BRNN + perplexity evaluation

eval.py

@@ -0,0 +1,47 @@
+from __future__ import print_function
+import numpy as np
+import tensorflow as tf
+
+import argparse
+import codecs
+import time
+import os
+from six.moves import cPickle
+
+from utils import TextLoader
+from model import Model
+
+from six import text_type
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--save_dir', type=str, default='save',
+                       help='model directory to store checkpointed models')
+    parser.add_argument('--text', type=str,
+                        help='filename of text to evaluate on')
+    args = parser.parse_args()
+    eval(args)
+
+def eval(args):
+    with open(os.path.join(args.save_dir, 'config.pkl'), 'rb') as f:
+        saved_args = cPickle.load(f)
+    saved_args.batch_size = 1
+    saved_args.seq_length = 200
+    with open(os.path.join(args.save_dir, 'chars_vocab.pkl'), 'rb') as f:
+        chars, vocab = cPickle.load(f)
+    model = Model(saved_args, False)
+
+    with codecs.open(args.text, 'r', encoding='utf-8') as f:
+        text = f.read()
+
+    with tf.Session() as sess:
+        tf.initialize_all_variables().run()
+        saver = tf.train.Saver(tf.all_variables())
+        ckpt = tf.train.get_checkpoint_state(args.save_dir)
+        if ckpt and ckpt.model_checkpoint_path:
+            saver.restore(sess, ckpt.model_checkpoint_path)
+            ppl = model.eval(sess, chars, vocab, text)
+            print('perplexity: {0}'.format(ppl))
+
+if __name__ == '__main__':
+    main()

model.py

@@ -1,5 +1,6 @@
 import tensorflow as tf
 from tensorflow.python.ops import rnn_cell
+from tensorflow.python.ops import rnn
 from tensorflow.python.ops import seq2seq
 
 import numpy as np
@@ -10,61 +11,120 @@ def __init__(self, args, infer=False):
         if infer:
             args.batch_size = 1
             args.seq_length = 1
-
-        if args.model == 'rnn':
+        if args.cell == 'rnn':
             cell_fn = rnn_cell.BasicRNNCell
-        elif args.model == 'gru':
+        elif args.cell == 'gru':
             cell_fn = rnn_cell.GRUCell
-        elif args.model == 'lstm':
+        elif args.cell == 'lstm':
             cell_fn = rnn_cell.BasicLSTMCell
         else:
             raise Exception("model type not supported: {}".format(args.model))
 
-        cell = cell_fn(args.rnn_size, state_is_tuple=True)
-
-        self.cell = cell = rnn_cell.MultiRNNCell([cell] * args.num_layers, state_is_tuple=True)
-
-        self.input_data = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
-        self.targets = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
-        self.initial_state = cell.zero_state(args.batch_size, tf.float32)
-
-        with tf.variable_scope('rnnlm'):
-            softmax_w = tf.get_variable("softmax_w", [args.rnn_size, args.vocab_size])
-            softmax_b = tf.get_variable("softmax_b", [args.vocab_size])
-            with tf.device("/cpu:0"):
-                embedding = tf.get_variable("embedding", [args.vocab_size, args.rnn_size])
-                inputs = tf.split(1, args.seq_length, tf.nn.embedding_lookup(embedding, self.input_data))
-                inputs = [tf.squeeze(input_, [1]) for input_ in inputs]
-
-        def loop(prev, _):
-            prev = tf.matmul(prev, softmax_w) + softmax_b
-            prev_symbol = tf.stop_gradient(tf.argmax(prev, 1))
-            return tf.nn.embedding_lookup(embedding, prev_symbol)
+        def build_unirnn(args):
+            cell = cell_fn(args.rnn_size, state_is_tuple=True)
+
+            self.cell = cell = rnn_cell.MultiRNNCell([cell] * args.num_layers, state_is_tuple=True)
+            self.cell2 = cell_dummy = cell_fn(args.rnn_size, state_is_tuple=True)
+
+            self.input_data = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
+            self.targets = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
+            self.initial_state = cell.zero_state(args.batch_size, tf.float32)
+            self.initial_state2 = cell_dummy.zero_state(args.batch_size, tf.float32)
+
+            with tf.variable_scope('rnnlm'):
+                softmax_w = tf.get_variable("softmax_w", [args.rnn_size, args.vocab_size])
+                softmax_b = tf.get_variable("softmax_b", [args.vocab_size])
+                with tf.device("/cpu:0"):
+                    embedding = tf.get_variable("embedding", [args.vocab_size, args.rnn_size])
+                    input_embeddings = tf.nn.embedding_lookup(embedding, self.input_data)
+                    inputs = tf.unpack(input_embeddings, axis=1)
+            def loop(prev, _):
+                prev = tf.matmul(prev, softmax_w) + softmax_b
+                prev_symbol = tf.stop_gradient(tf.argmax(prev, 1))
+                return tf.nn.embedding_lookup(embedding, prev_symbol)
+
+            outputs, self.last_state = seq2seq.rnn_decoder(inputs, self.initial_state, cell, loop_function=loop if infer else None, scope='rnnlm')
+            # dummy state in this case
+            self.last_state2 = cell_dummy.zero_state(args.batch_size, tf.float32)
+            output = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size])
+            self.logits = tf.matmul(output, softmax_w) + softmax_b
+
+        def build_birnn(args):
+            self.cell = fw_cell = cell_fn(args.rnn_size, state_is_tuple=True)
+            self.cell2 = bw_cell = cell_fn(args.rnn_size, state_is_tuple=True)
+
+            self.input_data = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
+            self.targets = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
+            self.initial_state = fw_cell.zero_state(args.batch_size, tf.float32)
+            self.initial_state2 = bw_cell.zero_state(args.batch_size, tf.float32)
+
+            with tf.variable_scope('rnnlm'):
+                softmax_w = tf.get_variable("softmax_w", [2*args.rnn_size, args.vocab_size])
+                softmax_b = tf.get_variable("softmax_b", [args.vocab_size])
+                with tf.device("/cpu:0"):
+                    embedding = tf.get_variable("embedding", [args.vocab_size, args.rnn_size])
+                    input_embeddings = tf.nn.embedding_lookup(embedding, self.input_data)
+                    inputs = tf.unpack(input_embeddings, axis=1)
+
+            outputs, self.last_state, self.last_state2 = rnn.bidirectional_rnn(fw_cell,
+                                                                               bw_cell,
+                                                                               inputs,
+                                                                               initial_state_fw=self.initial_state,
+                                                                               initial_state_bw=self.initial_state2,
+                                                                               dtype=tf.float32)
+            output = tf.reshape(tf.concat(1, outputs), [-1, 2*args.rnn_size])
+            self.logits = tf.matmul(output, softmax_w) + softmax_b
+
+        if args.model == 'uni':
+            build_unirnn(args)
+        else:
+            build_birnn(args)
 
-        outputs, last_state = seq2seq.rnn_decoder(inputs, self.initial_state, cell, loop_function=loop if infer else None, scope='rnnlm')
-        output = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size])
-        self.logits = tf.matmul(output, softmax_w) + softmax_b
         self.probs = tf.nn.softmax(self.logits)
-        loss = seq2seq.sequence_loss_by_example([self.logits],
+        self.loss = seq2seq.sequence_loss_by_example([self.logits],
                 [tf.reshape(self.targets, [-1])],
                 [tf.ones([args.batch_size * args.seq_length])],
                 args.vocab_size)
-        self.cost = tf.reduce_sum(loss) / args.batch_size / args.seq_length
-        self.final_state = last_state
+        self.cost = tf.reduce_sum(self.loss) / args.batch_size / args.seq_length
         self.lr = tf.Variable(0.0, trainable=False)
         tvars = tf.trainable_variables()
         grads, _ = tf.clip_by_global_norm(tf.gradients(self.cost, tvars),
                 args.grad_clip)
         optimizer = tf.train.AdamOptimizer(self.lr)
         self.train_op = optimizer.apply_gradients(zip(grads, tvars))
 
+    def eval(self, sess, chars, vocab, text):
+        batch_size = 200
+        state = sess.run(self.cell.zero_state(1, tf.float32))
+        state2 = sess.run(self.cell2.zero_state(1, tf.float32))
+        x = [vocab[c] if c in vocab else vocab['UNK'] for c in text]
+        x = [vocab['<S>']] + x + [vocab['</S>']]
+        total_len = len(x) - 1
+        # pad x so the batch_size divides it
+        while len(x) % 200 != 1:
+            x.append(vocab[' '])
+        y = np.array(x[1:]).reshape((-1, batch_size))
+        x = np.array(x[:-1]).reshape((-1, batch_size))
+
+        total_loss = 0.0
+        for i in range(x.shape[0]):
+            feed = {self.input_data: x[i:i+1, :], self.targets: y[i:i+1, :],
+                    self.initial_state: state, self.initial_state2: state2}
+            [state, state2, loss] = sess.run([self.last_state, self.last_state2, self.loss], feed)
+            total_loss += loss.sum()
+        # need to subtract off loss from padding tokens
+        total_loss -= loss[total_len % batch_size - batch_size:].sum()
+        avg_entropy = total_loss / len(text)
+        return np.exp(avg_entropy)  # this is the perplexity
+
     def sample(self, sess, chars, vocab, num=200, prime='The ', sampling_type=1):
         state = sess.run(self.cell.zero_state(1, tf.float32))
+        state2 = sess.run(self.cell2.zero_state(1, tf.float32))
         for char in prime[:-1]:
             x = np.zeros((1, 1))
             x[0, 0] = vocab[char]
-            feed = {self.input_data: x, self.initial_state:state}
-            [state] = sess.run([self.final_state], feed)
+            feed = {self.input_data: x, self.initial_state: state, self.initial_state2: state2}
+            [state, state2] = sess.run([self.last_state, self.last_state2], feed)
 
         def weighted_pick(weights):
             t = np.cumsum(weights)
@@ -77,7 +137,7 @@ def weighted_pick(weights):
             x = np.zeros((1, 1))
             x[0, 0] = vocab[char]
             feed = {self.input_data: x, self.initial_state:state}
-            [probs, state] = sess.run([self.probs, self.final_state], feed)
+            [probs, state, state2] = sess.run([self.probs, self.last_state, self.last_state2], feed)
             p = probs[0]
 
             if sampling_type == 0:

train.py

@@ -20,8 +20,10 @@ def main():
                        help='size of RNN hidden state')
     parser.add_argument('--num_layers', type=int, default=2,
                        help='number of layers in the RNN')
-    parser.add_argument('--model', type=str, default='lstm',
+    parser.add_argument('--cell', type=str, default='lstm',
                        help='rnn, gru, or lstm')
+    parser.add_argument('--model', type=str, default='uni',
+                       help='uni, or bi')
     parser.add_argument('--batch_size', type=int, default=50,
                        help='minibatch size')
     parser.add_argument('--seq_length', type=int, default=50,
@@ -91,14 +93,15 @@ def train(args):
             sess.run(tf.assign(model.lr, args.learning_rate * (args.decay_rate ** e)))
             data_loader.reset_batch_pointer()
             state = sess.run(model.initial_state)
+            state2 = sess.run(model.initial_state2)
             for b in range(data_loader.num_batches):
                 start = time.time()
                 x, y = data_loader.next_batch()
-                feed = {model.input_data: x, model.targets: y}
-                for i, (c, h) in enumerate(model.initial_state):
-                    feed[c] = state[i].c
-                    feed[h] = state[i].h
-                train_loss, state, _ = sess.run([model.cost, model.final_state, model.train_op], feed)
+                if b == 0 and e == 0:
+                    feed = {model.input_data: x, model.targets: y}
+                else:
+                    feed = {model.input_data: x, model.targets: y, model.initial_state: state, model.initial_state2: state2}
+                train_loss, state, state2, _ = sess.run([model.cost, model.last_state, model.last_state2, model.train_op], feed)
                 end = time.time()
                 print("{}/{} (epoch {}), train_loss = {:.3f}, time/batch = {:.3f}" \
                     .format(e * data_loader.num_batches + b,

utils.py

@@ -4,6 +4,7 @@
 from six.moves import cPickle
 import numpy as np
 
+
 class TextLoader():
     def __init__(self, data_dir, batch_size, seq_length, encoding='utf-8'):
         self.data_dir = data_dir
@@ -28,13 +29,14 @@ def preprocess(self, input_file, vocab_file, tensor_file):
         with codecs.open(input_file, "r", encoding=self.encoding) as f:
             data = f.read()
         counter = collections.Counter(data)
+        counter.update(('<S>', '</S>', 'UNK'))  # add tokens for start end and unk
         count_pairs = sorted(counter.items(), key=lambda x: -x[1])
         self.chars, _ = zip(*count_pairs)
         self.vocab_size = len(self.chars)
         self.vocab = dict(zip(self.chars, range(len(self.chars))))
         with open(vocab_file, 'wb') as f:
             cPickle.dump(self.chars, f)
-        self.tensor = np.array(list(map(self.vocab.get, data)))
+        self.tensor = np.array(list(map(self.vocab.get, ['<S>'] + list(data) + ['</S>'])))
         np.save(tensor_file, self.tensor)
 
     def load_preprocessed(self, vocab_file, tensor_file):