rnn_train.py

# encoding: UTF-8
# Copyright 2017 Google.com
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import tensorflow as tf
from tensorflow.contrib import layers
from tensorflow.contrib import rnn  # rnn stuff temporarily in contrib, moving back to code in TF 1.1
import os
import time
import math
import numpy as np
import my_txtutils as txt
tf.set_random_seed(0)

# model parameters
#
# Usage:
#   Training only:
#         Leave all the parameters as they are
#         Disable validation to run a bit faster (set validation=False below)
#         You can follow progress in Tensorboard: tensorboard --log-dir=log
#   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
#         training data (shakedir = "shakespeare/t*.txt" for example)
#
SEQLEN = 30
BATCHSIZE = 150
ALPHASIZE = txt.ALPHASIZE
INTERNALSIZE = 512
NLAYERS = 3
learning_rate = 0.001  # fixed learning rate
dropout_pkeep = 0.8    # some dropout

# load data, either shakespeare, or the Python source of Tensorflow itself
shakedir = "testdata/*.txt"
#shakedir = "../tensorflow/**/*.py"
codetext, valitext, bookranges = txt.read_data_files(shakedir, 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 ]
# 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 ]
# input state
Hin = tf.placeholder(tf.float32, [None, INTERNALSIZE*NLAYERS], name='Hin')  # [ BATCHSIZE, INTERNALSIZE * NLAYERS]

# using a NLAYERS=3 layers of GRU cells, unrolled SEQLEN=30 times
# dynamic_rnn infers SEQLEN from the size of the inputs Xo

# How to properly apply dropout in RNNs: see README.md
cells = [rnn.GRUCell(INTERNALSIZE) for _ in range(NLAYERS)]
# "naive dropout" implementation
dropcells = [rnn.DropoutWrapper(cell,input_keep_prob=pkeep) for cell in cells]
multicell = rnn.MultiRNNCell(dropcells, state_is_tuple=False)
multicell = rnn.DropoutWrapper(multicell, output_keep_prob=pkeep)  # dropout for the softmax layer

Yr, H = tf.nn.dynamic_rnn(multicell, Xo, dtype=tf.float32, initial_state=Hin)
# Yr: [ BATCHSIZE, 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, SEQLEN, ALPHASIZE ] => [ BATCHSIZE x 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 sequence time step or a minibatch item is the same thing.

Yflat = tf.reshape(Yr, [-1, INTERNALSIZE])    # [ BATCHSIZE x SEQLEN, INTERNALSIZE ]
Ylogits = layers.linear(Yflat, ALPHASIZE)     # [ BATCHSIZE x SEQLEN, ALPHASIZE ]
Yflat_ = tf.reshape(Yo_, [-1, ALPHASIZE])     # [ BATCHSIZE x SEQLEN, ALPHASIZE ]
loss = tf.nn.softmax_cross_entropy_with_logits(logits=Ylogits, labels=Yflat_)  # [ BATCHSIZE x SEQLEN ]
loss = tf.reshape(loss, [batchsize, -1])      # [ BATCHSIZE, SEQLEN ]
Yo = tf.nn.softmax(Ylogits, name='Yo')        # [ BATCHSIZE x SEQLEN, ALPHASIZE ]
Y = tf.argmax(Yo, 1)                          # [ BATCHSIZE x SEQLEN ]
Y = tf.reshape(Y, [batchsize, -1], name="Y")  # [ BATCHSIZE, 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("log/" + timestamp + "-training")
validation_writer = tf.summary.FileWriter("log/" + 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("checkpoints"):
    os.mkdir("checkpoints")
saver = tf.train.Saver(max_to_keep=1000)
#write_version=tf.train.SaverDef.V1, max_to_keep=1

# for display: init the progress bar
DISPLAY_FREQ = 50
_50_BATCHES = DISPLAY_FREQ * BATCHSIZE * SEQLEN
progress = txt.Progress(DISPLAY_FREQ, size=111+2, msg="Training on next "+str(DISPLAY_FREQ)+" batches")

# init
istate = np.zeros([BATCHSIZE, 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, BATCHSIZE, SEQLEN, nb_epochs=10):

    # train on one minibatch
    feed_dict = {X: x, Y_: y_, Hin: istate, lr: learning_rate, pkeep: dropout_pkeep, batchsize: BATCHSIZE}
    _, y, ostate = sess.run([train_step, Y, H], feed_dict=feed_dict)

    # log training data for Tensorboard display a mini-batch of sequences (every 50 batches)
    if step % _50_BATCHES == 0:
        feed_dict = {X: x, Y_: y_, Hin: istate, pkeep: 1.0, batchsize: BATCHSIZE}  # no dropout for validation
        y, l, bl, acc, smm = sess.run([Y, seqloss, batchloss, accuracy, summaries], feed_dict=feed_dict)
        txt.print_learning_learned_comparison(x, y, l, bookranges, bl, acc, epoch_size, step, epoch)
        summary_writer.add_summary(smm, step)

    # 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:
        saved_file = saver.save(sess, 'checkpoints/rnn_train_' + timestamp, global_step=step)
        print("Saved file: " + saved_file)

    # display progress bar
    progress.step(reset=step % _50_BATCHES == 0)

    # loop state around
    istate = ostate
    step += BATCHSIZE * SEQLEN

# all runs: SEQLEN = 30, BATCHSIZE = 100, ALPHASIZE = 98, INTERNALSIZE = 512, NLAYERS = 3
# run 1477669632 decaying learning rate 0.001-0.0001-1e7 dropout 0.5: not good
# run 1477670023 lr=0.001 no dropout: very good

# Tensorflow runs:
# 1485434262
#   trained on shakespeare/t*.txt only. Validation on 1K sequences
#   validation loss goes up from step 5M (overfitting because of small dataset)
# 1485436038
#   trained on shakespeare/t*.txt only. Validation on 5K sequences
#   On 5K sequences validation accuracy is slightly higher and loss slightly lower
#   => sequence breaks do introduce inaccuracies but the effect is small
# 1485437956
#   Trained on shakespeare/*.txt. Validation on 1K sequences
#   On this much larger dataset, validation loss still decreasing after 6 epochs (step 35M)
# 1495447371
#   Trained on shakespeare/*.txt no dropout, 30 epochs
#   Validation loss starts going up after 10 epochs (overfitting)
# 1495440473
#   Trained on shakespeare/*.txt "naive dropout" pkeep=0.8, 30 epochs
#   Dropout brings the validation loss under control, preventing it from
#   going up but the effect is small.