def __init__(self, rootDir, batchSize, seqLen): self.batchSize = batchSize self.seqLen = seqLen self.rootDir = rootDir # load data, either shakespeare, or the Python source of Tensorflow itself self.textdir = self.rootDir + "/*.txt" #shakedir = "../tensorflow/**/*.py" self.codetext, self.valitext, self.bookranges = txt.read_data_files(self.textdir, validation=True) self.epoch_size = len(self.codetext) // (self.batchSize * self.seqLen) # display some stats on the data txt.print_data_stats(len(self.codetext), len(self.valitext), self.epoch_size)
def validate_on_network(auth):
with tf.Session() as sess:
new_saver = tf.train.import_meta_graph(auth + '.meta')
new_saver.restore(sess, auth)
valitext, _, __ = my_txtutils.read_data_files(directory, validation=False)
VALI_SEQLEN = 1 * 64 # Sequence length for validation. State will be wrong at the start of each sequence.
bsize = len(valitext) // VALI_SEQLEN
vali_x, vali_y, _ = next(
my_txtutils.rnn_minibatch_sequencer(valitext, bsize, VALI_SEQLEN, 1)) # all data in 1 batch
vali_nullstate = np.zeros([bsize, INTERNALSIZE * NLAYERS])
feed_dict = {'inputs/X:0': vali_x, 'target/Y_:0': vali_y, 'model/pkeep:0': 1.0,
'hidden_state/Hin:0': vali_nullstate, 'model/batchsize:0': bsize}
ls, acc = sess.run(["display_data/batchloss:0", "display_data/accuracy:0"], feed_dict=feed_dict)
my_txtutils.print_validation_stats(ls, acc)
# A good choice of parameters ensures that the testing and validation curves stay close
# To see the curves drift apart ("overfitting") try to use an insufficient amount of
# training data (shakedir = "shakespeare/t*.txt" for example)
#
SEQLEN = 30
BATCHSIZE = 100
ALPHASIZE = txt.ALPHASIZE
INTERNALSIZE = 512
NLAYERS = 3
learning_rate = 0.001 # fixed learning rate
dropout_pkeep = 1.0 # no dropout
# load data, either shakespeare, or the Python source of Tensorflow itself
shakedir = "dickens/*.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 (see FAQ in README.md)
#
lr = tf.placeholder(tf.float32, name='lr') # learning rate
pkeep = tf.placeholder(tf.float32, name='pkeep') # dropout parameter
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 ]
# Training and experimentation (default):
# Keep validation enabled
# You can now play with the parameters anf follow the effects in Tensorboard
# A good choice of parameters ensures that the testing and validation curves stay close
# To see the curves drift apart ("overfitting") try to use an insufficient amount of
#
SEQLEN = 200
BATCHSIZE = 80
ALPHASIZE = txt.ALPHASIZE
INTERNALSIZE = 512
NLAYERS = 3
learning_rate = 0.001 # fixed learning rate
dropout_pkeep = 0.8 # some dropout
bibledir = "bible/*.txt"
codetext, valitext, bookranges = txt.read_data_files(bibledir, validation=True)
# display some stats on the data
epoch_size = len(codetext) // (BATCHSIZE * SEQLEN)
txt.print_data_stats(len(codetext), len(valitext), epoch_size)
#
# the model (see FAQ in README.md)
#
lr = tf.placeholder(tf.float32, name='lr') # learning rate
pkeep = tf.placeholder(tf.float32, name='pkeep') # dropout parameter
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 ]
BATCH_SIZE = 100
ALPHA_SIZE = txt.ALPHASIZE
INTERNAL_SIZE = 512
NLAYERS = 3
learning_rate = 0.001
dropout_pkeep = 0.5 # No dropout
# Load training data
# Shakespeare
# train_data = "shakespeare/*.txt"
# Java code
train_data = "../Desktop/android/frameworks/**/*.java"
codetxt, valitext, bookranges = txt.read_data_files(train_data,
validation=True)
# Model
lr = tf.placeholder(tf.float32, name='lr')
pkeep = tf.placeholder(tf.float32, name='pkeep')
batchsize = tf.placeholder(tf.int32, name='batchsize')
# Input
X = tf.placeholder(tf.uint8, [None, None], name='X') # [BATCH_SIZE, SEQ_LEN]
Xo = tf.one_hot(X, ALPHA_SIZE, 1.0, 0.0) # [BATCH_SIZE, SEQ_LEN, ALPHA_SIZE]
# Output
Y_ = tf.placeholder(tf.uint8, [None, None], name='Y_') # [BATCH_SIZE, SEQ_LEN]
Yo_ = tf.one_hot(Y_, ALPHA_SIZE, 1.0, 0.0) # [BATCH_SIZE, SEQ_LEN, ALPHA_SIZE]
# Input state
# training data (input_dir = "shakespeare/t*.txt" for example)
#
SEQLEN = 30
BATCHSIZE = 200
ALPHASIZE = txt.ALPHASIZE
INTERNALSIZE = 512
NLAYERS = 3
learning_rate = 0.001 # fixed learning rate
dropout_pkeep = 0.8 # some dropout
epoch_count = 13
project_name = "shakespeare"
# load data, from the project subfolder or the Python source of Tensorflow itself
input_dir = "{}/*.txt".format(project_name)
#input_dir = "../tensorflow/**/*.py"
codetext, valitext, bookranges = txt.read_data_files(input_dir,
validation=True)
# display some stats on the data
epoch_size = len(codetext) // (BATCHSIZE * SEQLEN)
txt.print_data_stats(len(codetext), len(valitext), epoch_size)
#
# the model (see FAQ in README.md)
#
lr = tf.placeholder(tf.float32, name='lr') # learning rate
pkeep = tf.placeholder(tf.float32, name='pkeep') # dropout parameter
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 ]
SEQ_LEN = 64 # Number of characters per sequence
NUM_EPOCHS = 50 # Number of epochs
BATCH_SIZE = 250 # Sequences per batch
NUM_OF_GRUS = 1024 # Number of GRU cells per layer
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
# You can now play with the parameters and follow the effects in Tensorboard
# A good choice of parameters ensures that the testing and validation curves stay close
# To see the curves drift apart ("overfitting") try to use an insufficient amount of
# training data
#
SEQLEN = 30
BATCHSIZE = 200
ALPHASIZE = txt.ALPHASIZE
INTERNALSIZE = 512
NLAYERS = 3
learning_rate = 0.001 # fixed learning rate
dropout_pkeep = 0.8 # some dropout
# validation = True
validation = False
codetext, valitext, fileranges = txt.read_data_files(args.globby, validation)
# display some stats on the data
epoch_size = len(codetext) // (BATCHSIZE * SEQLEN)
txt.print_data_stats(len(codetext), len(valitext), epoch_size)
#
# the model (see FAQ in README.md)
#
lr = tf.placeholder(tf.float32, name='lr') # learning rate
pkeep = tf.placeholder(tf.float32, name='pkeep') # dropout parameter
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 ]
timee = datetime.datetime.now() # ADAPTED FROM 'DEEP LEARNING WITHOUT A PHD' FROM GOOGLE # model parameters BATCHSIZE = 200 ALPHASIZE = txt.ALPHASIZE INTERNALSIZE = 512 NLAYERS = 3 SEQLEN = 30 learning_rate = 0.001 # fixed learning rate dropout_pkeep = 0.8 # some dropout # load data docdir = "all/*.txt" code_text, validation_text, bookranges = txt.read_data_files(docdir, validation=True) # display some stats on the data epoch_size = len(code_text) // (BATCHSIZE * SEQLEN) txt.print_data_stats(len(code_text), len(validation_text), epoch_size) lr = tf.placeholder(tf.float32, name='lr') # learning rate pkeep = tf.placeholder(tf.float32, name='pkeep') # dropout parameter batchsize = tf.placeholder(tf.int32, name='batchsize') # batch size parameter # 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 ]
def main(_):
# load data, either shakespeare, or the Python source of Tensorflow itself
shakedir = FLAGS.text_dir
# shakedir = "../tensorflow/**/*.py"
codetext, valitext, bookranges = txt.read_data_files(shakedir,
validation=True)
# display some stats on the data
epoch_size = len(codetext) // (FLAGS.train_batch_size * FLAGS.seqlen)
txt.print_data_stats(len(codetext), len(valitext), epoch_size)
#
# the model (see FAQ in README.md)
#
lr = tf.placeholder(tf.float32, name='lr') # learning rate
pkeep = tf.placeholder(tf.float32, name='pkeep') # dropout parameter
batchsize = tf.placeholder(tf.int32, name='batchsize')
# inputs
X = tf.placeholder(tf.uint8, [None, None],
name='X') # [ BATCHSIZE, FLAGS.seqlen ]
Xo = tf.one_hot(X, ALPHASIZE, 1.0,
0.0) # [ BATCHSIZE, FLAGS.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, FLAGS.seqlen ]
Yo_ = tf.one_hot(Y_, ALPHASIZE, 1.0,
0.0) # [ BATCHSIZE, FLAGS.seqlen, ALPHASIZE ]
# input state
Hin = tf.placeholder(tf.float32, [None, INTERNALSIZE * NLAYERS],
name='Hin') # [ BATCHSIZE, INTERNALSIZE * NLAYERS]
# using a NLAYERS=3 layers of GRU cells, unrolled FLAGS.seqlen=30 times
# dynamic_rnn infers FLAGS.seqlen from the size of the inputs Xo
onecell = rnn.GRUCell(INTERNALSIZE)
dropcell = rnn.DropoutWrapper(onecell, input_keep_prob=pkeep)
multicell = rnn.MultiRNNCell([dropcell] * NLAYERS, state_is_tuple=False)
multicell = rnn.DropoutWrapper(multicell, output_keep_prob=pkeep)
Yr, H = tf.nn.dynamic_rnn(multicell,
Xo,
dtype=tf.float32,
initial_state=Hin)
# Yr: [ BATCHSIZE, FLAGS.seqlen, INTERNALSIZE ]
# H: [ BATCHSIZE, INTERNALSIZE*NLAYERS ] # this is the last state in the sequence
H = tf.identity(H, name='H') # just to give it a name
# Softmax layer implementation:
# Flatten the first two dimension of the output [ BATCHSIZE, FLAGS.seqlen, ALPHASIZE ] => [ BATCHSIZE x FLAGS.seqlen, ALPHASIZE ]
# then apply softmax readout layer. This way, the weights and biases are shared across unrolled time steps.
# From the readout point of view, a value coming from a cell or a minibatch is the same thing
Yflat = tf.reshape(
Yr, [-1, INTERNALSIZE]) # [ BATCHSIZE x FLAGS.seqlen, INTERNALSIZE ]
Ylogits = layers.linear(
Yflat, ALPHASIZE) # [ BATCHSIZE x FLAGS.seqlen, ALPHASIZE ]
Yflat_ = tf.reshape(
Yo_, [-1, ALPHASIZE]) # [ BATCHSIZE x FLAGS.seqlen, ALPHASIZE ]
loss = tf.nn.softmax_cross_entropy_with_logits(
logits=Ylogits, labels=Yflat_) # [ BATCHSIZE x FLAGS.seqlen ]
loss = tf.reshape(loss, [batchsize, -1]) # [ BATCHSIZE, FLAGS.seqlen ]
Yo = tf.nn.softmax(Ylogits,
name='Yo') # [ BATCHSIZE x FLAGS.seqlen, ALPHASIZE ]
Y = tf.argmax(Yo, 1) # [ BATCHSIZE x FLAGS.seqlen ]
Y = tf.reshape(Y, [batchsize, -1], name="Y") # [ BATCHSIZE, FLAGS.seqlen ]
train_step = tf.train.AdamOptimizer(lr).minimize(loss)
# stats for display
seqloss = tf.reduce_mean(loss, 1)
batchloss = tf.reduce_mean(seqloss)
accuracy = tf.reduce_mean(
tf.cast(tf.equal(Y_, tf.cast(Y, tf.uint8)), tf.float32))
loss_summary = tf.summary.scalar("batch_loss", batchloss)
acc_summary = tf.summary.scalar("batch_accuracy", accuracy)
summaries = tf.summary.merge([loss_summary, acc_summary])
# Init Tensorboard stuff. This will save Tensorboard information into a different
# folder at each run named 'log/<timestamp>/'. Two sets of data are saved so that
# you can compare training and validation curves visually in Tensorboard.
timestamp = str(math.trunc(time.time()))
summary_writer = tf.summary.FileWriter(
os.path.join(FLAGS.summaries_dir, timestamp + "-training"))
validation_writer = tf.summary.FileWriter(
os.path.join(FLAGS.summaries_dir, timestamp + "-validation"))
# Init for saving models. They will be saved into a directory named 'checkpoints'.
# Only the last checkpoint is kept.
if not os.path.exists(FLAGS.checkpoint_dir):
os.mkdir(FLAGS.checkpoint_dir)
saver = tf.train.Saver(max_to_keep=1)
# for display: init the progress bar
DISPLAY_FREQ = 50
_50_BATCHES = DISPLAY_FREQ * FLAGS.train_batch_size * FLAGS.seqlen
progress = txt.Progress(DISPLAY_FREQ,
size=111 + 2,
msg="Training on next " + str(DISPLAY_FREQ) +
" batches")
# init
istate = np.zeros([FLAGS.train_batch_size,
INTERNALSIZE * NLAYERS]) # initial zero input state
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
step = 0
# training loop
for x, y_, epoch in txt.rnn_minibatch_sequencer(codetext,
FLAGS.train_batch_size,
FLAGS.seqlen,
nb_epochs=1000):
# train on one minibatch
feed_dict = {
X: x,
Y_: y_,
Hin: istate,
lr: FLAGS.learning_rate,
pkeep: FLAGS.dropout_pkeep,
batchsize: FLAGS.train_batch_size
}
_, y, ostate, smm = sess.run([train_step, Y, H, summaries],
feed_dict=feed_dict)
# save training data for Tensorboard
summary_writer.add_summary(smm, step)
# display a visual validation of progress (every 50 batches)
if step % _50_BATCHES == 0:
feed_dict = {
X: x,
Y_: y_,
Hin: istate,
pkeep: 1.0,
batchsize: FLAGS.train_batch_size
} # no dropout for validation
y, l, bl, acc = sess.run([Y, seqloss, batchloss, accuracy],
feed_dict=feed_dict)
txt.print_learning_learned_comparison(x, y, l, bookranges, bl, acc,
epoch_size, step, epoch)
# run a validation step every 50 batches
# The validation text should be a single sequence but that's too slow (1s per 1024 chars!),
# so we cut it up and batch the pieces (slightly inaccurate)
# tested: validating with 5K sequences instead of 1K is only slightly more accurate, but a lot slower.
if step % _50_BATCHES == 0 and len(valitext) > 0:
VALI_SEQLEN = 1 * 1024 # Sequence length for validation. State will be wrong at the start of each sequence.
bsize = len(valitext) // VALI_SEQLEN
txt.print_validation_header(len(codetext), bookranges)
vali_x, vali_y, _ = next(
txt.rnn_minibatch_sequencer(valitext, bsize, VALI_SEQLEN,
1)) # all data in 1 batch
vali_nullstate = np.zeros([bsize, INTERNALSIZE * NLAYERS])
feed_dict = {
X: vali_x,
Y_: vali_y,
Hin: vali_nullstate,
pkeep: 1.0, # no dropout for validation
batchsize: bsize
}
ls, acc, smm = sess.run([batchloss, accuracy, summaries],
feed_dict=feed_dict)
txt.print_validation_stats(ls, acc)
# save validation data for Tensorboard
validation_writer.add_summary(smm, step)
# display a short text generated with the current weights and biases (every 150 batches)
if step // 3 % _50_BATCHES == 0:
txt.print_text_generation_header()
ry = np.array([[txt.convert_from_alphabet(ord("K"))]])
rh = np.zeros([1, INTERNALSIZE * NLAYERS])
for k in range(1000):
ryo, rh = sess.run([Yo, H],
feed_dict={
X: ry,
pkeep: 1.0,
Hin: rh,
batchsize: 1
})
rc = txt.sample_from_probabilities(
ryo, topn=10 if epoch <= 1 else 2)
print(chr(txt.convert_to_alphabet(rc)), end="")
ry = np.array([[rc]])
txt.print_text_generation_footer()
# save a checkpoint (every 500 batches)
if step // 10 % _50_BATCHES == 0:
saver.save(sess,
FLAGS.checkpoint_dir + '/rnn_train_' + timestamp,
global_step=step)
# display progress bar
progress.step(reset=step % _50_BATCHES == 0)
# loop state around
istate = ostate
step += FLAGS.train_batch_size * FLAGS.seqlen
# option in tf.nn.rnn_cell.MultiRNNCell. This option is enabled by default. # It produces faster code (by ~10%) but handling the state as a tuple is bit # more cumbersome. Search for comments containing "state_is_tuple=True" for # details. SEQLEN = 30 BATCHSIZE = 100 ALPHASIZE = txt.ALPHASIZE INTERNALSIZE = 512 NLAYERS = 3 learning_rate = 0.001 # fixed learning rate # load data, either shakespeare, or the Python source of Tensorflow itself shakedir = "shakespeare/*.txt" charlesdir = "charles/*.txt" codetext, valitext, bookranges = txt.read_data_files(shakedir, validation=False) codetext2, valitext2, bookranges2 = txt.read_data_files(charlesdir, 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 ]
print("-"*80)
print("Model parameters:")
print("Seqlen {:>5d}".format(SEQLEN))
print("BatchSize {:>5d}".format(BATCHSIZE))
print("AlphaSize {:>5d}".format(ALPHASIZE))
print("InternalSize {:>5d}".format(INTERNALSIZE))
print("NLayers {:>5d}".format(NLAYERS))
print("Learning Rate {:>5.3f}".format(learning_rate))
print("Dropout PKeep {:>5.1f}".format(dropout_pkeep))
print("Num. epochs {:>5d}".format(args.epochs))
print("-"*80)
print("")
read_data_time = datetime.now()
codetext, valitext, bookranges = txt.read_data_files(args.input_dir, validation=args.novalidate,
validation_percentage = args.validation_percentage)
print_elapsed_time("reading data files", read_data_time)
# display some stats on the data
epoch_size = len(codetext) // (BATCHSIZE * SEQLEN)
txt.print_data_stats(len(codetext), len(valitext), epoch_size)
#
# the model (see FAQ in README.md)
#
lr = tf.placeholder(tf.float32, name='lr') # learning rate
pkeep = tf.placeholder(tf.float32, name='pkeep') # dropout parameter
batchsize = tf.placeholder(tf.int32, name='batchsize')
# inputs
X = tf.placeholder(tf.uint8, [None, None], name='X') # [ BATCHSIZE, SEQLEN ]
