TensorFlow Sequence Classification

blog_tensorflow_sequence_classification.py

# Example for my blog post at:
# https://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 SequenceClassification:

    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.contrib.rnn.GRUCell(self._num_hidden)
        network = tf.contrib.rnn.DropoutWrapper(
            network, output_keep_prob=self.dropout)
        network = tf.contrib.rnn.MultiRNNCell([network] * self._num_layers)
        output, _ = tf.nn.dynamic_rnn(network, self.data, dtype=tf.float32)
        # Select last output.
        output = tf.transpose(output, [1, 0, 2])
        last = tf.gather(output, int(output.get_shape()[0]) - 1)
        # Softmax layer.
        weight, bias = self._weight_and_bias(
            self._num_hidden, int(self.target.get_shape()[1]))
        prediction = tf.nn.softmax(tf.matmul(last, weight) + bias)
        return prediction

    @lazy_property
    def cost(self):
        cross_entropy = -tf.reduce_sum(self.target * tf.log(self.prediction))
        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, 1), tf.argmax(self.prediction, 1))
        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 main():
    # We treat images as sequences of pixel rows.
    train, test = sets.Mnist()
    _, rows, row_size = train.data.shape
    num_classes = train.target.shape[1]
    data = tf.placeholder(tf.float32, [None, rows, row_size])
    target = tf.placeholder(tf.float32, [None, num_classes])
    dropout = tf.placeholder(tf.float32)
    model = SequenceClassification(data, target, dropout)
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())
    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))


if __name__ == '__main__':
    main()

And another thing I've been wondering about: shouldn't the data in the following line be self.data instead?
output, _ = rnn.dynamic_rnn(network, data, dtype=tf.float32)

@wirth6 Sorry for the taking so long. @lazy_property causes the method to act like a property, so you can access it without parentheses. Moreover, the function is only evaluated once, when it's accessed for the first time. The result is stored an directly returned for later accesses. This is useful since we don't want to create this part of the TensorFlow graph once, but access the resulting tensor multiple times. For more information, please refer to my post Structuring Your TensorFlow Models.

Regarding your second question, you're right. It unexpectedly worked anyways since the __name__ == '__main__' block created a global data. In the updated example, I moved that code into a main() function so that can't happen anymore.

Great tutorial....
One problem I was facing when I was trying on other data set is when I added sequence_length to dynamic_rnn it was not training.
output, _ = tf.nn.dynamic_rnn(network, self.data, sequence_length=self.seq_len, dtype=tf.float32)
Can you please suggest me what is going wrong.

Hello I have a problem in line train, test = sets.Mnist() I found that I must install sets first as your product in here https://github.com/danijar/sets . @danijar Can you tell me how to config and install it?, Please I need help. I have successfully install it but still failed in import the modul Mnist() . I hope that you can create the requirement first for your tutorial so I can know what should I install before start to learn. The error is ImportError: cannot import name Mnist while I am using sudo pip install sets==0.3.2 to install it.
-Thanks-

Hi, great blog post, thank you!

Shouldn't lines 35-38 be replaced with

network = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.GRUCell(self._num_hidden) for _ in range(self._num_layers)])
network = tf.contrib.rnn.DropoutWrapper(network, output_keep_prob=1 - self.dropout)

to fix dropout (which is currently reversed) and to give each layer its own weights?

I tried to do this with my own data, which is a numpy array with 30 rows and each row has a vector of 20 elements. However, the skilearn datasets are very different from my own so I was not able to adapt to this code. The main error comes from data.shape and target.shape. when I delete the data and target and manually define the num_classes, the next error comes when I try to run session and it gives me this error: AttributeError: 'numpy.ndarray' object has no attribute 'target'

I'm trying to learn tensorflow and am running version 1.2.1. Every time I try to run any example with a construction like:
network = tf.contrib.rnn.GRUCell(self._num_hidden)
network = tf.contrib.rnn.DropoutWrapper(
network, output_keep_prob=self.dropout)
network = tf.contrib.rnn.MultiRNNCell([network] * self._num_layers)

I get an error like:
ValueError: Trying to share variable rnn/multi_rnn_cell/cell_0/gru_cell/gates/kernel, but specified shape (400, 400) and found shape (228, 400).

If I remove the DropoutWrapper() call, everything is fine. I can't seem to find any example that doesn't do this, so I'm having trouble learning how to construct a cell with dropout.

hey Joelkr Instead of using
network = tf.contrib.rnn.GRUCell(self._num_hidden)
network = tf.contrib.rnn.DropoutWrapper(
network, output_keep_prob=self.dropout)
network = tf.contrib.rnn.MultiRNNCell([network] * self._num_layers)

define each cell separately and then join them together so they dont share variables like the following :
cells = []
for _ in range(self._num_layers):
cell = tf.contrib.rnn.GRUCell(self._num_hidden) # Or LSTMCell(num_units)
cell = tf.contrib.rnn.DropoutWrapper(
cell, output_keep_prob=1.0 - self.dropout)
cells.append(cell)
network = tf.contrib.rnn.MultiRNNCell(cells)

Link is dead ?
https://danijar.com/introduction-to-recurrent-networks-in-tensorflow/

Hey guys, i saw last return zero tensor, not real last hidden.

        output, _ = tf.nn.dynamic_rnn(network, self.data, dtype=tf.float32)
        # Select last output.
        output = tf.transpose(output, [1, 0, 2])
        last = tf.gather(output, int(output.get_shape()[0]) - 1)

@danijar

I have two more questions about this project, hoping that someone who's still reading this blog would be able to answer.

1. For a while I was sure about this, but now I'm a bit worried: do I understand it correctly, that one "row" in data represent a feature vector at time t, and the next row represents a feature vector at time t+1?
2. Would it be possible to include an embedding layer below the GRU layers that would transform rows of binaries with high dimensionality (size) to a lower dimensionality, and if so, how?

Thank you in advance for any help.