Skip to content

Commit

Permalink
save
Browse files Browse the repository at this point in the history
  • Loading branch information
@matthieu637
matthieu637 committed Apr 26, 2012
1 parent 678b15d commit a9290b0
Show file tree
Hide file tree
Showing 7 changed files with 109 additions and 87 deletions.
50 changes: 32 additions & 18 deletions src/articles/digit_reco/discretize.py
View file
Expand Up @@ -10,7 +10,7 @@
from multilayerp import MultilayerPerceptron from multilayerp import MultilayerPerceptron
from random import shuffle from random import shuffle
import matplotlib.pyplot as plt import matplotlib.pyplot as plt
from data import DataFile from data import DataFile, DataFileR
from mpl_toolkits.mplot3d import Axes3D from mpl_toolkits.mplot3d import Axes3D
from utils import index_max from utils import index_max


Expand All @@ -34,39 +34,50 @@ def discretis(ll, nbDiscretized=nbDiscre):
nbr_network = 1 nbr_network = 1
momentum = 0.9 momentum = 0.9
lrate = 0.1 lrate = 0.1
nbShape = 10 nbShape = 3
nbEpoch = 1000 nbEpoch = 1000
nbTry = 10
display_interval = range(nbEpoch)[::6] display_interval = range(nbEpoch)[::6]




#create inputs/outputs to learn
# examples = DataFile("digit_shape.txt", mode)

examples = DataFileR("iris.txt")
momentum = 0.2

nbInputs = len(examples.inputs[0])
nbHidden = 5
nbOutputs = len(examples.outputs[0])
nbShape = nbOutputs

#create all networks #create all networks
networks = [{} for _ in range(nbr_network)] networks = [{} for _ in range(nbr_network)]


for i in range(nbr_network): for i in range(nbr_network):
first_order = MultilayerPerceptron(20, 5, 10, learning_rate=lrate, momentum=momentum, grid=mode) first_order = MultilayerPerceptron(nbInputs, nbHidden, nbOutputs, learning_rate=lrate, momentum=momentum, grid=mode)
high_order_10 = MultilayerPerceptron(5, 10, 35, learning_rate=lrate, momentum=momentum, grid=mode) high_order_10 = MultilayerPerceptron(nbHidden, nbHidden*2, nbInputs+nbHidden+nbOutputs, learning_rate=lrate, momentum=momentum, grid=mode)


networks[i] = {'first_order' : first_order, networks[i] = {'first_order' : first_order,
'high_order_10' : high_order_10} 'high_order_10' : high_order_10}


#create inputs/outputs to learn
examples = DataFile("digit_shape.txt", mode)




learned = [[0 for _ in range(10)] for _ in range(nbDiscre**5)]
learned = [[0 for _ in range(nbShape)] for _ in range(nbDiscre**nbHidden)]


#3 curves #3 curves
dis = [[0 for _ in range(nbEpoch)] for _ in range(nbShape)] dis = [[0 for _ in range(nbEpoch)] for _ in range(nbShape)]
dis2 = [[0 for _ in range(nbEpoch)] for _ in range(nbShape)] dis2 = [[0 for _ in range(nbEpoch)] for _ in range(nbShape)]
div = [[0 for _ in range(nbEpoch)] for _ in range(nbShape)] div = [[0 for _ in range(nbEpoch)] for _ in range(nbShape)]
valid = [[] for _ in range(10)] valid = [[] for _ in range(nbShape)]


#learning #learning
for epoch in range(nbEpoch): for epoch in range(nbEpoch):
for network in networks: for network in networks:
l_exx = list(range(nbShape)) l_exx = list(range(len(examples.inputs)))
shuffle(l_exx) shuffle(l_exx)
for ex in l_exx: for ex in l_exx[0:nbTry]:
#RMS #RMS
network['first_order'].calc_output(examples.inputs[ex]) network['first_order'].calc_output(examples.inputs[ex])


Expand All @@ -76,9 +87,12 @@ def discretis(ll, nbDiscretized=nbDiscre):


network['high_order_10'].calc_output(network['first_order'].stateHiddenNeurons) network['high_order_10'].calc_output(network['first_order'].stateHiddenNeurons)


learned[discretis(network['first_order'].stateHiddenNeurons)][ex] += 1


im = index_max(examples.outputs[ex]) im = index_max(examples.outputs[ex])

learned[discretis(network['first_order'].stateHiddenNeurons)][im] += 1

div[im][epoch] += 1 div[im][epoch] += 1
dis[im][epoch] += discretis(network['first_order'].stateHiddenNeurons) dis[im][epoch] += discretis(network['first_order'].stateHiddenNeurons)
dis2[im][epoch] += index_max(network['first_order'].stateOutputNeurons) dis2[im][epoch] += index_max(network['first_order'].stateOutputNeurons)
Expand All @@ -97,7 +111,7 @@ def discretis(ll, nbDiscretized=nbDiscre):
print(epoch) print(epoch)


#divided by the number of networks #divided by the number of networks
for i in range(10): for i in range(nbShape):
for j in range(nbEpoch): for j in range(nbEpoch):
if(div[i][j] != 0): if(div[i][j] != 0):
dis[i][j] /= div[i][j] dis[i][j] /= div[i][j]
Expand All @@ -107,7 +121,7 @@ def discretis(ll, nbDiscretized=nbDiscre):


fig = plt.figure() fig = plt.figure()
ax = fig.add_subplot(111, projection='3d') ax = fig.add_subplot(111, projection='3d')
for j in range(10): for j in range(nbShape):
ax.scatter([dis[j][k] for k in valid[j]], [j] * len(valid[j]), valid[j], color=colors[j], marker='x') ax.scatter([dis[j][k] for k in valid[j]], [j] * len(valid[j]), valid[j], color=colors[j], marker='x')


ax.set_xlabel('DISCRETIZED VALUE') ax.set_xlabel('DISCRETIZED VALUE')
Expand All @@ -118,7 +132,7 @@ def discretis(ll, nbDiscretized=nbDiscre):


fig = plt.figure() fig = plt.figure()
ax = fig.add_subplot(111, projection='3d') ax = fig.add_subplot(111, projection='3d')
for j in range(10): for j in range(nbShape):
ax.scatter([dis2[j][k] for k in valid[j]], [j] * len(valid[j]), valid[j], color=colors[j], marker='x') ax.scatter([dis2[j][k] for k in valid[j]], [j] * len(valid[j]), valid[j], color=colors[j], marker='x')


ax.set_xlabel('DISCRETIZED VALUE') ax.set_xlabel('DISCRETIZED VALUE')
Expand Down Expand Up @@ -147,19 +161,19 @@ def discretis(ll, nbDiscretized=nbDiscre):
for j in range(nbShape): for j in range(nbShape):
plt.title('shape :%d' % j) plt.title('shape :%d' % j)
plt.plot(valid[j], [dis[j][k] for k in valid[j]], '.', label="hidden") plt.plot(valid[j], [dis[j][k] for k in valid[j]], '.', label="hidden")
plt.plot(valid[j], [dis2[j][k]*(nbDiscre**5)/10 for k in valid[j]], '.', label="output") plt.plot(valid[j], [dis2[j][k]*(nbDiscre**nbHidden)/nbShape for k in valid[j]], '.', label="output")


plt.legend(loc='best', frameon=False) plt.legend(loc='best', frameon=False)
plt.show() plt.show()


stade = 0 stade = 0
for i in range(len(learned)): for i in range(len(learned)):
r = max(learned[i]) r = max(learned[i])
if(r > 10): if(r > nbShape):
cl = list(learned[i]) cl = list(learned[i])
cl.remove(r) cl.remove(r)
if(max(cl) > 10): if(max(cl) > nbShape):
stade += 1 stade += 1
plt.bar(range(stade*12+10)[stade*12::], learned[i]) plt.bar(range(stade*12+nbShape)[stade*12::], learned[i])
plt.show() plt.show()


38 changes: 23 additions & 15 deletions src/articles/digit_reco/handwritten_main.py
View file
Expand Up @@ -10,12 +10,12 @@
from multilayerp import MultilayerPerceptron from multilayerp import MultilayerPerceptron
from random import shuffle from random import shuffle
import matplotlib.pyplot as plt import matplotlib.pyplot as plt
from data import DataFile from data import DataFile, DataFileR
from utils import index_max, compare, compare_f from utils import index_max, compare, compare_f


if __name__ == '__main__': if __name__ == '__main__':
mode = MultilayerPerceptron.R0to1 mode = MultilayerPerceptron.R0to1
nbr_network = 3 nbr_network = 5
momentum = 0.9 momentum = 0.9
lrate = 0.1 lrate = 0.1
nbEpoch = 1000 nbEpoch = 1000
Expand All @@ -24,22 +24,30 @@
display_interval2 = [0, 25, 50, 100, 200, 500, 999] display_interval2 = [0, 25, 50, 100, 200, 500, 999]




#create inputs/outputs to learn
# examples = DataFile("digit_shape_16.txt")
# examples = DataFile("digit_handwritten_16.txt")
examples = DataFileR("iris.txt")

nbInput = len(examples.inputs[0])
nbHidden = 16 * 4
nbHidden = 5
nbOutput = len(examples.outputs[0])

#create all networks #create all networks
networks = [{} for _ in range(nbr_network)] networks = [{} for _ in range(nbr_network)]


for i in range(nbr_network): for i in range(nbr_network):
first_order = MultilayerPerceptron(16 * 16, 16 * 4, 10, learning_rate=lrate, momentum=momentum, grid=mode) first_order = MultilayerPerceptron(nbInput, nbHidden, nbOutput, learning_rate=lrate, momentum=momentum, grid=mode)
high_order_10 = MultilayerPerceptron(16 * 4, 16 * 4 * 2, 16 * 16 + 16 * 4 + 10, learning_rate=lrate, momentum=momentum, grid=mode) high_order_10 = MultilayerPerceptron(nbHidden, nbHidden * 2, nbInput + nbHidden + nbOutput, learning_rate=lrate, momentum=momentum, grid=mode)
high_order_5 = MultilayerPerceptron(16 * 4, 16 * 4, 16 * 16 + 16 * 4 + 10, learning_rate=lrate, momentum=momentum, grid=mode) high_order_5 = MultilayerPerceptron(nbHidden, nbHidden, nbInput + nbHidden + nbOutput, learning_rate=lrate, momentum=momentum, grid=mode)




networks[i] = {'first_order' : first_order, networks[i] = {'first_order' : first_order,
'high_order_10' : high_order_10, 'high_order_10' : high_order_10,
'high_order_5':high_order_5} 'high_order_5':high_order_5}


#create inputs/outputs to learn
# examples = DataFile("digit_shape_16.txt")
examples = DataFile("digit_handwritten_16.txt")


#3 curves #3 curves
rms_plot = {'first_order' : [] , rms_plot = {'first_order' : [] ,
Expand Down Expand Up @@ -83,11 +91,11 @@
if(index_max(network['first_order'].stateOutputNeurons) != index_max(examples.outputs[ex])): if(index_max(network['first_order'].stateOutputNeurons) != index_max(examples.outputs[ex])):
err_one_network['first_order'] += 1 err_one_network['first_order'] += 1


err_one_network['high_order_20'] += 1 - compare(examples.inputs[ex], network['high_order_10'].stateOutputNeurons[0:16 * 16]) err_one_network['high_order_20'] += 1 - compare(examples.inputs[ex], network['high_order_10'].stateOutputNeurons[0:nbInput])
if( not compare_f(network['first_order'].stateHiddenNeurons, if(not compare_f(network['first_order'].stateHiddenNeurons,
network['high_order_10'].stateOutputNeurons[16 * 16:16 * 16 + 16 * 4], 0.3) ): network['high_order_10'].stateOutputNeurons[nbInput:nbInput + nbHidden], 0.3)):
err_one_network['high_order_5'] += 1 err_one_network['high_order_5'] += 1
if(index_max(network['high_order_10'].stateOutputNeurons[16 * 16 + 16 * 4:16 * 16 + 16 * 4 + 10]) != if(index_max(network['high_order_10'].stateOutputNeurons[nbInput + nbHidden:nbInput + nbHidden + nbOutput]) !=
index_max(network['first_order'].stateOutputNeurons)): index_max(network['first_order'].stateOutputNeurons)):
err_one_network['high_order_10'] += 1 err_one_network['high_order_10'] += 1


Expand All @@ -99,9 +107,9 @@




#add plot #add plot
rms_plot['first_order'].append(sum_rms['first_order']/ (nbTry * nbr_network)) rms_plot['first_order'].append(sum_rms['first_order'] / (nbTry * nbr_network))
rms_plot['high_order_10'].append(sum_rms['high_order_10']/ (nbTry * nbr_network)) rms_plot['high_order_10'].append(sum_rms['high_order_10'] / (nbTry * nbr_network))
rms_plot['high_order_5'].append(sum_rms['high_order_5']/ (nbTry * nbr_network)) rms_plot['high_order_5'].append(sum_rms['high_order_5'] / (nbTry * nbr_network))


err_plot['first_order'].append(err_one_network['first_order'] / (nbTry * nbr_network)) err_plot['first_order'].append(err_one_network['first_order'] / (nbTry * nbr_network))
err_plot['high_order_10'].append(err_one_network['high_order_10'] / (nbTry * nbr_network)) err_plot['high_order_10'].append(err_one_network['high_order_10'] / (nbTry * nbr_network))
Expand Down
46 changes: 20 additions & 26 deletions ...digit_reco/handwritten_main_perceptron.py → src/articles/digit_reco/p_vs_mlp.py
View file
Expand Up @@ -14,31 +14,38 @@
import matplotlib.pyplot as plt import matplotlib.pyplot as plt
from data import DataFile, DataFileR from data import DataFile, DataFileR




if __name__ == '__main__': if __name__ == '__main__':
mode = MultilayerPerceptron.R0to1 mode = MultilayerPerceptron.R0to1
nbr_network = 5 nbr_network = 5
momentum = 0.5 momentum = 0.5
lrate = 0.15 lrate = 0.15
nbEpoch = 600 nbEpoch = 400
nbTry = 20 nbTry = 20
display_interval = range(nbEpoch)[::6] display_interval = range(nbEpoch)[::6]



#create inputs/outputs to learn
examples = DataFileR("iris.txt", mode)
# examples = DataFile("digit_handwritten_16.txt", mode)

nbInputs = len(examples.inputs[0])
nbOutputs = len(examples.outputs[0])


#create all networks #create all networks
networks = [{} for _ in range(nbr_network)] networks = [{} for _ in range(nbr_network)]


for i in range(nbr_network): for i in range(nbr_network):
first_order = [PerceptronR0to1(4, learning_rate=lrate, momentum=momentum, first_order = [PerceptronR0to1(nbInputs, learning_rate=lrate, momentum=momentum,
temperature=1., init_w_randomly=True, enable_bias=True) for _ in range(3)] temperature=1., init_w_randomly=True, enable_bias=True) for _ in range(nbOutputs)]
high_order_10 = MultilayerPerceptron(4, 40, 3, learning_rate=lrate, momentum=momentum, grid=mode) high_order_10 = MultilayerPerceptron(nbInputs, 3+nbInputs//4, nbOutputs, learning_rate=lrate, momentum=momentum, grid=mode)

high_order_10.init_weights_randomly(-1, 1)


networks[i] = {'first_order' : first_order, networks[i] = {'first_order' : first_order,
'high_order_10' : high_order_10} 'high_order_10' : high_order_10}


#create inputs/outputs to learn
examples = DataFileR("iris.txt", mode)
# examples = DataFile("digit_handwritten_16.txt", mode)

#3 curves #3 curves
err_plot = {'first_order' : [] , err_plot = {'first_order' : [] ,
'high_order_10' : []} 'high_order_10' : []}
Expand All @@ -53,32 +60,19 @@
shuffle(l_exx) shuffle(l_exx)
for ex in l_exx[0:nbTry]: for ex in l_exx[0:nbTry]:
resf1 = [network['first_order'][i].calc_output(examples.inputs[ex]) resf1 = [network['first_order'][i].calc_output(examples.inputs[ex])
for i in range(3)] for i in range(nbOutputs)]
network['high_order_10'].calc_output(examples.inputs[ex]) network['high_order_10'].calc_output(examples.inputs[ex])


# ww = []
#
# for i in range(10):
# ww.extend(network['first_order'][i].weights)
#
# resh = [network['high_order_10'][i].calc_output(ww)
# for i in range(10)]

if(index_max(resf1) != index_max(examples.outputs[ex])): if(index_max(resf1) != index_max(examples.outputs[ex])):
err_one_network['first_order'] += 1 err_one_network['first_order'] += 1


if(index_max(network['high_order_10'].stateOutputNeurons) != index_max(examples.outputs[ex])): if(index_max(network['high_order_10'].stateOutputNeurons) != index_max(examples.outputs[ex])):
err_one_network['high_order_10'] += 1 err_one_network['high_order_10'] += 1


# if(index_max(resh) != index_max(resf1)):
# err_one_network['high_order_10'] += 1


#learn
# for i in range(10):
# network['high_order_10'][i].train(ww, resf1[i])
network['high_order_10'].train(examples.inputs[ex], network['high_order_10'].train(examples.inputs[ex],
examples.outputs[ex]) examples.outputs[ex])
for i in range(3): for i in range(nbOutputs):
network['first_order'][i].train(examples.inputs[ex], network['first_order'][i].train(examples.inputs[ex],
examples.outputs[ex][i]) examples.outputs[ex][i])


Expand Down
14 changes: 8 additions & 6 deletions src/articles/digital_reco/feedback/merging.py
View file
Expand Up @@ -156,12 +156,12 @@ def ampli(l, n):
networks = [{} for _ in range(nbr_network)] networks = [{} for _ in range(nbr_network)]


for i in range(nbr_network): for i in range(nbr_network):
control = MultilayerPerceptron(16 * 16, 100, 10, learning_rate=0.15, momentum=momentum, grid=mode, seed(i)
temperature=1, random=False, enable_bias=True) control = MultilayerPerceptron(16 * 16, 100, 10, learning_rate=0.15, momentum=momentum, grid=mode)
control.init_weights_randomly(-1, 1) control.init_weights_randomly(-1, 1)


first_order = AdHock(control)
high_order_h = MultilayerPerceptron(100, 20, 2, learning_rate=0.1, momentum=0., grid=mode) high_order_h = MultilayerPerceptron(100, 20, 2, learning_rate=0.1, momentum=0., grid=mode)
first_order = AdHock(control)
# high_order_h.init_weights_randomly(-1, 1) # high_order_h.init_weights_randomly(-1, 1)


networks[i] = {'first_order' : first_order, networks[i] = {'first_order' : first_order,
Expand All @@ -179,6 +179,8 @@ def ampli(l, n):
'control': [], 'control': [],
'diff': []} 'diff': []}


seed(100)

#learning #learning
for epoch in range(nbEpoch): for epoch in range(nbEpoch):
perfo = {'first_order' : 0. , perfo = {'first_order' : 0. ,
Expand Down Expand Up @@ -240,9 +242,9 @@ def ampli(l, n):
print("score : ", sum(y_perfo['diff']) / len(y_perfo['diff'])) print("score : ", sum(y_perfo['diff']) / len(y_perfo['diff']))


plt.title("Feedback by merging") plt.title("Feedback by merging")
plt.plot(display_interval , y_perfo['first_order'][3::5], label="first-order network", linewidth=1) # plt.plot(display_interval , y_perfo['first_order'][3::5], label="first-order network", linewidth=1)
plt.plot(display_interval , y_perfo['high_order_h'][3::5], label="high-order network (high learning rate)") # plt.plot(display_interval , y_perfo['high_order_h'][3::5], label="high-order network (high learning rate)")
plt.plot(display_interval , y_perfo['wager_proportion'][3::5], label="proportion of high wagers") # plt.plot(display_interval , y_perfo['wager_proportion'][3::5], label="proportion of high wagers")
plt.plot(display_interval , y_perfo['control'][3::5], label="control network", linewidth=2) plt.plot(display_interval , y_perfo['control'][3::5], label="control network", linewidth=2)
plt.plot(display_interval , y_perfo['feedback'][3::5], label="feedback", linewidth=2) plt.plot(display_interval , y_perfo['feedback'][3::5], label="feedback", linewidth=2)
plt.ylabel('SUCCESS RATIO') plt.ylabel('SUCCESS RATIO')
Expand Down
20 changes: 10 additions & 10 deletions src/articles/digital_reco/feedback/nth_winner.py
View file
Expand Up @@ -8,7 +8,7 @@


from multilayerp import MultilayerPerceptron from multilayerp import MultilayerPerceptron
from utils import index_max, index_max_nth from utils import index_max, index_max_nth
from random import shuffle from random import shuffle, seed
import matplotlib.pyplot as plt import matplotlib.pyplot as plt
from data import DataFile from data import DataFile


Expand All @@ -24,13 +24,11 @@
networks = [{} for _ in range(nbr_network)] networks = [{} for _ in range(nbr_network)]


for i in range(nbr_network): for i in range(nbr_network):
seed(i)
first_order = MultilayerPerceptron(16 * 16, 100, 10, learning_rate=0.15, momentum=momentum, grid=mode) first_order = MultilayerPerceptron(16 * 16, 100, 10, learning_rate=0.15, momentum=momentum, grid=mode)
high_order_h = MultilayerPerceptron(100, 100, 10, learning_rate=0.1, momentum=0.5, grid=mode)


first_order.init_weights_randomly(-1, 1) first_order.init_weights_randomly(-1, 1)
high_order_h.init_weights_randomly(-1, 1)


high_order_h = MultilayerPerceptron(100, 100, 10, learning_rate=0.1, momentum=0.5, grid=mode)
networks[i] = {'first_order' : first_order, networks[i] = {'first_order' : first_order,
'high_order_h' : high_order_h} 'high_order_h' : high_order_h}


Expand All @@ -49,6 +47,8 @@
max2 = [0 for _ in range(nbEpoch)] max2 = [0 for _ in range(nbEpoch)]
div = [0 for _ in range(nbEpoch)] div = [0 for _ in range(nbEpoch)]


seed(100)

#learning #learning
for epoch in range(nbEpoch): for epoch in range(nbEpoch):
perfo = {'first_order' : 0. , perfo = {'first_order' : 0. ,
Expand All @@ -74,7 +74,7 @@
if(index_max(network['first_order'].stateOutputNeurons) == index_max(examples.outputs[ex])): if(index_max(network['first_order'].stateOutputNeurons) == index_max(examples.outputs[ex])):
perfo['first_order'] += 1 perfo['first_order'] += 1
max2[epoch] += max(network['first_order'].stateOutputNeurons) max2[epoch] += max(network['first_order'].stateOutputNeurons)
print(max(network['first_order'].stateOutputNeurons)) # print(max(network['first_order'].stateOutputNeurons))
div[epoch] += 1 div[epoch] += 1
perfo['max'] += max(network['first_order'].stateOutputNeurons) perfo['max'] += max(network['first_order'].stateOutputNeurons)


Expand Down Expand Up @@ -172,11 +172,11 @@


plt.title("Performance of first-order and higher-order networks with feedback ( nth W-T-A )") plt.title("Performance of first-order and higher-order networks with feedback ( nth W-T-A )")
plt.plot(display_interval , y_perfo['first_order'][3::5], label="first-order network", linewidth=2) plt.plot(display_interval , y_perfo['first_order'][3::5], label="first-order network", linewidth=2)
plt.plot(display_interval , y_perfo['wager_proportion'][3::5], label="proportion of high wagers") # plt.plot(display_interval , y_perfo['wager_proportion'][3::5], label="proportion of high wagers")
plt.plot(display_interval , y_perfo['feedback'][3::5], label="feedback", linewidth=2) plt.plot(display_interval , y_perfo['feedback'][3::5], label="feedback", linewidth=2)
plt.plot(display_interval , y_perfo['high_order_h'][3::5], label="high-order network (high learning rate)") # plt.plot(display_interval , y_perfo['high_order_h'][3::5], label="high-order network (high learning rate)")
plt.plot(display_interval , y_perfo['max'][3::5], label="most active neuron", linewidth=2) # plt.plot(display_interval , y_perfo['max'][3::5], label="most active neuron", linewidth=2)
plt.plot(display_interval , max2[3::5], label="most active neuron (good answer)", linewidth=2) # plt.plot(display_interval , max2[3::5], label="most active neuron (good answer)", linewidth=2)
plt.ylabel('SUCCESS RATIO') plt.ylabel('SUCCESS RATIO')
plt.xlabel("EPOCHS") plt.xlabel("EPOCHS")
plt.axis((0, nbEpoch, 0, 1.)) plt.axis((0, nbEpoch, 0, 1.))
Expand Down

0 comments on commit a9290b0

Please sign in to comment.