ffnet:Feed-forward neural network for Python

是 否 +2 论坛币

k人 参与回答

经管之家送您一份

应届毕业生专属福利!

求职就业群

赵安豆老师微信：zhaoandou666

经管之家联合CDA

送您一个全额奖学金名额~ !

立即领取

感谢您参与论坛问题回答

经管之家送您两个论坛币！

+2 论坛币

1. ### Digits recognition example for ffnet ###
   
2. 
   
3. # Training file (data/ocr.dat) contains 68 patterns - first 58
   
4. # are used for training and last 10 are used for testing.
   
5. # Each pattern contains 64 inputs which define 8x8 bitmap of
   
6. # the digit and last 10 numbers are the targets (10 targets for 10 digits).
   
7. # Layered network architecture is used here: (64, 10, 10, 10).
   
8. 
   
9. from ffnet import ffnet, mlgraph, readdata
   
10. 
    
11. # Generate standard layered network architecture and create network
    
12. conec = mlgraph((64,10,10,10))
    
13. net = ffnet(conec)
    
14. 
    
15. # Read data file
    
16. print "READING DATA..."
    
17. data = readdata( 'data/ocr.dat', delimiter = ' ' )
    
18. input =  data[:, :64] #first 64 columns - bitmap definition
    
19. target = data[:, 64:] #the rest - 10 columns for 10 digits
    
20. 
    
21. # Train network with scipy tnc optimizer - 58 lines used for training
    
22. print "TRAINING NETWORK..."
    
23. net.train_tnc(input[:58], target[:58], maxfun = 2000, messages=1)
    
24. 
    
25. # Test network - remaining 10 lines used for testing
    
26. print
    
27. print "TESTING NETWORK..."
    
28. output, regression = net.test(input[58:], target[58:], iprint = 2)
    
29. 
    
30. ############################################################
    
31. # Make a plot of a chosen digit along with the network guess
    
32. try:
    
33.     from pylab import *
    
34.     from random import randint
    
35. 
    
36.     digitpat = randint(58, 67) #Choose testing pattern to plot
    
37. 
    
38.     subplot(211)
    
39.     imshow(input[digitpat].reshape(8,8), interpolation = 'nearest')
    
40. 
    
41.     subplot(212)
    
42.     N = 10  # number of digits / network outputs
    
43.     ind = arange(N)   # the x locations for the groups
    
44.     width = 0.35       # the width of the bars
    
45.     bar(ind, net(input[digitpat]), width, color='b') #make a plot
    
46.     xticks(ind+width/2., ('1', '2', '3', '4', '5', '6', '7', '8', '9', '0'))
    
47.     xlim(-width,N-width)
    
48.     axhline(linewidth=1, color='black')
    
49.     title("Trained network (64-10-10-10) guesses a digit above...")
    
50.     xlabel("Digit")
    
51.     ylabel("Network outputs")
    
52. 
    
53.     show()
    
54. except ImportError, e:
    
55.     print "Cannot make plots. For plotting install matplotlib.\n%s" % e

复制代码

https://pypi.python.org/pypi/ffnet

扫码加我 拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

分享0 收藏1 回帖

关键词：network forward python Neural ward standard training network testing example

1. ### Sine training example for ffnet ###
   
2. 
   
3. from ffnet import ffnet
   
4. from math import pi, sin, cos
   
5. 
   
6. # Let's define network connectivity by hand:
   
7. conec = [(1, 2), (1, 3), (1, 4), (1, 5), (2, 6), (3, 6), (4, 6), (5, 6), \
   
8.          (0, 2), (0, 3), (0, 4), (0, 5), (0, 6)]
   
9. # Note 1: Biases in ffnet are handled as the connections
   
10. #         from special node numbered 0. Input nodes cannot be biased.
    
11. # Note 2: Node numbering and order of links in conec is meaningless,
    
12. #         but the connections have to be from source to target.
    
13. # Note 3: The same connectivity can be obtained using mlgraph function
    
14. #         provided with ffnet (layered architecture (1,4,1)).
    
15. 
    
16. # Network creation
    
17. net = ffnet(conec)
    
18. 
    
19. # Generation of training data (sine values for x from 0 to 2*pi)
    
20. patterns = 16
    
21. input = [ [ 0. ] ] + [ [ k*2*pi/patterns ] for k in xrange(1, patterns + 1) ]
    
22. target = [ [ sin(x[0]) ] for x in input ]
    
23. 
    
24. # Training network
    
25. #first find good starting point with genetic algorithm (not necessary, but may be helpful)
    
26. print "FINDING STARTING WEIGHTS WITH GENETIC ALGORITHM..."
    
27. net.train_genetic(input, target, individuals=20, generations=500)
    
28. #then train with scipy tnc optimizer
    
29. print "TRAINING NETWORK..."
    
30. net.train_tnc(input, target, maxfun = 5000, messages=1)
    
31. 
    
32. # Testing network
    
33. print
    
34. print "TESTNG NETWORK..."
    
35. output, regression = net.test(input, target, iprint = 2)
    
36. 
    
37. #################
    
38. # Make some plots
    
39. try:
    
40.     from pylab import *
    
41.     points = 128
    
42.     xaxis = [ 0. ] + [ k*2*pi/points for k in xrange(1, points + 1) ]
    
43.     sine = [ sin(x) for x in xaxis ]
    
44.     cosine = [ cos(x) for x in xaxis ]
    
45.     netsine = [ net([x])[0] for x in xaxis]
    
46.     netcosine = [ net.derivative([x])[0][0] for x in xaxis ]
    
47. 
    
48.     subplot(211)
    
49.     plot(xaxis, sine, 'b--', xaxis, netsine, 'k-')
    
50.     legend(('sine', 'network output'))
    
51.     grid(True)
    
52.     title('Outputs of trained network.')
    
53. 
    
54.     subplot(212)
    
55.     plot(xaxis, cosine, 'b--', xaxis, netcosine, 'k-')
    
56.     legend(('cosine', 'network derivative'))
    
57.     grid(True)
    
58.     show()
    
59. except ImportError:
    
60.     print "Cannot make plots. For plotting install matplotlib..."

复制代码

1. ### XOR problem example for ffnet ###
   
2. 
   
3. from ffnet import ffnet, mlgraph
   
4. 
   
5. # Generate standard layered network architecture and create network
   
6. conec = mlgraph((2,2,1))
   
7. net = ffnet(conec)
   
8. 
   
9. # Define training data
   
10. input = [[0.,0.], [0.,1.], [1.,0.], [1.,1.]]
    
11. target  = [[1.], [0.], [0.], [1.]]
    
12. 
    
13. # Train network
    
14. #first find good starting point with genetic algorithm (not necessary, but may be helpful)
    
15. print "FINDING STARTING WEIGHTS WITH GENETIC ALGORITHM..."
    
16. net.train_genetic(input, target, individuals=20, generations=500)
    
17. #then train with scipy tnc optimizer
    
18. print "TRAINING NETWORK..."
    
19. net.train_tnc(input, target, maxfun = 1000, messages=1)
    
20. 
    
21. # Test network
    
22. print
    
23. print "TESTING NETWORK..."
    
24. output, regression = net.test(input, target, iprint = 2)
    
25. 
    
26. # Save/load/export network
    
27. from ffnet import savenet, loadnet, exportnet
    
28. print "Network is saved..."
    
29. savenet(net, "xor.net")
    
30. print "Network is reloaded..."
    
31. net = loadnet("xor.net")
    
32. print "Network is tested again, but nothing is printed..."
    
33. output, regression = net.test(input, target, iprint = 0)
    
34. print
    
35. print "Exporting trained network to the fortran source..."
    
36. exportnet(net, "xor.f")
    
37. print "Done..."
    
38. print "Look at the generated xor.f file."
    
39. print "Note: you must compile xor.f along with the ffnet.f"
    
40. print "file which can be found in ffnet sources."

复制代码

1. ### Stock Problem Example for ffnet ###
   
2. 
   
3. # Data and description for this example is borrowed from:
   
4. # http://www.scientific-consultants.com/nnbd.html
   
5. #
   
6. #
   
7. # Training data consists of Black-Scholes option prices
   
8. # for volatility levels running from 20 percent to 200 percent,
   
9. # for time remaining running from 5 to 15 days, and for strike price
   
10. # running from 70 dollars to 130 dollars. The stock price was set to
    
11. # 100 dollars and the interest rate to 0 percent when generating
    
12. # the data.
    
13. #
    
14. # The data is "difficult" in that (for a neural network to
    
15. # practically emulate Black-Scholes) a very tight fit is required.
    
16. # The R-squared should be at least 0.999999 or better, and the largest
    
17. # absolute error must be less than 0.05 dollars (the price increment
    
18. # for most options) or, better yet, less than 0.01 dollars.
    
19. #
    
20. #
    
21. # So let's try.
    
22. # Attention: training might be a long process since we train a big network.
    
23. 
    
24. from ffnet import ffnet, mlgraph, readdata
    
25. from numpy import array
    
26. 
    
27. # Generate standard layered network architecture and create network
    
28. conec = mlgraph((3,22,12,1))
    
29. net = ffnet(conec)
    
30. 
    
31. # Read training data omitting first column and first line
    
32. print "READING DATA..."
    
33. data = readdata( 'data/black-scholes.dat',
    
34.                  usecols  = (1, 2, 3, 4),
    
35.                  skiprows =  1)
    
36. input =  data[:, :3] #first 3 columns
    
37. target = data[:, -1] #last column
    
38. 
    
39. print "TRAINING NETWORK..."
    
40. import sys; sys.stdout.flush() #Just to ensure dislpaing the above messages here
    
41. net.train_tnc(input, target, maxfun = 5000, messages=1)
    
42. 
    
43. # Test network
    
44. print
    
45. print "TESTING NETWORK..."
    
46. output, regression = net.test(input, target, iprint = 0)
    
47. Rsquared = regression[0][2]
    
48. maxerr = abs( array(output).reshape( len(output) ) - array(target) ).max()
    
49. print "R-squared:           %s  (should be >= 0.999999)" %str(Rsquared)
    
50. print "max. absolute error: %s  (should be <= 0.05)" %str(maxerr)
    
51. print
    
52. print "Is ffnet ready for a stock?"
    
53. 
    
54. #####################################
    
55. # Make plot if matplotlib is avialble
    
56. try:
    
57.     from pylab import *
    
58.     plot( target, 'b--' )
    
59.     plot( output, 'k-' )
    
60.     legend(('target', 'output'))
    
61.     xlabel('pattern'); ylabel('price')
    
62.     title('Outputs vs. target of trained network.')
    
63.     grid(True)
    
64.     show()
    
65. except ImportError, e:
    
66.     print "Cannot make plots. For plotting install matplotlib.\n%s" % e

复制代码

1. ### Multiprocessing training example for ffnet ###
   
2. 
   
3. from ffnet import ffnet, mlgraph
   
4. from scipy import rand
   
5. 
   
6. # Generate random training data (large)
   
7. input = rand(10000, 10)
   
8. target = rand(10000, 1)
   
9. 
   
10. # Define net (large one)
    
11. conec = mlgraph((10,300,1))
    
12. net = ffnet(conec)
    
13. 
    
14. # Test training speed-up
    
15. # Note that the below *if* is necessary only on Windows
    
16. if __name__=='__main__':   
    
17.     from time import time
    
18.     from multiprocessing import cpu_count
    
19.    
    
20.     # Preserve original weights
    
21.     weights0 = net.weights.copy()
    
22.    
    
23.     print "TRAINING, this can take a while..."
    
24.     for n in range(1, cpu_count()+1):
    
25.         net.weights[:] = weights0  #Start always from the same point
    
26.         t0 = time()
    
27.         net.train_tnc(input, target, nproc = n, maxfun=50, messages=0)
    
28.         t1 = time()
    
29.         print '%s processes: %s s' %(n, t1-t0)

复制代码

辛苦辛苦

感谢 分享