Merge pull request PaddlePaddle#31 from AIPG/wojtuss/bilstm

bilstm topology added to text_classification model (Senta)

fluid/text_classification/infer.py

@@ -1,7 +1,4 @@
-import sys
 import time
-import unittest
-import contextlib
 import numpy as np
 import argparse
 
@@ -17,7 +14,13 @@ def parse_args():
         '--batch_size',
         type=int,
         default=128,
-        help='The size of a batch. (default: %(default)d, usually: 128 for "bow" and "gru", 4 for "cnn" and "lstm")')
+        help='The size of a batch. (default: %(default)d, usually: 128 for "bow" and "gru", 4 for "cnn", "lstm" and "bilstm").')
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default='imdb',
+        choices=['imdb', 'data'],
+        help="Dataset to be used: 'imdb' or 'data' (from 'data' subdirectory).")
     parser.add_argument(
         '--device',
         type=str,
@@ -77,7 +80,7 @@ def infer(args):
         wpses = [0] * total_passes
         acces = [0] * total_passes
         word_dict, train_reader, test_reader = utils.prepare_data(
-            "imdb", self_dict=False, batch_size=args.batch_size,
+            args.dataset, self_dict=False, batch_size=args.batch_size,
             buf_size=50000)
         pass_acc = 0.0
         pass_data_len = 0
@@ -100,6 +103,7 @@ def infer(args):
                             fetch_list=fetch_targets,
                             return_numpy=True)
                 batch_time = time.time() - start
+                # TODO: add class accuracy measurement as in Senta
                 word_count = len([w for d in data for w in d[0]])
                 batch_times[pass_id] += batch_time
                 word_counts[pass_id] += word_count

fluid/text_classification/nets.py

@@ -93,6 +93,55 @@ def lstm_net(data,
     return avg_cost, acc, prediction
 
 
+def bilstm_net(data,
+               label,
+               dict_dim,
+               emb_dim=128,
+               hid_dim=128,
+               hid_dim2=96,
+               class_dim=2,
+               emb_lr=30.0):
+    """
+    Bi-Lstm net
+    """
+    # embedding layer
+    emb = fluid.layers.embedding(
+        input=data,
+        size=[dict_dim, emb_dim],
+        param_attr=fluid.ParamAttr(learning_rate=emb_lr))
+
+    # bi-lstm layer
+    fc0 = fluid.layers.fc(input=emb, size=hid_dim * 4)
+
+    rfc0 = fluid.layers.fc(input=emb, size=hid_dim * 4)
+
+    lstm_h, c = fluid.layers.dynamic_lstm(
+        input=fc0, size=hid_dim * 4, is_reverse=False)
+
+    rlstm_h, c = fluid.layers.dynamic_lstm(
+        input=rfc0, size=hid_dim * 4, is_reverse=True)
+
+    # extract last layer
+    lstm_last = fluid.layers.sequence_last_step(input=lstm_h)
+    rlstm_last = fluid.layers.sequence_last_step(input=rlstm_h)
+
+    lstm_last_tanh = fluid.layers.tanh(lstm_last)
+    rlstm_last_tanh = fluid.layers.tanh(rlstm_last)
+
+    # concat layer
+    lstm_concat = fluid.layers.concat(input=[lstm_last, rlstm_last], axis=1)
+
+    # full connect layer
+    fc1 = fluid.layers.fc(input=lstm_concat, size=hid_dim2, act='tanh')
+    # softmax layer
+    prediction = fluid.layers.fc(input=fc1, size=class_dim, act='softmax')
+    cost = fluid.layers.cross_entropy(input=prediction, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+    acc = fluid.layers.accuracy(input=prediction, label=label)
+
+    return avg_cost, acc, prediction
+
+
 def gru_net(data,
             label,
             dict_dim,

fluid/text_classification/scripts/README.md

@@ -24,6 +24,7 @@ train_bow.sh
 train_cnn.sh
 train_gru.sh
 train_lstm.sh
+train_bilstm.sh
 ```
 
 ## Inference
@@ -35,4 +36,5 @@ infer_bow.sh
 infer_cnn.sh
 infer_gru.sh
 infer_lstm.sh
+infer_bilstm.sh
 ```

fluid/text_classification/scripts/infer_bilstm.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+time python ../infer.py \
+	--device CPU \
+	--model_path bilstm_model/epoch0 \
+	--num_passes 100 \
+	--profile
+

fluid/text_classification/scripts/train_bilstm.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+time python ../train.py \
+	--device CPU \
+	--model_save_dir bilstm_model \
+	--num_passes 1 \
+	bilstm
+

fluid/text_classification/train.py

@@ -10,24 +10,32 @@
 from nets import bow_net
 from nets import cnn_net
 from nets import lstm_net
+from nets import bilstm_net
 from nets import gru_net
 
-nets = {'bow': bow_net, 'cnn': cnn_net, 'lstm': lstm_net, 'gru': gru_net}
+nets = {'bow': bow_net, 'cnn': cnn_net, 'lstm': lstm_net, 'bilstm': bilstm_net,
+        'gru': gru_net}
 # learning rates
-lrs = {'bow': 0.002, 'cnn': 0.01, 'lstm': 0.05, 'gru': 0.05}
+lrs = {'bow': 0.002, 'cnn': 0.01, 'lstm': 0.05, 'bilstm':0.002, 'gru': 0.05}
 
 def parse_args():
     parser = argparse.ArgumentParser("Run inference.")
     parser.add_argument(
         'topology',
         type=str,
-        choices=['bow', 'cnn', 'lstm', 'gru'],
+        choices=['bow', 'cnn', 'lstm', 'bilstm', 'gru'],
         help='Topology used for the model (bow/cnn/lstm/gru).')
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default='imdb',
+        choices=['imdb', 'data'],
+        help="Dataset to be used: 'imdb' or 'data' (from 'data' subdirectory).")
     parser.add_argument(
         '--batch_size',
         type=int,
         default=128,
-        help='The size of a batch. (default: %(default)d, usually: 128 for "bow" and "gru", 4 for "cnn" and "lstm")')
+        help='The size of a batch. (default: %(default)d, usually: 128 for "bow" and "gru", 4 for "cnn", "lstm" and "bilstm").')
     parser.add_argument(
         '--device',
         type=str,
@@ -122,7 +130,7 @@ def train(train_reader,
 
 def train_net(args):
     word_dict, train_reader, test_reader = utils.prepare_data(
-        "imdb", self_dict=False, batch_size=128, buf_size=50000)
+        args.dataset, self_dict=False, batch_size=128, buf_size=50000)
 
     train(
         train_reader,

fluid/text_classification/utils.py

@@ -1,6 +1,8 @@
 import sys
 import time
 import numpy as np
+import random
+import os
 
 import paddle
 import paddle.fluid as fluid
@@ -48,18 +50,43 @@ def data2tensor(data, place):
     return {"words": input_seq, "label": y_data}
 
 
+def data_reader(file_path, word_dict, is_shuffle=True):
+    """
+    Convert word sequence into slot
+    """
+    unk_id = len(word_dict)
+    all_data = []
+    with open(file_path, "r") as fin:
+        for line in fin:
+            cols = line.strip().split("\t")
+            label = int(cols[0])
+            wids = [word_dict[x] if x in word_dict else unk_id
+                    for x in cols[1].split(" ")]
+            all_data.append((wids, label))
+    if is_shuffle:
+        random.shuffle(all_data)
+
+    def reader():
+        for doc, label in all_data:
+            yield doc, label
+    return reader
+
+
 def prepare_data(data_type="imdb",
                  self_dict=False,
                  batch_size=128,
                  buf_size=50000):
     """
     prepare data
     """
+    script_path = os.path.dirname(__file__)
     if self_dict:
         word_dict = load_vocab(data_type + ".vocab")
     else:
         if data_type == "imdb":
             word_dict = paddle.dataset.imdb.word_dict()
+        elif data_type == "data":
+            word_dict = load_vocab(script_path + "/data/train.vocab")
         else:
             raise RuntimeError("No such dataset")
 
@@ -68,12 +95,18 @@ def prepare_data(data_type="imdb",
             paddle.reader.shuffle(
                 paddle.dataset.imdb.train(word_dict), buf_size=buf_size),
             batch_size=batch_size)
-
         test_reader = paddle.batch(
             paddle.reader.shuffle(
                 paddle.dataset.imdb.test(word_dict), buf_size=buf_size),
             batch_size=batch_size)
+    elif data_type == "data":
+        train_reader = paddle.batch(
+            data_reader(script_path + "/data/train_data/corpus.train", word_dict, True),
+            batch_size=batch_size)
+        test_reader = paddle.batch(
+            data_reader(script_path + "/data/test_data/corpus.test", word_dict, False),
+            batch_size=batch_size)
     else:
-        raise RuntimeError("no such dataset")
+        raise RuntimeError("No such dataset")
 
     return word_dict, train_reader, test_reader