RNN的介绍

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66

67

68

69

70

71

72

73

74

75

76

77

78

79

80

81

82

83

84

85

86

87

88

89

90

91

92

93

94

95

96

97

98

99

100

101

102

103

104

105

106

107

108

109

110

111

112

"""

Minimal character-level Vanilla RNN model. Written by Andrej Karpathy (@karpathy)

BSD License

"""

import numpy as np

# data I/O

data = open(‘input.txt‘, ‘r‘).read() # should be simple plain text file

chars = list(set(data))

data_size, vocab_size = len(data), len(chars)

print ‘data has %d characters, %d unique.‘ % (data_size, vocab_size)

char_to_ix = { ch:i for i,ch in enumerate(chars) }

ix_to_char = { i:ch for i,ch in enumerate(chars) }

# hyperparameters

hidden_size = 100 # size of hidden layer of neurons

seq_length = 25 # number of steps to unroll the RNN for

learning_rate = 1e-1

# model parameters

Wxh = np.random.randn(hidden_size, vocab_size)*0.01 # input to hidden

Whh = np.random.randn(hidden_size, hidden_size)*0.01 # hidden to hidden

Why = np.random.randn(vocab_size, hidden_size)*0.01 # hidden to output

bh = np.zeros((hidden_size, 1)) # hidden bias

by = np.zeros((vocab_size, 1)) # output bias

def lossFun(inputs, targets, hprev):

  """

  inputs,targets are both list of integers.

  hprev is Hx1 array of initial hidden state

  returns the loss, gradients on model parameters, and last hidden state

  """

  xs, hs, ys, ps = {}, {}, {}, {}

  hs[-1] = np.copy(hprev)

  loss = 0

  # forward pass

  for t in xrange(len(inputs)):

    xs[t] = np.zeros((vocab_size,1)) # encode in 1-of-k representation

    xs[t][inputs[t]] = 1

    hs[t] = np.tanh(np.dot(Wxh, xs[t]) + np.dot(Whh, hs[t-1]) + bh) # hidden state

    ys[t] = np.dot(Why, hs[t]) + by # unnormalized log probabilities for next chars

    ps[t] = np.exp(ys[t]) / np.sum(np.exp(ys[t])) # probabilities for next chars

    loss += -np.log(ps[t][targets[t],0]) # softmax (cross-entropy loss)

  # backward pass: compute gradients going backwards

  dWxh, dWhh, dWhy = np.zeros_like(Wxh), np.zeros_like(Whh), np.zeros_like(Why)

  dbh, dby = np.zeros_like(bh), np.zeros_like(by)

  dhnext = np.zeros_like(hs[0])

  for t in reversed(xrange(len(inputs))):

    dy = np.copy(ps[t])

    dy[targets[t]] -= 1 # backprop into y. see http://cs231n.github.io/neural-networks-case-study/#grad if confused here

    dWhy += np.dot(dy, hs[t].T)

    dby += dy

    dh = np.dot(Why.T, dy) + dhnext # backprop into h

    dhraw = (1 - hs[t] * hs[t]) * dh # backprop through tanh nonlinearity

    dbh += dhraw

    dWxh += np.dot(dhraw, xs[t].T)

    dWhh += np.dot(dhraw, hs[t-1].T)

    dhnext = np.dot(Whh.T, dhraw)

  for dparam in [dWxh, dWhh, dWhy, dbh, dby]:

    np.clip(dparam, -5, 5, out=dparam) # clip to mitigate exploding gradients

  return loss, dWxh, dWhh, dWhy, dbh, dby, hs[len(inputs)-1]

def sample(h, seed_ix, n):

  """

  sample a sequence of integers from the model

  h is memory state, seed_ix is seed letter for first time step

  """

  x = np.zeros((vocab_size, 1))

  x[seed_ix] = 1

  ixes = []

  for t in xrange(n):

    h = np.tanh(np.dot(Wxh, x) + np.dot(Whh, h) + bh)

    y = np.dot(Why, h) + by

    p = np.exp(y) / np.sum(np.exp(y))

    ix = np.random.choice(range(vocab_size), p=p.ravel())

    x = np.zeros((vocab_size, 1))

    x[ix] = 1

    ixes.append(ix)

  return ixes

n, p = 0, 0

mWxh, mWhh, mWhy = np.zeros_like(Wxh), np.zeros_like(Whh), np.zeros_like(Why)

mbh, mby = np.zeros_like(bh), np.zeros_like(by) # memory variables for Adagrad

smooth_loss = -np.log(1.0/vocab_size)*seq_length # loss at iteration 0

while True:

  # prepare inputs (we‘re sweeping from left to right in steps seq_length long)

  if p+seq_length+1 >= len(data) or n == 0:

    hprev = np.zeros((hidden_size,1)) # reset RNN memory

    p = 0 # go from start of data

  inputs = [char_to_ix[ch] for ch in data[p:p+seq_length]]

  targets = [char_to_ix[ch] for ch in data[p+1:p+seq_length+1]]

  # sample from the model now and then

  if n % 100 == 0:

    sample_ix = sample(hprev, inputs[0], 200)

    txt = ‘‘.join(ix_to_char[ix] for ix in sample_ix)

    print ‘----\n %s \n----‘ % (txt, )

  # forward seq_length characters through the net and fetch gradient

  loss, dWxh, dWhh, dWhy, dbh, dby, hprev = lossFun(inputs, targets, hprev)

  smooth_loss = smooth_loss * 0.999 + loss * 0.001

  if n % 100 == 0: print ‘iter %d, loss: %f‘ % (n, smooth_loss) # print progress

  # perform parameter update with Adagrad

  for param, dparam, mem in zip([Wxh, Whh, Why, bh, by],

                                [dWxh, dWhh, dWhy, dbh, dby],

                                [mWxh, mWhh, mWhy, mbh, mby]):

    mem += dparam * dparam

    param += -learning_rate * dparam / np.sqrt(mem + 1e-8) # adagrad update

  p += seq_length # move data pointer

n += 1 # iteration counter