add textclassifiler (#4547)

PaddleRec/text_classifiler/README.md

+# Text_Classifiler
+
+以下是本例的简要目录结构及说明：
+
+```text
+.
+├── README.md            # 文档
+├── train.py             # 训练脚本
+└── net.py               # 网络结构
+
+```
+
+
+## 简介
+
+文本分类是机器学习中经典的NLP和推荐任务，包括但不限于情感分类，标签推荐等等。以CNN为例介绍文本分类的任务。
+
+## 数据
+
+模型的输入数据为若干ID（表示词或者其他id）和一个label表示分类的标签。
+
+
+原始数据格式分为两列，第一列是输入数据，第二列是分类标签
+
+```
+特 喜欢 这种 好看的 狗狗                 1
+这 真是 惊艳 世界 的 中国 黑科技          1
+环境 特别 差 ，脏兮兮 的，再也 不去 了     0
+```
+本例采用随机数据作为demo样例
+
+
+## 训练命令
+```
+python train.py
+```
+
+## 未来工作
+
+添加公开数据集和测试结果

PaddleRec/text_classifiler/net.py

+import paddle.fluid as fluid
+
+
+def cnn_net(dict_dim=100,
+            max_len=10,
+            cnn_dim=32,
+            cnn_filter_size=128,
+            emb_dim=8,
+            hid_dim=128,
+            class_dim=2,
+            is_prediction=False):
+    """
+    Conv net
+    """
+    data = fluid.data(name="input", shape=[None, max_len], dtype='int64')
+    label = fluid.data(name="label", shape=[None, 1], dtype='int64')
+    seq_len = fluid.data(name="seq_len", shape=[None], dtype='int64')
+    # embedding layer
+    emb = fluid.embedding(input=data, size=[dict_dim, emb_dim])
+    emb = fluid.layers.sequence_unpad(emb, length=seq_len)
+    # convolution layer
+    conv = fluid.nets.sequence_conv_pool(
+        input=emb,
+        num_filters=cnn_dim,
+        filter_size=cnn_filter_size,
+        act="tanh",
+        pool_type="max")
+
+    # full connect layer
+    fc_1 = fluid.layers.fc(input=[conv], size=hid_dim)
+    # softmax layer
+    prediction = fluid.layers.fc(input=[fc_1], size=class_dim, act="softmax")
+    #if is_prediction:
+    #    return prediction
+    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

PaddleRec/text_classifiler/train.py

+import net
+import numpy as np
+import paddle.fluid as fluid
+
+
+def gen_data(dict_dim=100, class_size=2, batch_size=32, max_len=10):
+    return {
+        "input": np.random.randint(
+            dict_dim, size=(batch_size, max_len)).astype('int64'),
+        "seq_len": np.random.randint(
+            1, high=max_len, size=(batch_size)).astype('int64'),
+        "label": np.random.randint(
+            class_size, size=(batch_size, 1)).astype('int64')
+    }
+
+
+main_program = fluid.default_startup_program()
+startup_program = fluid.default_main_program()
+dict_dim = 100
+with fluid.program_guard(main_program, startup_program):
+    cost = net.cnn_net(dict_dim=dict_dim)
+    optimizer = fluid.optimizer.SGD(learning_rate=0.01)
+    optimizer.minimize(cost)
+    exe = fluid.Executor(fluid.CPUPlace())
+    exe.run(startup_program)
+    step = 100
+    for i in range(step):
+        cost_val = exe.run(main_program,
+                           feed=gen_data(),
+                           fetch_list=[cost.name])
+        print("step%d cost=%f" % (i, cost_val[0]))