TensorFlow Sequence Labelling

blog_tensorflow_sequence_labelling.py

# Example for my blog post at:
# http://danijar.com/introduction-to-recurrent-networks-in-tensorflow/
import functools
import sets
import tensorflow as tf


def lazy_property(function):
    attribute = '_' + function.__name__

    @property
    @functools.wraps(function)
    def wrapper(self):
        if not hasattr(self, attribute):
            setattr(self, attribute, function(self))
        return getattr(self, attribute)
    return wrapper


class SequenceLabelling:

    def __init__(self, data, target, dropout, num_hidden=200, num_layers=3):
        self.data = data
        self.target = target
        self.dropout = dropout
        self._num_hidden = num_hidden
        self._num_layers = num_layers
        self.prediction
        self.error
        self.optimize

    @lazy_property
    def prediction(self):
        # Recurrent network.
        network = tf.nn.rnn_cell.GRUCell(self._num_hidden)
        network = tf.nn.rnn_cell.DropoutWrapper(
            network, output_keep_prob=self.dropout)
        network = tf.nn.rnn_cell.MultiRNNCell([network] * self._num_layers)
        output, _ = tf.nn.dynamic_rnn(network, data, dtype=tf.float32)
        # Softmax layer.
        max_length = int(self.target.get_shape()[1])
        num_classes = int(self.target.get_shape()[2])
        weight, bias = self._weight_and_bias(self._num_hidden, num_classes)
        # Flatten to apply same weights to all time steps.
        output = tf.reshape(output, [-1, self._num_hidden])
        prediction = tf.nn.softmax(tf.matmul(output, weight) + bias)
        prediction = tf.reshape(prediction, [-1, max_length, num_classes])
        return prediction

    @lazy_property
    def cost(self):
        cross_entropy = -tf.reduce_sum(
            self.target * tf.log(self.prediction), [1, 2])
        cross_entropy = tf.reduce_mean(cross_entropy)
        return cross_entropy

    @lazy_property
    def optimize(self):
        learning_rate = 0.003
        optimizer = tf.train.RMSPropOptimizer(learning_rate)
        return optimizer.minimize(self.cost)

    @lazy_property
    def error(self):
        mistakes = tf.not_equal(
            tf.argmax(self.target, 2), tf.argmax(self.prediction, 2))
        return tf.reduce_mean(tf.cast(mistakes, tf.float32))

    @staticmethod
    def _weight_and_bias(in_size, out_size):
        weight = tf.truncated_normal([in_size, out_size], stddev=0.01)
        bias = tf.constant(0.1, shape=[out_size])
        return tf.Variable(weight), tf.Variable(bias)


def read_dataset():
    dataset = sets.Ocr()
    dataset = sets.OneHot(dataset.target, depth=2)(dataset, columns=['target'])
    dataset['data'] = dataset.data.reshape(
        dataset.data.shape[:-2] + (-1,)).astype(float)
    train, test = sets.Split(0.66)(dataset)
    return train, test


if __name__ == '__main__':
    train, test = read_dataset()
    _, length, image_size = train.data.shape
    num_classes = train.target.shape[2]
    data = tf.placeholder(tf.float32, [None, length, image_size])
    target = tf.placeholder(tf.float32, [None, length, num_classes])
    dropout = tf.placeholder(tf.float32)
    model = SequenceLabelling(data, target, dropout)
    sess = tf.Session()
    sess.run(tf.initialize_all_variables())
    for epoch in range(10):
        for _ in range(100):
            batch = train.sample(10)
            sess.run(model.optimize, {
                data: batch.data, target: batch.target, dropout: 0.5})
        error = sess.run(model.error, {
            data: test.data, target: test.target, dropout: 1})
        print('Epoch {:2d} error {:3.1f}%'.format(epoch + 1, 100 * error))

facing problem in line line 107
dataset = sets.Ocr()
AttributeError: 'module' object has no attribute 'Ocr'

running this gist (after installing your sets package) I get an error around 40% but I don't understand why. Looking at the ocr data from sets it looks like the input just says gommandin over and over, with a -1 in the column [2] when the sequence repeats, nonetheless the network seems incapable of recognizing this. It seems like the network should have 2 very good chances to solve this problem (once by memorizing gommandin, and once by actually learning the ocr letters) but at least it fails quite spectacularly on my system. I've tried fiddling with the parameters (epoch, learning, dropout etc...) but nothing seems to have a good effect. Any ideas as to what I'm missing or misunderstanding?

@pfaucon I'm learning the RNN too. From the official tensorflow tutorial, they said "If you want to do optimization to minimize the cross entropy, AND you're softmaxing after your last layer, you should use tf.nn.softmax_cross_entropy_with_logits instead of doing it yourself, because it covers numerically unstable corner cases in the mathematically right way." I think maybe that's the reason for your error.

@adelsalehali1982 It works for me with the newest TensorFlow and sets versions.

@mrathi12 Take a look at the placeholders to see the input format. It's a tensor of shape batch_size x sequence_length x image_size for the sequences and a tensor of shape batch_size x sequence_length x num_classes for the targets. You can pass nested Python lists or Numpy arrays into the sess.run() call.

@pfaucon The parameters are not tuned at all, but the RNN starts learning and everything runs through:

Epoch  1 error 65.2%
Epoch  2 error 55.6%
Epoch  3 error 50.6%
Epoch  4 error 48.6%
Epoch  5 error 47.0%
Epoch  6 error 44.3%
Epoch  7 error 44.2%
Epoch  8 error 43.4%
Epoch  9 error 42.6%
Epoch 10 error 42.9%

@YiruS I implement the activation and cost manually so it's easier to understand what's going on. But I'd worth pointing out tf.nn.softmax_cross_entropy_with_logits() for performance and numerical stability reasons.

excuse me,i 'm a new hand in tensorflow, thanks for your bolg which helps me a lot. And i just want to know how can i save the model i trained and use it in another job? looking forward to your reply!

@YiruS: I find this a bit confusing, as the description of the softmax_cross_entropy_with_logits function states "This op expects unscaled logits, since it performs a softmax on logits internally for efficiency. Do not call this op with the output of softmax, as it will produce incorrect results." and here self.prediction is the output of a softmax, so according to the documentation we shouldn't use this function here.

Hi, Thanks very much for the example. It helps me a lot for creating my own training model.
I don't fully understand the setting of weight and bias. Is that only 1 weight and 1 bias are needed even for multiple layers network?

@adelsalehali1982 You may want to install [sets] like this:

$ pip install -e git+https://github.com/danijar/sets/#egg=sets

Now, [AttributeError: 'module' object has no attribute 'Ocr'] problem should disappear.

Thank you for the post.
I am learning RNN but I have a ValueError when I run this code:

I have a hard time to debug this. Could you please help with it?

ValueError: Dimensions must be equal, but are 400 and 328 for 'rnn/while/rnn/multi_rnn_cell/cell_0/cell_0/gru_cell/MatMul_2' (op: 'MatMul') with input shapes: [?,400], [328,400].

@zdarktknight try the below. I refactored the original implementation and commented heavily for easy understanding. Danijar's sets library makes the job super easy as allows us to focus purely on understanding the algorithm. Good luck.
https://github.com/surfertas/deep_learning/blob/master/experiments/1-char_sequence_labeling_lstm.ipynb

@QoT, @adelsalehali1982 I suspect it is due to the name shadowing issue with the standard python library. I'm facing the same problem here. No idea how every other guys run the code successfully...I'm using python 2.7.13, and virtualenv, in a virtual environment dedicated for tensorflow. Running the following command tries to compile danijar's sets library from the source, am I guessing right?

$ pip install -e git+https://github.com/danijar/sets/#egg=sets

Frustratingly, it gives me the following error:

      ...
      File "/Users/jx/ProgramData/tensorflow/src/sets/setup.py", line 94, in finalize_options
        super().finalize_options()
    TypeError: super() takes at least 1 argument (0 given)
    
    ----------------------------------------
  Rolling back uninstall of sets
Command "/Users/jx/ProgramData/tensorflow/bin/python2.7 -c "import setuptools, 
tokenize;__file__='/Users/jx/ProgramData/tensorflow/src/sets/setup.py';f=getattr(tokenize, 'open', open)
(__file__);code=f.read().replace('\r\n', '\n');f.close();exec(compile(code, __file__, 'exec'))" develop --no-deps" failed with error code 1 in 
/Users/jx/ProgramData/tensorflow/src/sets/

@sufertas, What you did is awesome. Thank you!!