In [ ]:
# Import libraries
import tensorflow as tf
import numpy as np
import collections
import os
In [ ]:
# Load text data
data = open(os.path.join('datasets', 'text_ass_6.txt'), 'r').read() # must be simple plain text file
print('Text data:',data)
chars = list(set(data))
print('\nSingle characters:',chars)
data_len, vocab_size = len(data), len(chars)
print('\nText data has %d characters, %d unique.' % (data_len, vocab_size))
char_to_ix = { ch:i for i,ch in enumerate(chars) }
ix_to_char = { i:ch for i,ch in enumerate(chars) }
print('\nMapping characters to numbers:',char_to_ix)
print('\nMapping numbers to characters:',ix_to_char)
The goal is to define with TensorFlow a vanilla recurrent neural network (RNN) model:
$$ \begin{aligned} h_t &= \textrm{tanh}(W_h h_{t-1} + W_x x_t + b_h)\\ y_t &= W_y h_t + b_y \end{aligned} $$to predict a sequence of characters. $x_t \in \mathbb{R}^D$ is the input character of the RNN in a dictionary of size $D$. $y_t \in \mathbb{R}^D$ is the predicted character (through a distribution function) by the RNN system. $h_t \in \mathbb{R}^H$ is the memory of the RNN, called hidden state at time $t$. Its dimensionality is arbitrarly chosen to $H$. The variables of the system are $W_h \in \mathbb{R}^{H\times H}$, $W_x \in \mathbb{R}^{H\times D}$, $W_y \in \mathbb{R}^{D\times H}$, $b_h \in \mathbb{R}^D$, and $b_y \in \mathbb{R}^D$.
The number of time steps of the RNN is $T$, that is we will learn a sequence of data of length $T$: $x_t$ for $t=0,...,T-1$.
In [ ]:
# hyperparameters of RNN
batch_size = 3 # batch size
batch_len = data_len // batch_size # batch length
T = 5 # temporal length
epoch_size = (batch_len - 1) // T # nb of iterations to get one epoch
D = vocab_size # data dimension = nb of unique characters
H = 5*D # size of hidden state, the memory layer
print('data_len=',data_len,' batch_size=',batch_size,' batch_len=',
batch_len,' T=',T,' epoch_size=',epoch_size,' D=',D)
Initialize input variables of the computational graph:
(1) Xin of size batch_size x T x D and type tf.float32. Each input character is encoded on a vector of size D.
(2) Ytarget of size batch_size x T and type tf.int64. Each target character is encoded by a value in {0,...,D-1}.
(3) hin of size batch_size x H and type tf.float32
In [ ]:
# input variables of computational graph (CG)
Xin = tf.placeholder(tf.float32, [batch_size,T,D]); #print('Xin=',Xin) # Input
Ytarget = tf.placeholder(tf.int64, [batch_size,T]); #print('Y_=',Y_) # target
hin = tf.placeholder(tf.float32, [batch_size,H]); #print('hin=',hin.get_shape())
Define the variables of the computational graph:
(1) $W_x$ is a random variable of shape D x H with normal distribution of variance $\frac{6}{D+H}$
(2) $W_h$ is an identity matrix multiplies by constant $0.01$
(3) $W_y$ is a random variable of shape H x D with normal distribution of variance $\frac{6}{D+H}$
(4) $b_h$, $b_y$ are zero vectors of size H, and D
In [ ]:
# Model variables
Wx = tf.Variable(tf.random_normal([D,H], stddev=tf.sqrt(6./tf.to_float(D+H)))); print('Wx=',Wx.get_shape())
Wh = tf.Variable(0.01*np.identity(H, np.float32)); print('Wh=',Wh.get_shape())
Wy = tf.Variable(tf.random_normal([H,D], stddev=tf.sqrt(6./tf.to_float(H+D)))); print('Wy=',Wy.get_shape())
bh = tf.Variable(tf.zeros([H])); print('bh=',bh.get_shape())
by = tf.Variable(tf.zeros([D])); print('by=',by.get_shape())
Implement the recursive formula:
$$ \begin{aligned} h_t &= \textrm{tanh}(W_h h_{t-1} + W_x x_t + b_h)\\ y_t &= W_y h_t + b_y \end{aligned} $$with $h_{t=0}=hin$.
Hints:
(1) You may use functions tf.split(), enumerate(), tf.squeeze(), tf.matmul(), tf.tanh(), tf.transpose(), append(), pack().
(2) You may use a matrix Y of shape batch_size x T x D. We recall that Ytarget should have the shape batch_size x T.
In [ ]:
# Vanilla RNN implementation
Y = []
ht = hin
for t, xt in enumerate(tf.split(1, T, Xin)):
if batch_size>1:
xt = tf.squeeze(xt); #print('xt=',xt)
else:
xt = tf.squeeze(xt)[None,:]
ht = tf.matmul(ht, Wh); #print('ht1=',ht)
ht += tf.matmul(xt, Wx); #print('ht2=',ht)
ht += bh; #print('ht3=',ht)
ht = tf.tanh(ht); #print('ht4=',ht)
yt = tf.matmul(ht, Wy); #print('yt1=',yt)
yt += by; #print('yt2=',yt)
Y.append(yt)
#print('Y=',Y)
Y = tf.pack(Y);
if batch_size>1:
Y = tf.squeeze(Y);
Y = tf.transpose(Y, [1, 0, 2])
print('Y=',Y.get_shape())
print('Ytarget=',Ytarget.get_shape())
In [ ]:
# perplexity
logits = tf.reshape(Y,[batch_size*T,D])
weights = tf.ones([batch_size*T])
cross_entropy_perplexity = tf.nn.seq2seq.sequence_loss_by_example([logits],[Ytarget],[weights])
cross_entropy_perplexity = tf.reduce_sum(cross_entropy_perplexity) / batch_size
loss = cross_entropy_perplexity
In [ ]:
# Optimization
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)
In [ ]:
# Predict
idx_pred = tf.placeholder(tf.int64) # input seed
xtp = tf.one_hot(idx_pred,depth=D); #print('xtp1=',xtp.get_shape())
htp = tf.zeros([1,H])
Ypred = []
for t in range(T):
htp = tf.matmul(htp, Wh); #print('htp1=',htp)
htp += tf.matmul(xtp, Wx); #print('htp2=',htp)
htp += bh; #print('htp3=',htp) # (1, 100)
htp = tf.tanh(htp); #print('htp4=',htp) # (1, 100)
ytp = tf.matmul(htp, Wy); #print('ytp1=',ytp)
ytp += by; #print('ytp2=',ytp)
ytp = tf.nn.softmax(ytp); #print('yt1=',ytp)
ytp = tf.squeeze(ytp); #print('yt2=',ytp)
seed_idx = tf.argmax(ytp,dimension=0); #print('seed_idx=',seed_idx)
xtp = tf.one_hot(seed_idx,depth=D)[None,:]; #print('xtp2=',xtp.get_shape())
Ypred.append(seed_idx)
Ypred = tf.convert_to_tensor(Ypred)
In [ ]:
# Prepare train data matrix of size "batch_size x batch_len"
data_ix = [char_to_ix[ch] for ch in data[:data_len]]
train_data = np.array(data_ix)
print('original train set shape',train_data.shape)
train_data = np.reshape(train_data[:batch_size*batch_len], [batch_size,batch_len])
print('pre-processed train set shape',train_data.shape)
In [ ]:
# The following function tansforms an integer value d between {0,...,D-1} into an one hot vector, that is a
# vector of dimension D x 1 which has value 1 for index d-1, and 0 otherwise
from scipy.sparse import coo_matrix
def convert_to_one_hot(a,max_val=None):
N = a.size
data = np.ones(N,dtype=int)
sparse_out = coo_matrix((data,(np.arange(N),a.ravel())), shape=(N,max_val))
return np.array(sparse_out.todense())
Run the computational graph with batches of training data.
Predict the sequence of characters starting from the character "h".
Hints:
(1) Initial memory is $h_{t=0}$ is 0.
(2) Run the computational graph to optimize the perplexity loss, and to predict the the sequence of characters starting from the character "h".
In [ ]:
# Run CG
init = tf.initialize_all_variables()
sess = tf.Session()
sess.run(init)
h0 = np.zeros([batch_size,H])
indices = collections.deque()
costs = 0.0; epoch_iters = 0
for n in range(50):
# Batch extraction
if len(indices) < 1:
indices.extend(range(epoch_size))
costs = 0.0; epoch_iters = 0
i = indices.popleft()
batch_x = train_data[:,i*T:(i+1)*T]
batch_x = convert_to_one_hot(batch_x,D); batch_x = np.reshape(batch_x,[batch_size,T,D])
batch_y = train_data[:,i*T+1:(i+1)*T+1]
#print(batch_x.shape,batch_y.shape)
# Train
idx = char_to_ix['h'];
loss_value,_,Ypredicted = sess.run([loss,train_step,Ypred], feed_dict={Xin: batch_x, Ytarget: batch_y, hin: h0, idx_pred: [idx]})
# Perplexity
costs += loss_value
epoch_iters += T
perplexity = np.exp(costs/epoch_iters)
if not n%1:
idx_char = Ypredicted
txt = ''.join(ix_to_char[ix] for ix in list(idx_char))
print('\nn=',n,', perplexity value=',perplexity)
print('starting char=',ix_to_char[idx], ', predicted sequences=',txt)
sess.close()