Skip to content

Commit 6b64eb8

Browse files
webengwebeng
committed
various changes
1 parent 0eb5575 commit 6b64eb8

File tree

5 files changed

+3237
-584
lines changed

5 files changed

+3237
-584
lines changed

‎code/best_model.pkl

Lines changed: 1990 additions & 325 deletions
Large diffs are not rendered by default.

‎code/convolutional_mlp_test.py renamed to ‎code/convolutional_mlp_v2.py

Lines changed: 174 additions & 25 deletions
Original file line numberDiff line numberDiff line change
@@ -26,6 +26,7 @@
2626
import timeit
2727

2828
import numpy
29+
import cPickle
2930

3031
import theano
3132
import theano.tensor as T
@@ -34,6 +35,7 @@
3435

3536
from logistic_sgd_test import LogisticRegression, load_data
3637
from mlp import HiddenLayer
38+
from fetex_image import FetexImage
3739

3840

3941
class LeNetConvPoolLayer(object):
@@ -117,9 +119,9 @@ def __init__(self, rng, input, filter_shape, image_shape, poolsize=(2, 2)):
117119
# def evaluate_lenet5(learning_rate=0.1, n_epochs=100,
118120
# dataset='mnist.pkl.gz',
119121
# nkerns=[(96) , (256)], batch_size=300):
120-
def evaluate_lenet5(learning_rate=0.1, n_epochs=100,
122+
def evaluate_lenet5(learning_rate=0.1, n_epochs=2,
121123
dataset='mnist.pkl.gz',
122-
nkerns=[(25 / 1) , (25 / 1)], batch_size=4):
124+
nkerns=[(25 / 1) , (25 / 1)], batch_size=400):
123125

124126
""" Demonstrates lenet on MNIST dataset
125127
@@ -224,14 +226,14 @@ def evaluate_lenet5(learning_rate=0.1, n_epochs=100,
224226
rng,
225227
input=layer2_input,
226228
n_in=nkerns[1] * 13 * 13,
227-
n_out=2,
229+
n_out=400,
228230
activation=T.tanh
229231
)
230232

231233
# classify the values of the fully-connected sigmoidal layer
232234
#layer3 = LogisticRegression(input=layer2.output, n_in=500, n_out=10)
233235
#layer3 = LogisticRegression(input=layer2.output, n_in=500, n_out=2)
234-
layer4 = LogisticRegression(input=layer3.output, n_in=2, n_out=2)
236+
layer4 = LogisticRegression(input=layer3.output, n_in=400, n_out=21)
235237

236238
# the cost we minimize during training is the NLL of the model
237239
#cost = layer3.negative_log_likelihood(y)
@@ -352,9 +354,33 @@ def evaluate_lenet5(learning_rate=0.1, n_epochs=100,
352354
(epoch, minibatch_index + 1, n_train_batches,
353355
test_score * 100.))
354356

357+
output = open('../data/layer0.pkl', 'wb')
358+
cPickle.dump(layer0, output,protocol=-1)
359+
output.close()
360+
361+
output = open('../data/layer1.pkl', 'wb')
362+
cPickle.dump(layer1, output,protocol=-1)
363+
output.close()
364+
365+
# output = open('../data/layer2.pkl', 'wb')
366+
# cPickle.dump(layer2, output,protocol=-1)
367+
# output.close()
368+
369+
output = open('../data/layer3.pkl', 'wb')
370+
cPickle.dump(layer3, output,protocol=-1)
371+
output.close()
372+
373+
output = open('../data/layer4.pkl', 'wb')
374+
cPickle.dump(layer4, output,protocol=-1)
375+
output.close()
376+
355377
# save the best model
356378
# with open('best_model.pkl', 'w') as f:
357-
# cPickle.dump(classifier, f)
379+
# cPickle.dump(layer0, f)
380+
# cPickle.dump(layer1, f)
381+
# cPickle.dump(layer2, f)
382+
# cPickle.dump(layer3, f)
383+
# cPickle.dump(layer4, f)
358384

359385
if patience <= iter:
360386
done_looping = True
@@ -369,32 +395,155 @@ def evaluate_lenet5(learning_rate=0.1, n_epochs=100,
369395
os.path.split(__file__)[1] +
370396
' ran for %.2fm' % ((end_time - start_time) / 60.))
371397

372-
def predict():
373-
"""
374-
An example of how to load a trained model and use it
375-
to predict labels.
376-
"""
398+
# def predict(self, data):
399+
# """
400+
# the CNN expects inputs with Nsamples = self.batch_size.
401+
# In order to run 'predict' on an arbitrary number of samples we
402+
# pad as necessary.
403+
# """
404+
# if isinstance(data, list):
405+
# data = np.array(data)
406+
# if data.ndim == 1:
407+
# data = np.array([data])
408+
409+
# nsamples = data.shape[0]
410+
# n_batches = nsamples//self.batch_size
411+
# n_rem = nsamples%self.batch_size
412+
# if n_batches > 0:
413+
# preds = [list(self.predict_wrap(data[i*self.batch_size:(i+1)*self.batch_size]))\
414+
# for i in range(n_batches)]
415+
# else:
416+
# preds = []
417+
# if n_rem > 0:
418+
# z = np.zeros((self.batch_size, self.n_in * self.n_in))
419+
# z[0:n_rem] = data[n_batches*self.batch_size:n_batches*self.batch_size+n_rem]
420+
# preds.append(self.predict_wrap(z)[0:n_rem])
421+
422+
# return np.hstack(preds).flatten()
423+
424+
def cosine_distance(a, b):
425+
import numpy as np
426+
from numpy import linalg as LA
427+
dot_product = np.dot(a,b.T)
428+
cosine_distance = dot_product / (LA.norm(a) * LA.norm(b))
429+
return cosine_distance
430+
431+
def predict():
432+
433+
from sktheano_cnn import MetaCNN as CNN
434+
cnn = CNN()
435+
436+
pkl_file = open( '../data/train_set.pkl', 'rb')
437+
train_set = cPickle.load(pkl_file)
438+
439+
pkl_file = open( '../data/valid_set.pkl', 'rb')
440+
valid_set = cPickle.load(pkl_file)
441+
442+
pkl_file = open( '../data/test_set.pkl', 'rb')
443+
test_set = cPickle.load(pkl_file)
444+
445+
"""An example of how to load a trained model and use it
446+
to predict labels.
447+
"""
448+
449+
fe = FetexImage(verbose=True)
450+
# load the saved model
451+
classifier = cPickle.load(open('best_model.pkl'))
452+
453+
layer0 = cPickle.load(open('../data/layer0.pkl'))
454+
layer1 = cPickle.load(open('../data/layer1.pkl'))
455+
# layer2 = cPickle.load(open('../data/layer2.pkl'))
456+
layer3 = cPickle.load(open('../data/layer3.pkl'))
457+
layer4 = cPickle.load(open('../data/layer4.pkl'))
458+
459+
#layer0_input = x.reshape((batch_size, 3, 64, 64))
460+
461+
# predict = theano.function(
462+
# outputs=layer4.y_pred,
463+
# givens = {x : train_set_x[0] }
464+
# )
465+
466+
# compile a predictor function
467+
predict_model = theano.function(
468+
inputs=[classifier.input],
469+
outputs=classifier.y_pred)
470+
471+
# We can test it on some examples from test test
472+
dataset='mnist.pkl.gz'
473+
datasets = load_data(dataset)
377474

378-
# load the saved model
379-
classifier = cPickle.load(open('best_model.pkl'))
475+
test_set_x, test_set_y = datasets[2]
476+
test_set_x = test_set_x.get_value()
380477

381-
# compile a predictor function
382-
predict_model = theano.function(
383-
inputs=[classifier.input],
384-
outputs=classifier.y_pred)
478+
train_set_x, train_set_y = datasets[0]
479+
train_set_x = train_set_x.get_value()
480+
481+
pkl_file = open( '../data/X_original.pkl', 'rb')
482+
X_original = cPickle.load(pkl_file)
483+
484+
a = X_original[0]
485+
#fe.reconstructImage(a).show()
486+
487+
#predicted_values = predict_model([a])
488+
489+
get_input = theano.function(
490+
inputs=[classifier.input],
491+
outputs=classifier.input
492+
)
385493

386-
# We can test it on some examples from test test
387-
dataset='mnist.pkl.gz'
388-
datasets = load_data(dataset)
389-
test_set_x, test_set_y = datasets[2]
390-
test_set_x = test_set_x.get_value()
494+
a = get_input(train_set_x[0:1])
495+
#print a.shape
391496

392-
predicted_values = predict_model(test_set_x[:10])
393-
print ("Predicted values for the first 10 examples in test set:")
394-
print predicted_values
497+
x = T.matrix('x') # the data is presented as rasterized images
498+
predict = theano.function(
499+
inputs = [x],
500+
outputs=layer3.output
501+
)
502+
# givens = { x : train_set_x[0] }
503+
#train_set_x = train_set_x[0:400]
504+
#x = train_set_x.reshape((400, 3, 64, 64))
505+
x = train_set_x.reshape(np.zeros((400,3,64,64)))
506+
print predict(x)
507+
#predicted_values = predict_model([train_set_x[0]])
508+
#print predicted_values
509+
return "fffff"
510+
511+
512+
max_similarity = 0
513+
max_similarity_pos = -1
514+
#for i in xrange(1,len(train_set_x)):
515+
for i in xrange(1,1000):
516+
b = get_input([train_set_x[i]])
517+
d = cosine_distance(a, b)
518+
if d > max_similarity:
519+
max_similarity = d
520+
max_similarity_pos = i
521+
522+
fe.reconstructImage(X_original[max_similarity_pos]).show()
523+
524+
#a = a.flatten(order='F')
525+
# a = a * 256
526+
# a = numpy.array(a,dtype=numpy.uint8)
527+
528+
#b = b.flatten(order='F')
529+
# b = b * 256
530+
# b = numpy.array(b,dtype=numpy.uint8)
531+
532+
# a = get_input([a])
533+
# b = get_input([b])
534+
535+
print a.shape
536+
print b.shape
537+
print cosine_distance(a, b)
538+
539+
# #print get_input(test_set_x[0:1]).sum()
540+
541+
# print ("Predicted values for the first 10 examples in test set:")
542+
# print predicted_values
395543

396544
if __name__ == '__main__':
397-
evaluate_lenet5()
545+
#evaluate_lenet5()
546+
predict()
398547

399548

400549
def experiment(state, channel):

0 commit comments

Comments
 (0)