SEQLEN = 30
BATCHSIZE = 100
ALPHASIZE = txt.ALPHASIZE
INTERNALSIZE = 512
NLAYERS = 3
learning_rate = 0.001 # fixed learning rate
# load data, either davinci, or the Python source of Tensorflow itself
shakedir = "davinci/*.txt"
# shakedir = "../tensorflow/**/*.py"
codetext, valitext, bookranges = txt.read_data_files(shakedir,
validation=False)
# display some stats on the data
epoch_size = len(codetext) // (BATCHSIZE * SEQLEN)
txt.print_data_stats(len(codetext), len(valitext), epoch_size)
#
# the model
#
lr = tf.placeholder(tf.float32, name='lr') # learning rate
batchsize = tf.placeholder(tf.int32, name='batchsize')
# inputs
X = tf.placeholder(tf.uint8, [None, None], name='X') # [ BATCHSIZE, SEQLEN ]
Xo = tf.one_hot(X, ALPHASIZE, 1.0, 0.0) # [ BATCHSIZE, SEQLEN, ALPHASIZE ]
# expected outputs = same sequence shifted by 1 since we are trying to predict the next character
Y_ = tf.placeholder(tf.uint8, [None, None], name='Y_') # [ BATCHSIZE, SEQLEN ]
Yo_ = tf.one_hot(Y_, ALPHASIZE, 1.0, 0.0) # [ BATCHSIZE, SEQLEN, ALPHASIZE ]
cells = [rnn.GRUCell(INTERNALSIZE) for _ in range(NLAYERS)]
NUM_LAYERS = 3 # How many layers deep we are going
SET_LR = 0.001 # Small fixed learning rate
SET_PKEEP = 0.75 # Dropping 20% of neurons
# Seed our random number generator
tf.set_random_seed(0)
# Load our Star Wars Scripts.
filedir = "StarWarsScripts/*.txt"
traintext, validtext, scriptranges = txt.read_data_files(filedir,
validation=True)
# Print out information about our data
size_of_epoch = len(traintext) // (BATCH_SIZE * SEQ_LEN)
txt.print_data_stats(len(traintext), len(validtext), size_of_epoch)
# Create our TensorFlow Graph.
batchsize = tf.placeholder(tf.int32, name='batchsize')
lr = tf.placeholder(tf.float32, name='lr')
pkeep = tf.placeholder(tf.float32, name='pkeep')
X = tf.placeholder(tf.uint8, [None, None], name='X') # Input vector
Xo = tf.one_hot(
X, ALPHA_SIZE, 1.0,
0.0) # One Hots create vector size ALPHA_SIZE, all set 0 except character
Y_ = tf.placeholder(tf.uint8, [None, None], name='Y_') # Output tensor
Yo_ = tf.one_hot(Y_, ALPHA_SIZE, 1.0, 0.0) # OneHot our output also
Hin = tf.placeholder(tf.float32, [None, NUM_OF_GRUS * NUM_LAYERS],
name='Hin') # Recurrent input states
cells = [rnn.GRUCell(NUM_OF_GRUS)
for _ in range(NUM_LAYERS)] # Create all our GRU cells per layer