Seq2seq (#989)

ArnoldLIULJ · zsdonghao · commit 0fe70a734dd6 · 2019-06-02T12:37:03.000+08:00
* add seq2seq model; add seq2seq test * Revert "add seq2seq model; add seq2seq test" This reverts commit 6401d54. * UNDO last commit * Print list instead of tensor * Add comments * ADD batch testing at Inference * ADD batch testing at Inference * FIX typo and ADD some comments * FIX the Travis CI build * Delete unnecessary files * Delete unnecessary files * Delete unnecessary files
diff --git a/tensorlayer/models/seq2seq.py b/tensorlayer/models/seq2seq.py
@@ -0,0 +1,160 @@
+#! /usr/bin/python
+# -*- coding: utf-8 -*-
+
+import tensorflow as tf
+import tensorlayer as tl
+import numpy as np
+from tensorlayer.models import Model
+from tensorlayer.layers import Dense, Dropout, Input
+from tensorlayer.layers.core import Layer
+
+
+class Seq2seq(Model):
+    """vanilla stacked layer Seq2Seq model.
+
+    Parameters
+    ----------
+    decoder_seq_length: int
+        The length of your target sequence
+    cell_enc : str, tf.function
+        The RNN function cell for your encoder stack, e.g tf.keras.layers.GRUCell
+    cell_dec : str, tf.function
+        The RNN function cell for your decoder stack, e.g. tf.keras.layers.GRUCell
+    n_layer : int
+        The number of your RNN layers for both encoder and decoder block
+    embedding_layer : tl.Layer
+        A embedding layer, e.g. tl.layers.Embedding(vocabulary_size=voc_size, embedding_size=emb_dim)
+    name : str
+        The model name
+    
+    Examples
+    ---------
+    Classify stacked-layer Seq2Seq model, see `chatbot <https://github.com/tensorlayer/seq2seq-chatbot>`__
+
+    Returns
+    -------
+        static stacked-layer Seq2Seq model.
+    """
+
+    def __init__(self, decoder_seq_length, cell_enc, cell_dec, n_units=256, n_layer=3, embedding_layer=None, name=None):
+        super(Seq2seq, self).__init__(name=name)
+        self.embedding_layer = embedding_layer
+        self.vocabulary_size = embedding_layer.vocabulary_size
+        self.embedding_size = embedding_layer.embedding_size
+        self.n_layer = n_layer
+        self.enc_layers = []
+        self.dec_layers = []
+        for i in range(n_layer):
+            if (i == 0):
+                self.enc_layers.append(
+                    tl.layers.RNN(
+                        cell=cell_enc(units=n_units), in_channels=self.embedding_size, return_last_state=True
+                    )
+                )
+            else:
+                self.enc_layers.append(
+                    tl.layers.RNN(cell=cell_enc(units=n_units), in_channels=n_units, return_last_state=True)
+                )
+
+        for i in range(n_layer):
+            if (i == 0):
+                self.dec_layers.append(
+                    tl.layers.RNN(
+                        cell=cell_dec(units=n_units), in_channels=self.embedding_size, return_last_state=True
+                    )
+                )
+            else:
+                self.dec_layers.append(
+                    tl.layers.RNN(cell=cell_dec(units=n_units), in_channels=n_units, return_last_state=True)
+                )
+
+        self.reshape_layer = tl.layers.Reshape([-1, n_units])
+        self.dense_layer = tl.layers.Dense(n_units=self.vocabulary_size, in_channels=n_units)
+        self.reshape_layer_after = tl.layers.Reshape([-1, decoder_seq_length, self.vocabulary_size])
+        self.reshape_layer_individual_sequence = tl.layers.Reshape([-1, 1, self.vocabulary_size])
+
+    def inference(self, encoding, seq_length, start_token, top_n):
+        """Inference mode"""
+        """
+        Parameters
+        ----------
+        encoding : input tensor
+            The source sequences 
+        seq_length : int
+            The expected length of your predicted sequence.
+        start_token : int
+            <SOS> : The token of "start of sequence"
+        top_n : int
+            Random search algorithm based on the top top_n words sorted by the probablity. 
+        """
+        feed_output = self.embedding_layer(encoding[0])
+        state = [None for i in range(self.n_layer)]
+
+        for i in range(self.n_layer):
+            feed_output, state[i] = self.enc_layers[i](feed_output, return_state=True)
+        batch_size = len(encoding[0].numpy())
+        decoding = [[start_token] for i in range(batch_size)]
+        feed_output = self.embedding_layer(decoding)
+        for i in range(self.n_layer):
+            feed_output, state[i] = self.dec_layers[i](feed_output, initial_state=state[i], return_state=True)
+
+        feed_output = self.reshape_layer(feed_output)
+        feed_output = self.dense_layer(feed_output)
+        feed_output = self.reshape_layer_individual_sequence(feed_output)
+        feed_output = tf.argmax(feed_output, -1)
+        # [B, 1]
+        final_output = feed_output
+
+        for i in range(seq_length - 1):
+            feed_output = self.embedding_layer(feed_output)
+            for i in range(self.n_layer):
+                feed_output, state[i] = self.dec_layers[i](feed_output, initial_state=state[i], return_state=True)
+            feed_output = self.reshape_layer(feed_output)
+            feed_output = self.dense_layer(feed_output)
+            feed_output = self.reshape_layer_individual_sequence(feed_output)
+            ori_feed_output = feed_output
+            if (top_n is not None):
+                for k in range(batch_size):
+                    idx = np.argpartition(ori_feed_output[k][0], -top_n)[-top_n:]
+                    probs = [ori_feed_output[k][0][i] for i in idx]
+                    probs = probs / np.sum(probs)
+                    feed_output = np.random.choice(idx, p=probs)
+                    feed_output = tf.convert_to_tensor([[feed_output]], dtype=tf.int64)
+                    if (k == 0):
+                        final_output_temp = feed_output
+                    else:
+                        final_output_temp = tf.concat([final_output_temp, feed_output], 0)
+                feed_output = final_output_temp
+            else:
+                feed_output = tf.argmax(feed_output, -1)
+            final_output = tf.concat([final_output, feed_output], 1)
+
+        return final_output, state
+
+    def forward(self, inputs, seq_length=20, start_token=None, return_state=False, top_n=None):
+
+        state = [None for i in range(self.n_layer)]
+        if (self.is_train):
+            encoding = inputs[0]
+            enc_output = self.embedding_layer(encoding)
+
+            for i in range(self.n_layer):
+                enc_output, state[i] = self.enc_layers[i](enc_output, return_state=True)
+
+            decoding = inputs[1]
+            dec_output = self.embedding_layer(decoding)
+
+            for i in range(self.n_layer):
+                dec_output, state[i] = self.dec_layers[i](dec_output, initial_state=state[i], return_state=True)
+
+            dec_output = self.reshape_layer(dec_output)
+            denser_output = self.dense_layer(dec_output)
+            output = self.reshape_layer_after(denser_output)
+        else:
+            encoding = inputs
+            output, state = self.inference(encoding, seq_length, start_token, top_n)
+
+        if (return_state):
+            return output, state
+        else:
+            return output
diff --git a/tests/models/test_seq2seq_model.py b/tests/models/test_seq2seq_model.py
@@ -0,0 +1,96 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+import os
+import unittest
+
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
+
+import numpy as np
+import tensorflow as tf
+import tensorlayer as tl
+from tqdm import tqdm
+from sklearn.utils import shuffle
+from tensorlayer.models.seq2seq import Seq2seq
+from tests.utils import CustomTestCase
+from tensorlayer.cost import cross_entropy_seq
+
+
+class Model_SEQ2SEQ_Test(CustomTestCase):
+
+    @classmethod
+    def setUpClass(cls):
+
+        cls.batch_size = 16
+
+        cls.vocab_size = 20
+        cls.embedding_size = 32
+        cls.dec_seq_length = 5
+        cls.trainX = np.random.randint(20, size=(50, 6))
+        cls.trainY = np.random.randint(20, size=(50, cls.dec_seq_length + 1))
+        cls.trainY[:, 0] = 0  # start_token == 0
+
+        # Parameters
+        cls.src_len = len(cls.trainX)
+        cls.tgt_len = len(cls.trainY)
+
+        assert cls.src_len == cls.tgt_len
+
+        cls.num_epochs = 100
+        cls.n_step = cls.src_len // cls.batch_size
+
+    @classmethod
+    def tearDownClass(cls):
+        pass
+
+    def test_basic_simpleSeq2Seq(self):
+        model_ = Seq2seq(
+            decoder_seq_length=5,
+            cell_enc=tf.keras.layers.GRUCell,
+            cell_dec=tf.keras.layers.GRUCell,
+            n_layer=3,
+            n_units=128,
+            embedding_layer=tl.layers.Embedding(vocabulary_size=self.vocab_size, embedding_size=self.embedding_size),
+        )
+
+        optimizer = tf.optimizers.Adam(learning_rate=0.001)
+
+        for epoch in range(self.num_epochs):
+            model_.train()
+            trainX, trainY = shuffle(self.trainX, self.trainY)
+            total_loss, n_iter = 0, 0
+            for X, Y in tqdm(tl.iterate.minibatches(inputs=trainX, targets=trainY, batch_size=self.batch_size,
+                                                    shuffle=False), total=self.n_step,
+                             desc='Epoch[{}/{}]'.format(epoch + 1, self.num_epochs), leave=False):
+
+                dec_seq = Y[:, :-1]
+                target_seq = Y[:, 1:]
+
+                with tf.GradientTape() as tape:
+                    ## compute outputs
+                    output = model_(inputs=[X, dec_seq])
+
+                    output = tf.reshape(output, [-1, self.vocab_size])
+
+                    loss = cross_entropy_seq(logits=output, target_seqs=target_seq)
+
+                    grad = tape.gradient(loss, model_.all_weights)
+                    optimizer.apply_gradients(zip(grad, model_.all_weights))
+
+                total_loss += loss
+                n_iter += 1
+
+            model_.eval()
+            test_sample = trainX[0:2, :].tolist()
+
+            top_n = 1
+            for i in range(top_n):
+                prediction = model_([test_sample], seq_length=self.dec_seq_length, start_token=0, top_n=1)
+                print("Prediction: >>>>>  ", prediction, "\n Target: >>>>>  ", trainY[0:2, 1:], "\n\n")
+
+            # printing average loss after every epoch
+            print('Epoch [{}/{}]: loss {:.4f}'.format(epoch + 1, self.num_epochs, total_loss / n_iter))
+
+
+if __name__ == '__main__':
+    unittest.main()
