 and i % 100 == 0:
Wf_ = sess.run("Wf_var/read:0")
W1_ = u.unflatten_np(Wf_, fs[1:])[0]
display_network.display_network(W1_.T,
filename="pics/weights-%03d.png" %
(frame_count, ))
frame_count += 1
old_cost = cost0
old_i = i
u.record_time()
u.dump(costs, "new%d.csv" % (whitening_mode, ))
u.summarize_time()
train_labels = load_MNIST.load_MNIST_labels('data/train-labels-idx1-ubyte')
# implement MNIST sparse autoencoder
visibleSize = 28 * 28
hiddenSize = 196
sparsityParam = 0.1
lambda_ = 3e-3
beta = 3
patches = train_images[:, :10000]
opttheta = scipy.io.loadmat("opttheta.mat")["opttheta"].flatten()
W1 = opttheta[:hiddenSize * visibleSize].reshape((hiddenSize, visibleSize),
order='F')
# also need to transpose individual columns
display_network.display_network(transposeImageCols(W1.transpose()),
"sample2.jpg")
patchesTransposed = transposeImageCols(patches)
# J1 = lambda x: sparse_autoencoder.sparse_autoencoder_cost_matlab(x, visibleSize, hiddenSize, lambda_, sparsityParam, beta, patchesTransposed)
# cost,grad=J1(opttheta)
# print("My cost1 %.3f" %(cost))
# J2 = lambda x: sparse_autoencoder.sparse_autoencoder_cost(x, visibleSize, hiddenSize, lambda_, sparsityParam, beta, patchesTransposed)
# cost,grad=J2(opttheta)
# print("My cost2 %.3f" %(cost))
cost3 = sparse_autoencoder.sparse_autoencoder_cost_tf(opttheta, visibleSize,
hiddenSize, lambda_,
sparsityParam, beta,
patchesTransposed)
print("My cost4 %.3f" % (cost3))
import functools
import numpy as np
import pca
import sample_images
import display_network
if __name__=='__main__':
num_samples = 10000
num_samples = 10000
m = sample_images.load_matlab_images('IMAGES_RAW.mat')
patches = sample_images.sample(m, num_samples, size=(12,12), norm=None)
display_network.display_network('raw-patches.png', patches)
# ensure that patches have zero mean
mean = np.mean(patches, axis=0)
patches -= mean
assert np.allclose(np.mean(patches, axis=0), np.zeros(patches.shape[1]))
U, s, x_rot = pca.pca(patches)
covar = pca.covariance(x_rot)
display_network.array_to_file('covariance.png', covar)
# percentage of variance
# cumulative sum
pov = np.array(functools.reduce(
lambda t,l: t+[t[-1]+l] if len(t) > 0 else [l],
sae_cost = partial(sparse_autoencoder.cost,
visible_size=visible_size,
hidden_size=hidden_size,
weight_decay=weight_decay,
beta=beta,
sparsity_param=sparsity_param,
data=patches)
# Train!
trained, cost, d = scipy.optimize.lbfgsb.fmin_l_bfgs_b(sae_cost,
theta,
maxfun=max_iter,
m=1000,
factr=1.0,
pgtol=1e-100,
iprint=1)
'''
# numerical approximation (way too slow!)
trained, cost, d = scipy.optimize.lbfgsb.fmin_l_bfgs_b(
lambda x: sae_cost(x)[0], theta,
approx_grad=True,
maxfun=max_iter,
m=1000,
iprint=1)
'''
# Save the trained weights
W1, W2, b1, b2 = sparse_autoencoder.unflatten_params(
trained, hidden_size, visible_size)
display_network.display_network('weights.png', W1.T)
# debug (set to True in Ex 3)
debug = True
# ======================================================================
# Exercise 1: Load MNIST
# In this exercise, you will load the mnist dataset
# First download the dataset from the following website:
# Training Images: http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
# Training Labels: http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
# Loading Sample Images
# Loading 10K images from MNIST database
images = load_MNIST.load_MNIST_images('train-images.idx3-ubyte')
patches = images[:, 0:10000]
patches = patches[:, 1:200]
display_network.display_network(patches[:, 0:100])
# Now you will use the display network function to display different sets if the MNIST dataset
# The display is saved in the directory under the name weigths.
# Display 10, 50 and 100 datasets
### YOUR CODE HERE ###
# display 10
display_network.display_network(patches[:, 0:10], 'weights10.png')
# display 50
display_network.display_network(patches[:, 0:50], 'weights50.png')
# display 100
display_network.display_network(patches[:, 0:100], 'weights100.png')
# Obtain random parameters theta
# You need to implement utils_hw.initialize in utils_hw.py
def test_mnist():
'''Test the convolutional autoencder using MNIST.'''
# %%
# load MNIST as before
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
mean_img = np.mean(mnist.train.images, axis=0)
ae = eXclusiveAutoencoder(
input_dimensions = 784,
layers = [
{
'n_channels': 144,
'reconstructive_regularizer': 1.0,
'weight_decay': 1.0,
'sparse_regularizer': 1.0,
'sparsity_level': 0.05,
'exclusive_regularizer': 1.0,
'exclusive_type': 'logcosh',
'exclusive_logcosh_scale': 10.0,
'corrupt_prob': 1.0,
'tied_weight': True,
'encode':'sigmoid', 'decode':'linear',
'pathways': [
# range(0, 144),
range(0, 96),
range(48, 144),
],
},
],
init_encoder_weight = None,
init_decoder_weight = None,
init_encoder_bias = None,
init_decoder_bias = None,
)
# %%
learning_rate = 0.01
n_reload_per_epochs = 10
n_display_per_epochs = 10000
batch_size = 2000
n_epochs = 100000
optimizer_list = []
for layer_i in range(1):
optimizer_list.append(AMSGrad(learning_rate).minimize(ae['layerwise_cost'][layer_i]['total'], var_list=[
ae['encoder_weight'][layer_i],
ae['encoder_bias'][layer_i],
# ae['decoder_weights'][layer_i],
# ae['decoder_biases'][layer_i],
]))
# optimizer_full = tf.train.AdamOptimizer(learning_rate).minimize(ae['cost']['total'])
optimizer_list.append(AMSGrad(learning_rate).minimize(ae['cost']['total']))
# %%
# We create a session to use the graph
sess = tf.Session()
writer = tf.summary.FileWriter('logs', sess.graph)
sess.run(tf.global_variables_initializer())
# %%
# Fit all training data
for optimizer_i, (optimizer) in enumerate(optimizer_list):
for epoch_i in range(n_epochs):
if (epoch_i) % n_reload_per_epochs == 0:
batch_xs, batch_ys = mnist.train.next_batch(batch_size)
batch_x0 = np.array([img - mean_img for img in batch_xs[np.where(np.any(np.array([
batch_ys[:, 0],
# batch_ys[:, 1],
# batch_ys[:, 2],
# batch_ys[:, 3],
# batch_ys[:, 4],
# batch_ys[:, 5],
# batch_ys[:, 6],
# batch_ys[:, 7],
# batch_ys[:, 8],
# batch_ys[:, 9],
]) == 1, axis=0))]])
batch_x1 = np.array([img - mean_img for img in batch_xs[np.where(np.any(np.array([
# batch_ys[:, 0],
batch_ys[:, 1],
# batch_ys[:, 2],
# batch_ys[:, 3],
# batch_ys[:, 4],
# batch_ys[:, 5],
# batch_ys[:, 6],
# batch_ys[:, 7],
# batch_ys[:, 8],
# batch_ys[:, 9],
]) == 1, axis=0))]])
batch_x2 = np.array([img - mean_img for img in batch_xs[np.where(np.any(np.array([
# batch_ys[:, 0],
# batch_ys[:, 1],
batch_ys[:, 2],
# batch_ys[:, 3],
# batch_ys[:, 4],
# batch_ys[:, 5],
# batch_ys[:, 6],
# batch_ys[:, 7],
# batch_ys[:, 8],
# batch_ys[:, 9],
]) == 1, axis=0))]])
min_batch_size_x01 = np.min((batch_x0.shape[0], batch_x1.shape[0]))
batch_x01 = 0.5*(batch_x0[:min_batch_size_x01]+batch_x1[:min_batch_size_x01])
min_batch_size_x012 = np.min((batch_x0.shape[0], batch_x1.shape[0], batch_x2.shape[0]))
batch_x012 = 0.333*(batch_x0[:min_batch_size_x012]+batch_x1[:min_batch_size_x012]+batch_x2[:min_batch_size_x012])
batch_x1 = np.array([img - mean_img for img in batch_xs[np.where(np.any(np.array([
# batch_ys[:, 0],
batch_ys[:, 1],
# batch_ys[:, 2],
# batch_ys[:, 3],
# batch_ys[:, 4],
# batch_ys[:, 5],
# batch_ys[:, 6],
# batch_ys[:, 7],
# batch_ys[:, 8],
# batch_ys[:, 9],
]) == 1, axis=0))]])
batch_x2 = np.array([img - mean_img for img in batch_xs[np.where(np.any(np.array([
# batch_ys[:, 0],
# batch_ys[:, 1],
batch_ys[:, 2],
# batch_ys[:, 3],
# batch_ys[:, 4],
# batch_ys[:, 5],
# batch_ys[:, 6],
# batch_ys[:, 7],
# batch_ys[:, 8],
# batch_ys[:, 9],
]) == 1, axis=0))]])
min_batch_size_x12 = np.min((batch_x1.shape[0], batch_x2.shape[0]))
batch_x12 = 0.5*(batch_x1[:min_batch_size_x12]+batch_x2[:min_batch_size_x12])
#
train = []
# train.append(batch_x012)
train.append(batch_x01)
train.append(batch_x12)
# train.append(np.array([img - mean_img for img in batch_xs[np.where(np.any(np.array([
# batch_ys[:, 0],
# batch_ys[:, 1],
# # batch_ys[:, 2],
# # batch_ys[:, 3],
# # batch_ys[:, 4],
# # batch_ys[:, 5],
# # batch_ys[:, 6],
# # batch_ys[:, 7],
# # batch_ys[:, 8],
# # batch_ys[:, 9],
# ]) == 1, axis=0))]]))
# train.append(np.array([img - mean_img for img in batch_xs[np.where(np.any(np.array([
# # batch_ys[:, 0],
# batch_ys[:, 1],
# batch_ys[:, 2],
# # batch_ys[:, 3],
# # batch_ys[:, 4],
# # batch_ys[:, 5],
# # batch_ys[:, 6],
# # batch_ys[:, 7],
# # batch_ys[:, 8],
# # batch_ys[:, 9],
# ]) == 1, axis=0))]]))
sess.run(optimizer, feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
if (epoch_i+1) % n_display_per_epochs == 0:
if not optimizer is optimizer_list[-1]:
cost_total = sess.run(ae['layerwise_cost'][optimizer_i]['total'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_reconstruction_error = sess.run(ae['layerwise_cost'][optimizer_i]['reconstruction_error'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_sparsity = sess.run(ae['layerwise_cost'][optimizer_i]['sparsity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_exclusivity = sess.run(ae['layerwise_cost'][optimizer_i]['exclusivity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_weight_decay = sess.run(ae['layerwise_cost'][optimizer_i]['weight_decay'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
print('layer:', optimizer_i+1, ', epoch:', epoch_i+1, ', total cost:', cost_total, ', recon error:', cost_reconstruction_error, ', sparsity:', cost_sparsity, ', weight decay:', cost_weight_decay, ', exclusivity: ', cost_exclusivity)
else:
cost_total = sess.run(ae['cost']['total'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_reconstruction_error = sess.run(ae['cost']['reconstruction_error'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_sparsity = sess.run(ae['cost']['sparsity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_exclusivity = sess.run(ae['cost']['exclusivity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_weight_decay = sess.run(ae['cost']['weight_decay'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
print('layer: full,', 'epoch:', epoch_i+1, ', total cost:', cost_total, ', recon error:', cost_reconstruction_error, ', sparsity:', cost_sparsity, ', weight decay:', cost_weight_decay, ', exclusivity: ', cost_exclusivity)
n_examples = 5120
test_xs, test_ys = mnist.test.next_batch(n_examples)
test_xs = np.array([img - mean_img for img in test_xs[np.where(np.any(np.array([
test_ys[:, 0],
test_ys[:, 1],
test_ys[:, 2],
# test_ys[:, 3],
# test_ys[:, 4],
# test_ys[:, 5],
# test_ys[:, 6],
# test_ys[:, 7],
# test_ys[:, 8],
# test_ys[:, 9],
]) == 1, axis=0))][:144]])
if not optimizer is optimizer_list[-1]:
recon = sess.run(ae['layerwise_y'][layer_i], feed_dict={ae['x']: test_xs})
else:
recon = sess.run(ae['y'], feed_dict={ae['x']: test_xs})
weights = sess.run(ae['encoder_weight'][0])
# weights = np.transpose(weights, axes=(3,0,1,2))
# display_network(batch_x012[:144].transpose(), filename='mnist_batch_01.png')
display_network(batch_x01[:144].transpose(), filename='mnist_batch_01.png')
display_network(batch_x12[:144].transpose(), filename='mnist_batch_12.png')
display_network(test_xs.transpose(), filename='mnist_test.png')
display_network(recon.reshape((144,784)).transpose(), filename='mnist_results.png')
display_network(weights, filename='mnist_weights.png')
writer.close()
return ae
a function that accepts one label argument (e.g. only get digits 5-9),
a boolean (True if use only train/validation sets,
False for test set),
and a num_samples integer to choose only the first N samples.
Only reads mnist data from a special pickled mnist file.
Returns array of images with shape (784, num_images).
"""
training, valid, testing = read_mnist_file(filename)
if train:
t = np.array([e for e,l in izip(training[0], training[1]) if test(l)])
v = np.array([e for e,l in izip(valid[0], valid[1]) if test(l)])
images = np.vstack([t, v]).T
tl = np.array([l for e,l in izip(training[0], training[1]) if test(l)])
vl = np.array([l for e,l in izip(valid[0], valid[1]) if test(l)])
labels = np.hstack([tl, vl])
else:
t = testing
images = np.array([e for e,l in izip(t[0], t[1]) if test(l)]).T
labels = np.array([l for e,l in izip(t[0], t[1]) if test(l)])
assert images.shape[1] == len(labels)
assert images.shape[0] == 784
patches = images[:,:num_samples]
labels = labels[:num_samples]
assert patches.shape[0] == 784
return patches, labels
if __name__=='__main__':
train, valid, test = read_mnist_file('../data/mnist.pkl.gz')
display_network.display_network('mnist.png', train[0].T[:,:100])
# number of input units
visible_size = 28 * 28
# number of input units
hidden_size = 25
# desired average activation of the hidden units.
# (This was denoted by the Greek alphabet rho, which looks like a lower-case "p",
# in the lecture notes).
# weight decay parameter
lambda_ = 0.0001
# debug (set to True in Ex 3)
debug = False
##======================================================================
## Ex 1: Load MNIST
## In this example, you will load the mnist dataset
## First download the dataset from the following website:
##Training Images: http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
##Training Labels: http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
# Loading Sample Images
#Loading 10K images from MNIST database
images = load_MNIST.load_MNIST_images('../../data/mnist/train-images-idx3-ubyte')
patches = images[:, 0:10000]
patches = patches[:,1:200]
display_network.display_network(patches[:,1:100])
import random import display_network import numpy as np ##================================================================ ## Step 0a: Load data # Here we provide the code to load natural image data into x. # x will be a 144 * 10000 matrix, where the kth column x(:, k) corresponds to # the raw image data from the kth 12x12 image patch sampled. # You do not need to change the code below. patches = sample_images.sample_images_raw() num_samples = patches.shape[1] random_sel = random.sample(range(num_samples), 400) display_network.display_network(patches[:, random_sel], 'raw_pca.png') ##================================================================ ## Step 0b: Zero-mean the data (by row) # You can make use of the mean and repmat/bsxfun functions. # patches = patches - patches.mean(axis=0) patch_mean = patches.mean(axis=1) patches = patches - np.tile(patch_mean, (patches.shape[1], 1)).transpose() ##================================================================ ## Step 1a: Implement PCA to obtain xRot # Implement PCA to obtain xRot, the matrix in which the data is expressed # with respect to the eigenbasis of sigma, which is the matrix U. sigma = patches.dot(patches.transpose()) / patches.shape[1]
import numpy as np
import scipy.optimize
import matplotlib.pyplot as plt
from sparse_autoencoder import sparse_autoencoder_cost, initialize_parameters
from display_network import display_network
from load_MNIST import load_MNIST_images
# Loading 10K images from MNIST database
images = load_MNIST_images('data/mnist/train-images-idx3-ubyte')
patches = images[:, :10000]
n_patches = patches.shape[1] # Number of patches
# Randomly sample 200 patches and save as an image file
image = display_network(patches[:, [np.random.randint(n_patches) for i in range(200)]])
plt.figure()
plt.imsave('sparse_autoencoder_minist_patches.png', image, cmap=plt.cm.gray)
plt.imshow(image, cmap=plt.cm.gray)
visible_size = patches.shape[0] # Number of input units
hidden_size = 196 # Number of hidden units
weight_decay_param = 3e-3 # Weight decay parameter, which is the lambda in lecture notes
beta = 3 # Weight of sparsity penalty term
sparsity_param = 0.1 # Desired average activation of the hidden units.
# Randomly initialize the fitting parameters
theta = initialize_parameters(hidden_size, visible_size)
import numpy as np
import scipy.optimize
import matplotlib.pyplot as plt
from sparse_autoencoder import sparse_autoencoder_cost, initialize_parameters
from display_network import display_network
from load_MNIST import load_MNIST_images
# Loading 10K images from MNIST database
images = load_MNIST_images('data/mnist/train-images-idx3-ubyte')
patches = images[:, :10000]
n_patches = patches.shape[1] # Number of patches
# Randomly sample 200 patches and save as an image file
image = display_network(
patches[:, [np.random.randint(n_patches) for i in range(200)]])
plt.figure()
plt.imsave('sparse_autoencoder_minist_patches.png', image, cmap=plt.cm.gray)
plt.imshow(image, cmap=plt.cm.gray)
visible_size = patches.shape[0] # Number of input units
hidden_size = 196 # Number of hidden units
weight_decay_param = 3e-3 # Weight decay parameter, which is the lambda in lecture notes
beta = 3 # Weight of sparsity penalty term
sparsity_param = 0.1 # Desired average activation of the hidden units.
# Randomly initialize the fitting parameters
theta = initialize_parameters(hidden_size, visible_size)
input_size = 28 * 28 # Size of input vector (MNIST images are 28x28)
num_classes = 10 # Number of classes (MNIST images fall into 10 classes)
weight_decay = 1e-4 # Weight decay parameter
max_iter = 400
theta = 0.005 * np.random.randn(num_classes * input_size)
trained = softmax_train(input_size,
num_classes,
weight_decay,
data,
labels,
max_iter)
np.save('softmax.model', trained)
display_network.display_network('softmax.png', trained.T)
# test on the test data
test_data, test_labels = sample_images.get_mnist_data('../data/mnist.pkl.gz',
train=False,
num_examples=1e10)
predicted_labels = softmax_predict(trained, test_data)
assert len(predicted_labels) == len(test_labels)
print np.mean(predicted_labels == test_labels)
patches,_ = sample_images.get_mnist_data('../data/mnist.pkl.gz',
lambda l: l >= 5,
train=True,
num_samples=num_samples)
# set up L-BFGS args
theta = sparse_autoencoder.initialize_params(hidden_size, visible_size)
sae_cost = partial(sparse_autoencoder.cost,
visible_size=visible_size,
hidden_size=hidden_size,
weight_decay=weight_decay,
beta=beta,
sparsity_param=sparsity_param,
data=patches)
# Train!
trained, cost, d = scipy.optimize.lbfgsb.fmin_l_bfgs_b(sae_cost, theta,
maxfun=max_iter,
m=100,
factr=10.0,
pgtol=1e-8,
iprint=1)
# Save the trained weights
W1, W2, b1, b2 = sparse_autoencoder.unflatten_params(trained,
hidden_size,
visible_size)
display_network.display_network('mnist-features.png', W1.T)
np.save('mnist.5to9.model', trained)
# Set training options
options_ = {'maxiter': 1000, 'disp': True}
# Minimize cost function J by modifying parameters θ using L-BFGS-B optimization algo
result = scipy.optimize.minimize(J,
θ,
method='L-BFGS-B',
jac=True,
options=options_)
# Get optimized parameter vector
opt_θ = result.x
print(result)
############################################################################################
############################################################################################
############################################################################################
# =============================== Visualization ========================================== #
############################################################################################
# Slice fan-in weight vector of the hidden layer from optimized parameter vector
W1 = opt_θ[0:hidden_size * input_size].reshape(hidden_size,
input_size).transpose()
# Reshape the fan-in weight vector to obtain features learned by each hidden layer neuron
display_network.display_network(W1,
filename='weights-' + str(hidden_size) + '-' +
str(ρ) + '-' + str(λ) + '-' + str(β) + '.png')
############################################################################################
############################################################################################
def test_bc():
'''Test the convolutional autoencder using MNIST.'''
# %%
# load MNIST as before
n_channels = 144
filter_size = 28
stride_size = 7
input_size = 128
n_epochs = 10000
n_display = 100
n_reload = 100
patch_per_image = 10
patch_per_display = 16
test_xs, _ = read_hematoxylin_data_sets(input_size, patch_per_image)
mean_img = np.mean(test_xs, axis=0)
# mean_img = np.zeros((128, 128))
ae = eXclusiveConvolutionalAutoencoder(
input_shape = [None, input_size, input_size, 1],
layers = [
{
'n_channels': n_channels,
'reconstructive_regularizer': 1.0,
'weight_decay': 1.0,
'sparse_regularizer': 1.0,
'sparsity_level': 0.05,
'exclusive_regularizer': 1.0,
'tied_weight': True,
'filter_size': filter_size,
'stride_size': stride_size,
'corrupt_prob': 0.5,
'padding_type': 'SAME',
'encode':'sigmoid', 'decode':'linear',
'pathways': [
range(0, 128),
range(0, 256),
],
},
],
init_encoder_weights = None,
init_decoder_weights = None,
init_encoder_biases = None,
init_decoder_biases = None,
)
# %%
learning_rate = 0.01
optimizer_list = []
for layer_i in range(1):
optimizer_list.append(tf.train.AdamOptimizer(learning_rate).minimize(ae['layerwise_cost'][layer_i]['total'], var_list=[
ae['encoder_weights'][layer_i],
ae['encoder_biases'][layer_i],
# ae['decoder_weights'][layer_i],
# ae['decoder_biases'][layer_i],
]))
optimizer_full = tf.train.AdamOptimizer(learning_rate).minimize(ae['cost']['total'])
# %%
# We create a session to use the graph
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# %%
# Fit all training data
for layer_i in range(1):
for epoch_i in range(n_epochs):
if (epoch_i) % n_reload == 0:
batch_xs, batch_ys = read_hematoxylin_data_sets(input_size, patch_per_image)
train = []
train.append(np.array([img.reshape((input_size, input_size, 1)) - mean_img.reshape((input_size, input_size, 1)) for img in batch_xs[np.where(np.any(np.array([
batch_ys[:, 0],
# batch_ys[:, 1],
]) == 1, axis=0))]]))
train.append(np.array([img.reshape((input_size, input_size, 1)) - mean_img.reshape((input_size, input_size, 1)) for img in batch_xs[np.where(np.any(np.array([
# batch_ys[:, 0],
batch_ys[:, 1],
]) == 1, axis=0))]]))
sess.run(optimizer_list[layer_i], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
if (epoch_i+1) % n_display == 0:
cost_total = sess.run(ae['layerwise_cost'][layer_i]['total'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_reconstruction_error = sess.run(ae['layerwise_cost'][layer_i]['reconstruction_error'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_sparsity = sess.run(ae['layerwise_cost'][layer_i]['sparsity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_exclusivity = sess.run(ae['layerwise_cost'][layer_i]['exclusivity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_weight_decay = sess.run(ae['layerwise_cost'][layer_i]['weight_decay'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
print('layer:{}, epoch:{:d}, cost:{:f}, recon:{:f}, sparsity:{:f}, weight:{:f}, exclusivity:{:f}'.format(
layer_i+1,
epoch_i+1,
cost_total,
cost_reconstruction_error,
cost_sparsity,
cost_weight_decay,
cost_exclusivity))
test_xs, test_ys = read_hematoxylin_data_sets(input_size, patch_per_image)
test_xs = np.array([img.reshape((input_size, input_size, 1)) - mean_img.reshape((input_size, input_size, 1)) for img in test_xs[np.where(np.any(np.array([
test_ys[:, 0],
test_ys[:, 1],
]) == 1, axis=0))][np.random.randint(0, test_xs.shape[0], patch_per_display)]])
recon = sess.run(ae['layerwise_y'][layer_i], feed_dict={ae['x']: test_xs})
weights = sess.run(ae['encoder_weights'][0])
bias = sess.run(ae['encoder_biases'][0])
data = {'weights': weights, 'bias': bias}
with open('bc.pickle', 'wb') as fp:
pickle.dump(data, fp, protocol=pickle.HIGHEST_PROTOCOL)
display_network(weights.transpose((3,0,1,2)).reshape((n_channels, filter_size*filter_size)).transpose(), filename='bc_weights.png')
display_network(test_xs.reshape((patch_per_display,input_size*input_size)).transpose(), filename='bc_test.png')
display_network(recon.reshape((patch_per_display,input_size*input_size)).transpose(), filename='bc_results.png')
for epoch_i in range(n_epochs):
if (epoch_i) % n_reload == 0:
batch_xs, batch_ys = read_hematoxylin_data_sets(input_size, patch_per_image)
train = []
train.append(np.array([img.reshape((input_size, input_size, 1)) - mean_img.reshape((input_size, input_size, 1)) for img in batch_xs[np.where(np.any(np.array([
batch_ys[:, 0],
# batch_ys[:, 1],
]) == 1, axis=0))]]))
train.append(np.array([img.reshape((input_size, input_size, 1)) - mean_img.reshape((input_size, input_size, 1)) for img in batch_xs[np.where(np.any(np.array([
# batch_ys[:, 0],
batch_ys[:, 1],
]) == 1, axis=0))]]))
sess.run(optimizer_full, feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
if (epoch_i+1) % n_display == 0:
cost_total = sess.run(ae['cost']['total'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_reconstruction_error = sess.run(ae['cost']['reconstruction_error'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_sparsity = sess.run(ae['cost']['sparsity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_exclusivity = sess.run(ae['cost']['exclusivity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_exclusivity = sess.run(ae['cost']['exclusivity'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
cost_weight_decay = sess.run(ae['cost']['weight_decay'], feed_dict={ae['training_x'][0]: train[0], ae['training_x'][1]: train[1]})
print('layer:{}, epoch:{:d}, cost:{:f}, recon:{:f}, sparsity:{:f}, weight:{:f}, exclusivity:{:f}'.format(
'F',
epoch_i+1,
cost_total,
cost_reconstruction_error,
cost_sparsity,
cost_weight_decay,
cost_exclusivity))
test_xs, test_ys = read_hematoxylin_data_sets(input_size, patch_per_image)
test_xs = np.array([img.reshape((input_size, input_size, 1)) - mean_img.reshape((input_size, input_size, 1)) for img in test_xs[np.where(np.any(np.array([
test_ys[:, 0],
test_ys[:, 1],
]) == 1, axis=0))][np.random.randint(0, test_xs.shape[0], patch_per_display)]])
recon = sess.run(ae['y'], feed_dict={ae['x']: test_xs})
weights = sess.run(ae['encoder_weights'][0])
bias = sess.run(ae['encoder_biases'][0])
data = {'weights': weights, 'bias': bias}
with open('bc.pickle', 'wb') as fp:
pickle.dump(data, fp, protocol=pickle.HIGHEST_PROTOCOL)
display_network(weights.transpose((3,0,1,2)).reshape((n_channels, filter_size*filter_size)).transpose(), filename='bc_weights.png')
display_network(test_xs.reshape((patch_per_display,input_size*input_size)).transpose(), filename='bc_test.png')
display_network(recon.reshape((patch_per_display,input_size*input_size)).transpose(), filename='bc_results.png')
mndata.train_img_fname = 'train-images.idx3-ubyte'
mndata.train_lbl_fname = 'train-labels.idx1-ubyte'
mndata.load_testing()
X = mndata.test_images
X0 = np.asarray(X)[:1000, :] / 256.0
X = X0
# K-MEANS CLUSTERING & PREDICTION
K = 10
kmeans = KMeans(n_clusters=K).fit(X)
pred_label = kmeans.predict(X)
print(type(kmeans.cluster_centers_.T))
print(kmeans.cluster_centers_.T.shape)
# PLOT CLUSTER CENTERS
A = display_network(kmeans.cluster_centers_.T, K, 1)
f1 = plt.imshow(A, interpolation='nearest', cmap="jet")
f1.axes.get_xaxis().set_visible(False)
f1.axes.get_yaxis().set_visible(False)
plt.show()
# SAVE PLOT
# a colormap and a normalization instance
cmap = plt.cm.jet
norm = plt.Normalize(vmin=A.min(), vmax=A.max())
# map the normalized data to colors
# image is now RGBA (512x512x4)
image = cmap(norm(A))
scipy.misc.imsave('aa.png', image) # pip install pillow
# PLOT NEIGHBORS
J = lambda theta : sparse_autoencoder_cost(theta, input_size, hidden_size_L1, lambda_, sparsity_param, beta, train_data)
options = {'maxiter': maxiter, 'disp': True}
results = scipy.optimize.minimize(J, sae1_theta, method='L-BFGS-B', jac=True, options=options)
sae1_opt_theta = results['x']
print("Show the results of optimization as following.\n")
print(results)
# Visualize weights
visualize_weights = False
if visualize_weights:
W1 = sae1_opt_theta[0:hidden_size_L1*input_size].reshape((hidden_size_L1, input_size))
image = display_network(W1.T)
plt.figure()
plt.imshow(image, cmap=plt.cm.gray)
plt.show()
"""
STEP 2: Train the second sparse autoencoder
This trains the second sparse autoencoder on the first autoencoder featurse.
If you've correctly implemented sparse_autoencoder_cost, you don't need
to change anything here.
"""
sae1_features = feedforward_autoencoder(sae1_opt_theta, hidden_size_L1, input_size, train_data)
hidden_size=hidden_size,
weight_decay = weight_decay,
beta=beta,
sparsity_param=sparsity_param,
data=patches)
# Train!
trained, cost, d = scipy.optimize.lbfgsb.fmin_l_bfgs_b(sae_cost, theta,
maxfun=max_iter,
m=1000,
factr=1.0,
pgtol=1e-100,
iprint=1)
'''
# numerical approximation (way too slow!)
trained, cost, d = scipy.optimize.lbfgsb.fmin_l_bfgs_b(
lambda x: sae_cost(x)[0], theta,
approx_grad=True,
maxfun=max_iter,
m=1000,
iprint=1)
'''
# Save the trained weights
W1, W2, b1, b2 = sparse_autoencoder.unflatten_params(trained,
hidden_size,
visible_size)
display_network.display_network('weights.png', W1.T)
import random import display_network import numpy as np ##================================================================ ## Step 0a: Load data # Here we provide the code to load natural image data into x. # x will be a 144 * 10000 matrix, where the kth column x(:, k) corresponds to # the raw image data from the kth 12x12 image patch sampled. # You do not need to change the code below. patches = sample_images.sample_images_raw() num_samples = patches.shape[1] random_sel = random.sample(range(num_samples), 400) display_network.display_network(patches[:, random_sel], 'raw_pca.png') ##================================================================ ## Step 0b: Zero-mean the data (by row) # You can make use of the mean and repmat/bsxfun functions. # patches = patches - patches.mean(axis=0) patch_mean = patches.mean(axis=1) patches = patches - np.tile(patch_mean, (patches.shape[1], 1)).transpose() ##================================================================ ## Step 1a: Implement PCA to obtain xRot # Implement PCA to obtain xRot, the matrix in which the data is expressed # with respect to the eigenbasis of sigma, which is the matrix U. sigma = patches.dot(patches.transpose()) / patches.shape[1]
# debug (set to True in Ex 3)
debug = False
# ======================================================================
# Exercise 1: Load MNIST
# In this exercise, you will load the mnist dataset
# First download the dataset from the following website:
# Training Images: http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
# Training Labels: http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
# Loading Sample Images
# Loading 10K images from MNIST database
images = load_MNIST.load_MNIST_images('train-images.idx3-ubyte')
patches = images[:, 0:10000]
patches = patches[:, 1:200]
display_network.display_network(patches[:, 0:100])
# Now you will use the display network function to display different sets if the MNIST dataset
# The display is saved in the directory under the name weigths.
# Display 10, 50 and 100 datasets
### YOUR CODE HERE ###
# display 10
display_network.display_network(patches[:, 0:10], 'weights10.png')
# display 50
display_network.display_network(patches[:, 0:50], 'weights50.png')
# display 100
display_network.display_network(patches[:, 0:100], 'weights100.png')
# Obtain random parameters theta
# You need to implement utils_hw.initialize in utils_hw.py
print num_grad, grad
# Compare numerically computed gradients with the ones obtained from backpropagation
diff = np.linalg.norm(num_grad - grad) / np.linalg.norm(num_grad + grad)
print diff
print "Norm of the difference between numerical and analytical num_grad (should be < 1e-9)\n\n"
##======================================================================
## STEP 4: After verifying that your implementation of
# sparseAutoencoderCost is correct, You can start training your sparse
# autoencoder with minFunc (L-BFGS).
# Randomly initialize the parameters
theta = sparse_autoencoder.initialize(hidden_size, visible_size)
J = lambda x: sparse_autoencoder.sparse_autoencoder_cost(x, visible_size, hidden_size,
lambda_, sparsity_param,
beta, patches)
options_ = {'maxiter': 400, 'disp': True}
result = scipy.optimize.minimize(J, theta, method='L-BFGS-B', jac=True, options=options_)
opt_theta = result.x
print result
##======================================================================
## STEP 5: Visualization
W1 = opt_theta[0:hidden_size * visible_size].reshape(hidden_size, visible_size).transpose()
display_network.display_network(W1)
num_samples = 10000
patches, _ = sample_images.get_mnist_data('../data/mnist.pkl.gz',
train=True,
num_examples=num_samples)
# set up L-BFGS args
theta = sparse_autoencoder.initialize_params(hidden_size, visible_size)
sae_cost = partial(sparse_autoencoder.cost,
visible_size=visible_size,
hidden_size=hidden_size,
weight_decay = weight_decay,
beta=beta,
sparsity_param=sparsity_param,
data=patches)
# Train!
trained, cost, d = scipy.optimize.lbfgsb.fmin_l_bfgs_b(sae_cost, theta,
maxfun=max_iter,
m=100,
factr=1.0,
pgtol=1e-100,
iprint=1)
# Save the trained weights
W1, W2, b1, b2 = sparse_autoencoder.unflatten_params(trained,
hidden_size,
visible_size)
display_network.display_network('mnist-weights.png', W1.T)
def test_mnist():
'''Test the convolutional autoencder using MNIST.'''
# %%
# load MNIST as before
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
mean_img = np.mean(mnist.train.images, axis=0)
learning_rate = 0.001
batch_size = 1000
n_epochs = 10000
n_filter_size = 14
n_reload_per_epochs = 100
n_display_per_epochs = 1000
input_shape = [None, 28, 28, 1]
xae_layers = [
{
'n_channels': 144,
'reconstructive_regularizer': 1.0,
'weight_decay': 1.0,
'sparse_regularizer': 1.0,
'sparsity_level': 0.05,
'exclusive_regularizer': 1.0,
'tied_weight': True,
'conv_size': n_filter_size,
'conv_stride': 2,
'conv_padding': 'VALID',
'pool_size': 8,
'pool_stride': 1,
'pool_padding': 'VALID',
'corrupt_prob': 1.0,
'encode': 'lrelu',
'decode': 'linear',
'pathways': [
range(0, 96),
range(48, 144),
],
},
# {
# 'n_channels': 256,
# 'reconstructive_regularizer': 1.0,
# 'weight_decay': 1.0,
# 'sparse_regularizer': 1.0,
# 'sparsity_level': 0.05,
# 'exclusive_regularizer': 1.0,
# 'tied_weight': True,
# 'conv_size': 9,
# 'conv_stride': 1,
# 'conv_padding': 'VALID',
# # 'pool_size': 9,
# # 'pool_stride': 1,
# # 'pool_padding': 'VALID',
# 'corrupt_prob': 1.0,
# 'encode':'sigmoid', 'decode':'linear',
# 'pathways': [
# range(0, 96),
# range(48, 144),
# ],
# },
]
ae = eXclusiveConvolutionalAutoencoder(
input_shape=input_shape,
layers=xae_layers,
init_encoder_weight=None,
init_decoder_weight=None,
init_encoder_bias=None,
init_decoder_bias=None,
)
# %%
optimizer_list = []
for layer_i in range(len(xae_layers)):
optimizer_list.append(
tf.train.AdamOptimizer(learning_rate).minimize(
ae['layerwise_cost'][layer_i]['total'],
var_list=[
ae['encoder_weight'][layer_i],
ae['encoder_bias'][layer_i],
# ae['decoder_weight'][layer_i],
# ae['decoder_bias'][layer_i],
]))
optimizer_list.append(
tf.train.AdamOptimizer(learning_rate).minimize(ae['cost']['total']))
# %%
# We create a session to use the graph
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
# %%
# Fit all training data
for layer_i, (optimizer) in enumerate(optimizer_list):
for epoch_i in range(n_epochs):
if (epoch_i) % n_reload_per_epochs == 0:
batch_xs, batch_ys = mnist.train.next_batch(batch_size)
train = []
train.append(
np.array([
img.reshape((28, 28, 1)) - mean_img.reshape(
(28, 28, 1))
for img in batch_xs[np.where(
np.any(
np.array([
batch_ys[:, 0],
batch_ys[:, 1],
# batch_ys[:, 2],
# batch_ys[:, 3],
# batch_ys[:, 4],
# batch_ys[:, 5],
# batch_ys[:, 6],
# batch_ys[:, 7],
# batch_ys[:, 8],
# batch_ys[:, 9],
]) == 1,
axis=0))]
]))
train.append(
np.array([
img.reshape((28, 28, 1)) - mean_img.reshape(
(28, 28, 1))
for img in batch_xs[np.where(
np.any(
np.array([
# batch_ys[:, 0],
batch_ys[:, 1],
batch_ys[:, 2],
# batch_ys[:, 3],
# batch_ys[:, 4],
# batch_ys[:, 5],
# batch_ys[:, 6],
# batch_ys[:, 7],
# batch_ys[:, 8],
# batch_ys[:, 9],
]) == 1,
axis=0))]
]))
# sess.run(optimizer, feed_dict={ae['training_x'][0]: train[0], })
sess.run(optimizer,
feed_dict={
ae['training_x'][0]: train[0],
ae['training_x'][1]: train[1],
})
if (epoch_i + 1) % n_display_per_epochs == 0:
data_dict = {
ae['training_x'][0]: train[0],
ae['training_x'][1]: train[1],
}
if optimizer is optimizer_list[-1]:
cost_total = sess.run(ae['cost']['total'],
feed_dict=data_dict)
cost_reconstruction_error = sess.run(
ae['cost']['reconstruction_error'],
feed_dict=data_dict)
cost_sparsity = sess.run(ae['cost']['sparsity'],
feed_dict=data_dict)
cost_exclusivity = sess.run(ae['cost']['exclusivity'],
feed_dict=data_dict)
cost_weight_decay = sess.run(ae['cost']['weight_decay'],
feed_dict=data_dict)
else:
cost_total = sess.run(
ae['layerwise_cost'][layer_i]['total'],
feed_dict=data_dict)
cost_reconstruction_error = sess.run(
ae['layerwise_cost'][layer_i]['reconstruction_error'],
feed_dict=data_dict)
cost_sparsity = sess.run(
ae['layerwise_cost'][layer_i]['sparsity'],
feed_dict=data_dict)
cost_exclusivity = sess.run(
ae['layerwise_cost'][layer_i]['exclusivity'],
feed_dict=data_dict)
cost_weight_decay = sess.run(
ae['layerwise_cost'][layer_i]['weight_decay'],
feed_dict=data_dict)
print(
'layer:{}, epoch:{:5d}, cost:{:.6f}, error: {:.6f}, sparsity: {:.6f}, exclusivity: {:.6f}, weight decay: {:.6f}'
.format(
optimizer is optimizer_list[-1] and 'A' or layer_i + 1,
epoch_i + 1, cost_total, cost_reconstruction_error,
cost_sparsity, cost_weight_decay, cost_exclusivity))
n_examples = 5120
test_xs, test_ys = mnist.test.next_batch(n_examples)
test_xs = np.array([
img.reshape((28, 28, 1)) - mean_img.reshape((28, 28, 1))
for img in test_xs[np.where(
np.any(
np.array([
test_ys[:, 0],
test_ys[:, 1],
test_ys[:, 2],
# test_ys[:, 3],
# test_ys[:, 4],
# test_ys[:, 5],
# test_ys[:, 6],
# test_ys[:, 7],
# test_ys[:, 8],
# test_ys[:, 9],
]) == 1,
axis=0))][:256]
])
if optimizer is optimizer_list[-1]:
recon = sess.run(ae['y'], feed_dict={ae['x']: test_xs})
else:
recon = sess.run(ae['layerwise_y'][layer_i],
feed_dict={ae['x']: test_xs})
weights = sess.run(ae['encoder_weight'][0])
weights = weights.transpose((3, 0, 1, 2)).reshape(
(144, n_filter_size * n_filter_size)).transpose()
display_network(weights, filename='mnist_weight.png')
display_network(test_xs.reshape((256, 784)).transpose(),
filename='mnist_test.png')
display_network(recon.reshape((256, 784)).transpose(),
filename='mnist_results.png')
import functools
import numpy as np
import pca
import sample_images
import display_network
if __name__ == '__main__':
num_samples = 10000
num_samples = 10000
m = sample_images.load_matlab_images('IMAGES_RAW.mat')
patches = sample_images.sample(m, num_samples, size=(12, 12), norm=None)
display_network.display_network('raw-patches.png', patches)
# ensure that patches have zero mean
mean = np.mean(patches, axis=0)
patches -= mean
assert np.allclose(np.mean(patches, axis=0), np.zeros(patches.shape[1]))
U, s, x_rot = pca.pca(patches)
covar = pca.covariance(x_rot)
display_network.array_to_file('covariance.png', covar)
# percentage of variance
# cumulative sum
pov = np.array(
functools.reduce(lambda t, l: t + [t[-1] + l]
def train():
# STEP 0: Here we provide the relevant parameters values that will
# allow your sparse autoencoder to get good filters; you do not need to
# change the parameters below.
patch_size = 8
num_patches = 10000
visible_size = patch_size ** 2 # number of input units
hidden_size = 25 # number of hidden units
sparsity_param = 0.01 # desired average activation of the hidden units.
# (This was denoted by the Greek alphabet rho, which looks like a lower-case "p",
# in the lecture notes).
decay_lambda = 0.0001 # weight decay parameter
beta = 3 # weight of sparsity penalty term
# STEP 1: Implement sampleIMAGES
# After implementing sampleIMAGES, the display_network command should
# display a random sample of 200 patches from the dataset
patches = sample_images(patch_size, num_patches)
# list = [randint(0, patches.shape[0]-1) for i in xrange(64)]
# display_network(patches[list, :], 8)
# Obtain random parameters theta
# theta = initialize_parameters(visible_size, hidden_size)
# STEP 2: Implement sparseAutoencoderCost
#
# You can implement all of the components (squared error cost, weight decay term,
# sparsity penalty) in the cost function at once, but it may be easier to do
# it step-by-step and run gradient checking (see STEP 3) after each step. We
# suggest implementing the sparseAutoencoderCost function using the following steps:
#
# (a) Implement forward propagation in your neural network, and implement the
# squared error term of the cost function. Implement backpropagation to
# compute the derivatives. Then (using lambda=beta=0), run Gradient Checking
# to verify that the calculations corresponding to the squared error cost
# term are correct.
#
# (b) Add in the weight decay term (in both the cost function and the derivative
# calculations), then re-run Gradient Checking to verify correctness.
#
# (c) Add in the sparsity penalty term, then re-run Gradient Checking to
# verify correctness.
#
# Feel free to change the training settings when debugging your
# code. (For example, reducing the training set size or
# number of hidden units may make your code run faster; and setting beta
# and/or lambda to zero may be helpful for debugging.) However, in your
# final submission of the visualized weights, please use parameters we
# gave in Step 0 above.
# cost, grad = sparse_autoencoder_cost_and_grad(theta, visible_size, hidden_size,
# decay_lambda, sparsity_param, beta, patches)
# STEP 3: Gradient Checking
#
# Hint: If you are debugging your code, performing gradient checking on smaller models
# and smaller training sets (e.g., using only 10 training examples and 1-2 hidden
# units) may speed things up.
# First, lets make sure your numerical gradient computation is correct for a
# simple function. After you have implemented compute_numerical_gradient,
# run the following:
# check_numerical_gradient()
# Now we can use it to check your cost function and derivative calculations
# for the sparse autoencoder.
# func = lambda x: sparse_autoencoder_cost(x, visible_size, hidden_size,
# decay_lambda, sparsity_param, beta, patches)
# numgrad = compute_numerical_gradient(func, theta)
# Use this to visually compare the gradients side by side
# print numgrad, grad
# Compare numerically computed gradients with the ones obtained from backpropagation
# diff = np.linalg.norm(numgrad-grad)/np.linalg.norm(numgrad+grad)
# Should be small. In our implementation, these values are usually less than 1e-9.
# print diff
# STEP 4: After verifying that your implementation of
# sparse_autoencoder_cost is correct, You can start training your sparse
# autoencoder with minFunc (L-BFGS).
# Randomly initialize the parameters
# Use minimize interface, and set jac=True, so it can accept cost and grad together
theta = initialize_parameters(visible_size, hidden_size)
func_args = (visible_size, hidden_size, decay_lambda, sparsity_param, beta, patches)
res = minimize(sparse_autoencoder_cost_and_grad, x0=theta, args=func_args, method='L-BFGS-B',
jac=True, options={'maxiter': 400, 'disp': True})
# STEP 5: Visualization
w1 = res.x[0: hidden_size * visible_size].reshape((visible_size, hidden_size))
display_network(w1.T, 5, save_figure_path='../data/sparse_autoencoder.png')
visibleSize = 8*8 # number of input units
hiddenSize = 25 # number of hidden units
sparsityParam = 0.01 # desired average activation of the hidden units.
# (This was denoted by the Greek alphabet rho, which looks like a lower-case "p",
# in the lecture notes).
lambdaval = 0.0001 # weight decay parameter
beta = 3 # weight of sparsity penalty term
##======================================================================
## STEP 1: Implement sampleIMAGES
#
# After implementing sampleIMAGES, the display_network command should
# display a random sample of 200 patches from the dataset
patches = sampleIMAGES()
display_network(patches[:, randi(size(patches, 2), 200, 1).squeeze() ])
plt.show()
pdb.set_trace()
# Obtain random parameters theta
theta = initializeParameters(hiddenSize, visibleSize)
##======================================================================
## STEP 2: Implement sparseAutoencoderCost
#
# You can implement all of the components (squared error cost, weight decay term,
# sparsity penalty) in the cost function at once, but it may be easier to do
# it step-by-step and run gradient checking (see STEP 3) after each step. We
# suggest implementing the sparseAutoencoderCost function using the following steps:
#
# (a) Implement forward propagation in your neural network, and implement the
a boolean (True if use only train/validation sets,
False for test set),
and a num_samples integer to choose only the first N samples.
Only reads mnist data from a special pickled mnist file.
Returns array of images with shape (784, num_images).
"""
training, valid, testing = read_mnist_file(filename)
if train:
t = np.array([e for e, l in izip(training[0], training[1]) if test(l)])
v = np.array([e for e, l in izip(valid[0], valid[1]) if test(l)])
images = np.vstack([t, v]).T
tl = np.array(
[l for e, l in izip(training[0], training[1]) if test(l)])
vl = np.array([l for e, l in izip(valid[0], valid[1]) if test(l)])
labels = np.hstack([tl, vl])
else:
t = testing
images = np.array([e for e, l in izip(t[0], t[1]) if test(l)]).T
labels = np.array([l for e, l in izip(t[0], t[1]) if test(l)])
assert images.shape[1] == len(labels)
assert images.shape[0] == 784
patches = images[:, :num_samples]
labels = labels[:num_samples]
assert patches.shape[0] == 784
return patches, labels
if __name__ == '__main__':
train, valid, test = read_mnist_file('data/mnist.pkl.gz')
display_network.display_network('mnist.png', train[0].T[:, :100])
def train():
# STEP 0: Here we provide the relevant parameters values that will
# allow your sparse autoencoder to get good filters; you do not need to
# change the parameters below.
patch_size = 8
num_patches = 10000
visible_size = patch_size**2 # number of input units
hidden_size = 25 # number of hidden units
sparsity_param = 0.01 # desired average activation of the hidden units.
# (This was denoted by the Greek alphabet rho, which looks like a lower-case "p",
# in the lecture notes).
decay_lambda = 0.0001 # weight decay parameter
beta = 3 # weight of sparsity penalty term
# STEP 1: Implement sampleIMAGES
# After implementing sampleIMAGES, the display_network command should
# display a random sample of 200 patches from the dataset
patches = sample_images(patch_size, num_patches)
# list = [randint(0, patches.shape[0]-1) for i in xrange(64)]
# display_network(patches[list, :], 8)
# Obtain random parameters theta
# theta = initialize_parameters(visible_size, hidden_size)
# STEP 2: Implement sparseAutoencoderCost
#
# You can implement all of the components (squared error cost, weight decay term,
# sparsity penalty) in the cost function at once, but it may be easier to do
# it step-by-step and run gradient checking (see STEP 3) after each step. We
# suggest implementing the sparseAutoencoderCost function using the following steps:
#
# (a) Implement forward propagation in your neural network, and implement the
# squared error term of the cost function. Implement backpropagation to
# compute the derivatives. Then (using lambda=beta=0), run Gradient Checking
# to verify that the calculations corresponding to the squared error cost
# term are correct.
#
# (b) Add in the weight decay term (in both the cost function and the derivative
# calculations), then re-run Gradient Checking to verify correctness.
#
# (c) Add in the sparsity penalty term, then re-run Gradient Checking to
# verify correctness.
#
# Feel free to change the training settings when debugging your
# code. (For example, reducing the training set size or
# number of hidden units may make your code run faster; and setting beta
# and/or lambda to zero may be helpful for debugging.) However, in your
# final submission of the visualized weights, please use parameters we
# gave in Step 0 above.
# cost, grad = sparse_autoencoder_cost_and_grad(theta, visible_size, hidden_size,
# decay_lambda, sparsity_param, beta, patches)
# STEP 3: Gradient Checking
#
# Hint: If you are debugging your code, performing gradient checking on smaller models
# and smaller training sets (e.g., using only 10 training examples and 1-2 hidden
# units) may speed things up.
# First, lets make sure your numerical gradient computation is correct for a
# simple function. After you have implemented compute_numerical_gradient,
# run the following:
# check_numerical_gradient()
# Now we can use it to check your cost function and derivative calculations
# for the sparse autoencoder.
# func = lambda x: sparse_autoencoder_cost(x, visible_size, hidden_size,
# decay_lambda, sparsity_param, beta, patches)
# numgrad = compute_numerical_gradient(func, theta)
# Use this to visually compare the gradients side by side
# print numgrad, grad
# Compare numerically computed gradients with the ones obtained from backpropagation
# diff = np.linalg.norm(numgrad-grad)/np.linalg.norm(numgrad+grad)
# Should be small. In our implementation, these values are usually less than 1e-9.
# print diff
# STEP 4: After verifying that your implementation of
# sparse_autoencoder_cost is correct, You can start training your sparse
# autoencoder with minFunc (L-BFGS).
# Randomly initialize the parameters
# Use minimize interface, and set jac=True, so it can accept cost and grad together
theta = initialize_parameters(visible_size, hidden_size)
func_args = (visible_size, hidden_size, decay_lambda, sparsity_param, beta,
patches)
res = minimize(sparse_autoencoder_cost_and_grad,
x0=theta,
args=func_args,
method='L-BFGS-B',
jac=True,
options={
'maxiter': 400,
'disp': True
})
# STEP 5: Visualization
w1 = res.x[0:hidden_size * visible_size].reshape(
(visible_size, hidden_size))
display_network(w1.T, 5, save_figure_path='../data/sparse_autoencoder.png')
def __init__(self, scenario):
self.readScenario(scenario)
self.mLinkInsertMaxSpeedCapacity() # add max speed and capacity
C = self.mLink2WeightedAdjacency(self.mLink, fieldNo=6) # capacity
U = self.mLink2WeightedAdjacency(self.mLink, fieldNo=5) # max Speed
L = self.mLink2WeightedAdjacency(self.mLink,
fieldNo=3) # link distance
S = ifn.capacity2stochastic(C) # Markov stochastic
# first try at arbitrary kappa=100
pi = ifn.steadyStateMC(S, kappa=100) # node values
F = ifn.idealFlow(S, pi) # ideal flow
G = self.HadamardDivision(F, C) # congestion
maxCongestion = np.max(G)
if self.calibrationBasis == "flow":
# calibrate with new kappa to reach totalFlow
kappa = self.totalFlow
else: # calibrationBasis=="congestion"
# calibrate with new kappa to reach max congestion level
kappa = 100 * float(
self.maxAllowableCongestion) / maxCongestion # total flow
# compute ideal flow and congestion
pi = ifn.steadyStateMC(S, kappa) # node values
F = ifn.idealFlow(S, pi) # scaled ideal flow
G = self.HadamardDivision(F, C) # congestion
maxCongestion = np.max(G)
# compute link performances
self.mLink = self.addFlow2mLink(self.mLink, F) # fieldNo=7 flow
self.mLink = self.addFlow2mLink(self.mLink,
G) # fieldNo=8 congestion level
self.mLink = self.computeLinkPerformance(
self.mLink, self.travelTimeModel,
self.cloudNode) # fieldNo=9 to 11
# save output mLink
mR, mC = self.mLink.shape
fmt = "%d,%d,%d,%0.3f,%d,%0.3f,%d,%0.3f,%0.3f,%0.3f,%0.3f,%0.3f"
header = "LinkNo,Node1,Node2,Dist,Lanes,MaxSpeed,Capacity,Flow,Congestion,Speed,TravelTime,Delay"
mLink2 = self.mLink.T
with open(self.folder + self.scenarioName + '.csv', 'w') as fh:
for j in range(mC):
col = mLink2[j, :]
if j == 0:
np.savetxt(fh,
col.reshape(1, -1),
fmt=fmt,
header=header,
delimiter=',')
else:
np.savetxt(fh, col.reshape(1, -1), fmt=fmt, delimiter=',')
# network performance
avgSpeed = np.nanmean(self.mLink[9, :])
avgTravelTime = np.nanmean(self.mLink[10, :])
avgDelay = np.nanmean(self.mLink[11, :])
avgDist = np.nanmean(self.mLink[3, :])
# save network performance
with open(self.folder + self.scenarioName + '.net', 'w') as fh:
fh.write("totalFlow=" + str(kappa) + "\n") # in pcu/hour
fh.write("maxCongestion=" + str(maxCongestion) + "\n")
fh.write("avgSpeed=" + str(avgSpeed) + "\n") # in km/hour
fh.write("avgTravelTime=" + str(avgTravelTime) + "\n") # in hour
fh.write("avgDelay=" + str(avgDelay) + "\n") # in hour
fh.write("avgDist=" + str(avgDist) + "\n") # in meter
# report
print(self.scenarioName)
print("Network performance:")
print("\tTotal Flow = ", round(kappa, 2), " pcu/hour")
print("\tMax Congestion = ", round(maxCongestion, 4))
print("\tAvg Link Speed =", round(avgSpeed, 4), " km/hour")
print("\tAvg Link Travel Time = ",
round(1000 * 60 * avgTravelTime / avgDist, 4), " min/km")
print("\tAvg Link Delay = ", round(1000 * 3600 * avgDelay / avgDist,
4), " seconds/km")
print("Basis:")
print("\tAvg Link Distance = ", round(avgDist, 4), " m/link")
print("\tAvg Link Travel Time = ", round(3600 * avgTravelTime, 4),
" seconds/link")
print("\tAvg Link Delay = ", round(3600 * avgDelay, 4),
" seconds/link")
arrThreshold = [0.8, 0.9, 1]
plt = dn.display_network(self.mLink.T, 8, self.mNode.T, arrThreshold,
9) # display congestion
plt.show()
a function that accepts one label argument (e.g. only get digits 5-9),
a boolean (True if use only train/validation sets,
False for test set),
and a num_samples integer to choose only the first N samples.
Only reads mnist data from a special pickled mnist file.
Returns array of images with shape (784, num_images).
"""
training, valid, testing = read_mnist_file(filename)
if train:
t = np.array([e for e, l in izip(training[0], training[1]) if test(l)])
v = np.array([e for e, l in izip(valid[0], valid[1]) if test(l)])
images = np.vstack([t, v]).T
tl = np.array([l for e, l in izip(training[0], training[1]) if test(l)])
vl = np.array([l for e, l in izip(valid[0], valid[1]) if test(l)])
labels = np.hstack([tl, vl])
else:
t = testing
images = np.array([e for e, l in izip(t[0], t[1]) if test(l)]).T
labels = np.array([l for e, l in izip(t[0], t[1]) if test(l)])
assert images.shape[1] == len(labels)
assert images.shape[0] == 784
patches = images[:, :num_samples]
labels = labels[:num_samples]
assert patches.shape[0] == 784
return patches, labels
if __name__ == "__main__":
train, valid, test = read_mnist_file("data/mnist.pkl.gz")
display_network.display_network("mnist.png", train[0].T[:, :100])
