? 作者：韩信子@ShowMeAI
? 深度学习实战系列：https://www.showmeai.tech/tutorials/42
? TensorFlow 实战系列：https://www.showmeai.tech/tutorials/43
? 本文地址：https://www.showmeai.tech/article-detail/315
? 声明：版权所有，转载请联系平台与作者并注明出处
? 收藏ShowMeAI查看更多精彩内容
自 Transformers 出现以来，基于它的结构已经颠覆了自然语言处理和计算机视觉，带来各种非结构化数据业务场景和任务的巨大效果突破，接着大家把目光转向了结构化业务数据，它是否能在结构化表格数据上同样有惊人的效果表现呢？
答案是YES！亚马逊在论文中提出的 ?TabTransformer，是一种把结构调整后适应于结构化表格数据的网络结构，它更擅长于捕捉传统结构化表格数据中不同类型的数据信息，并将其结合以完成预估任务。下面ShowMeAI给大家讲解构建 TabTransformer 并将其应用于结构化数据上的过程。
? 环境设置
本篇使用到的深度学习框架为TensorFlow，大家需要安装2.7或更高版本， 我们还需要安装一下 ?TensorFlow插件addons，安装的过程大家可以通过下述命令完成：
pip install -U tensorflow tensorflow-addons
关于本篇代码实现中使用到的TensorFlow工具库，大家可以查看ShowMeAI制作的TensorFlow速查手册快学快用：
- AI垂直领域工具库速查表 | TensorFlow2建模速查&应用速查
接下来我们导入工具库
import mathimport numpy as npimport pandas as pdimport tensorflow as tffrom tensorflow import kerasfrom tensorflow.keras import layersimport tensorflow_addons as tfaimport matplotlib.pyplot as plt? 数据说明
ShowMeAI在本例中使用到的是 ?美国人口普查收入数据集，任务是根据人口基本信息预测其年收入是否可能超过 50,000 美元，是一个二分类问题。
数据集可以在以下地址下载：
? https://archive.ics.uci.edu/ml/datasets/Adult
? https://archive.ics.uci.edu/ml/machine-learning-databases/adult/
数据从美国1994年人口普查数据库抽取而来，可以用来预测居民收入是否超过50K/year。该数据集类变量为年收入是否超过50k，属性变量包含年龄、工种、学历、职业、人种等重要信息，值得一提的是，14个属性变量中有7个类别型变量。数据集各属性是：其中序号0~13是属性，14是类别。
| 字段序号 | 字段名 | 含义 | 类型 |
|---|---|---|---|
| 0 | age | 年龄 | Double |
| 1 | workclass | 工作类型* | string |
| 2 | fnlwgt | 序号 | string |
| 3 | education | 教育程度* | string |
| 4 | education_num | 受教育时间 | double |
| 5 | maritial_status | 婚姻状况* | string |
| 6 | occupation | 职业* | string |
| 7 | relationship | 关系* | string |
| 8 | race | 种族* | string |
| 9 | sex | 性别* | string |
| 10 | capital_gain | 资本收益 | string |
| 11 | capital_loss | 资本损失 | string |
| 12 | hours_per_week | 每周工作小时数 | double |
| 13 | native_country | 原籍* | string |
| 14(label) | income | 收入标签 | string |
我们先用pandas读取数据到dataframe中：
CSV_HEADER = [ "age", "workclass", "fnlwgt", "education", "education_num", "marital_status", "occupation", "relationship", "race", "gender", "capital_gain", "capital_loss", "hours_per_week", "native_country", "income_bracket",]train_data_url = ( "https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.data")train_data = pd.read_csv(train_data_url, header=None, names=CSV_HEADER)test_data_url = ( "https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.test")test_data = pd.read_csv(test_data_url, header=None, names=CSV_HEADER)print(f"Train dataset shape: {train_data.shape}")print(f"Test dataset shape: {test_data.shape}")Train dataset shape: (32561, 15)Test dataset shape: (16282, 15)我们做点数据清洗，把测试集第一条记录剔除（它不是有效的数据示例），把类标签中的尾随的“点”去掉。
test_data = test_data[1:]test_data.income_bracket = test_data.income_bracket.apply( lambda value: value.replace(".", ""))再把训练集和测试集存回单独的 CSV 文件中。
train_data_file = "train_data.csv"test_data_file = "test_data.csv"train_data.to_csv(train_data_file, index=False, header=False)test_data.to_csv(test_data_file, index=False, header=False)? 模型原理
TabTransformer的模型架构如下所示：
我们可以看到，类别型的特征，很适合在 embedding 后，送入 transformer 模块进行深度交叉组合与信息挖掘，得到的信息与右侧的连续值特征进行拼接，再送入全连接的 MLP 模块进行组合和完成最后的任务（分类或者回归）。
? 模型实现? 定义数据集元数据
要实现模型，我们先对输入数据字段，区分不同的类型（数值型特征与类别型特征）。我们会对不同类型的特征，使用不同的方式进行处理和完成特征工程（例如数值型的特征进行幅度缩放，类别型的特征进行编码处理）。
## 数值特征字段NUMERIC_FEATURE_NAMES = [ "age", "education_num", "capital_gain", "capital_loss", "hours_per_week",]## 类别型特征字段及其取值列表CATEGORICAL_FEATURES_WITH_VOCABULARY = { "workclass": sorted(list(train_data["workclass"].unique())), "education": sorted(list(train_data["education"].unique())), "marital_status": sorted(list(train_data["marital_status"].unique())), "occupation": sorted(list(train_data["occupation"].unique())), "relationship": sorted(list(train_data["relationship"].unique())), "race": sorted(list(train_data["race"].unique())), "gender": sorted(list(train_data["gender"].unique())), "native_country": sorted(list(train_data["native_country"].unique())),}## 权重字段WEIGHT_COLUMN_NAME = "fnlwgt"## 类别型字段名称CATEGORICAL_FEATURE_NAMES = list(CATEGORICAL_FEATURES_WITH_VOCABULARY.keys())## 所有的输入特征FEATURE_NAMES = NUMERIC_FEATURE_NAMES + CATEGORICAL_FEATURE_NAMES## 默认填充的取值COLUMN_DEFAULTS = [ [0.0] if feature_name in NUMERIC_FEATURE_NAMES + [WEIGHT_COLUMN_NAME] else ["NA"] for feature_name in CSV_HEADER]## 目标字段TARGET_FEATURE_NAME = "income_bracket"## 目标字段取值TARGET_LABELS = [" 50K"]? 配置超参数
我们为神经网络的结构和训练过程的超参数进行设置，如下。
# 学习率LEARNING_RATE = 0.001# 学习率衰减WEIGHT_DECAY = 0.0001# 随机失活 概率参数DROPOUT_RATE = 0.2# 批数据大小BATCH_SIZE = 265# 总训练轮次数NUM_EPOCHS = 15# transformer块的数量NUM_TRANSFORMER_BLOCKS = 3# 注意力头的数量NUM_HEADS = 4# 类别型embedding嵌入的维度EMBEDDING_DIMS = 16# MLP隐层单元数量MLP_HIDDEN_UNITS_FACTORS = [ 2, 1,]# MLP块的数量NUM_MLP_BLOCKS = 2? 实现数据读取管道
下面我们定义一个输入函数，它负责读取和解析文件，并对特征和标签处理，放入 tf.data.Dataset，以便后续训练和评估。
target_label_lookup = layers.StringLookup( vocabulary=TARGET_LABELS, mask_token=None, num_oov_indices=0)def prepare_example(features, target): target_index = target_label_lookup(target) weights = features.pop(WEIGHT_COLUMN_NAME) return features, target_index, weights# 从csv中读取数据def get_dataset_from_csv(csv_file_path, batch_size=128, shuffle=False): dataset = tf.data.experimental.make_csv_dataset( csv_file_path, batch_size=batch_size, column_names=CSV_HEADER, column_defaults=COLUMN_DEFAULTS, label_name=TARGET_FEATURE_NAME, num_epochs=1, header=False, na_value="?", shuffle=shuffle, ).map(prepare_example, num_parallel_calls=tf.data.AUTOTUNE, deterministic=False) return dataset.cache()? 模型构建与评估
def run_experiment( model, train_data_file, test_data_file, num_epochs, learning_rate, weight_decay, batch_size,): # 优化器 optimizer = tfa.optimizers.AdamW( learning_rate=learning_rate, weight_decay=weight_decay ) # 模型编译 model.compile( optimizer=optimizer, loss=keras.losses.BinaryCrossentropy(), metrics=[keras.metrics.BinaryAccuracy(name="accuracy")], ) # 训练集与验证集 train_dataset = get_dataset_from_csv(train_data_file, batch_size, shuffle=True) validation_dataset = get_dataset_from_csv(test_data_file, batch_size) # 模型训练 print("Start training the model...") history = model.fit( train_dataset, epochs=num_epochs, validation_data=validation_dataset ) print("Model training finished") # 模型评估 _, accuracy = model.evaluate(validation_dataset, verbose=0) print(f"Validation accuracy: {round(accuracy * 100, 2)}%") return history① 创建模型输入
基于 TensorFlow 的输入要求，我们将模型的输入定义为字典，其中『key/键』是特征名称，『value/值』为 keras.layers.Input具有相应特征形状的张量和数据类型。
def create_model_inputs(): inputs = {} for feature_name in FEATURE_NAMES: if feature_name in NUMERIC_FEATURE_NAMES: inputs[feature_name] = layers.Input( name=feature_name, shape=(), dtype=tf.float32 ) else: inputs[feature_name] = layers.Input( name=feature_name, shape=(), dtype=tf.string ) return inputs② 编码特征
我们定义一个encode_inputs函数，返回encoded_categorical_feature_list和 numerical_feature_list。我们将分类特征编码为嵌入，使用固定的embedding_dims对于所有功能， 无论他们的词汇量大小。 这是 Transformer 模型所必需的。
def encode_inputs(inputs, embedding_dims): encoded_categorical_feature_list = [] numerical_feature_list = [] for feature_name in inputs: if feature_name in CATEGORICAL_FEATURE_NAMES: # 获取类别型特征的不同取值(vocabulary) vocabulary = CATEGORICAL_FEATURES_WITH_VOCABULARY[feature_name] # 构建lookup table去构建 类别型取值 和 索引 的相互映射 lookup = layers.StringLookup( vocabulary=vocabulary, mask_token=None, num_oov_indices=0, output_mode="int", ) # 类别型字符串取值 转为 整型索引 encoded_feature = lookup(inputs[feature_name]) # 构建embedding层 embedding = layers.Embedding( input_dim=len(vocabulary), output_dim=embedding_dims ) # 为索引构建embedding嵌入 encoded_categorical_feature = embedding(encoded_feature) encoded_categorical_feature_list.append(encoded_categorical_feature) else: # 数值型特征 numerical_feature = tf.expand_dims(inputs[feature_name], -1) numerical_feature_list.append(numerical_feature) return encoded_categorical_feature_list, numerical_feature_list③ MLP模块实现
网络中不可或缺的部分是 MLP 全连接板块，下面是它的简单实现：
def create_mlp(hidden_units, dropout_rate, activation, normalization_layer, name=None): mlp_layers = [] for units in hidden_units: mlp_layers.append(normalization_layer), mlp_layers.append(layers.Dense(units, activation=activation)) mlp_layers.append(layers.Dropout(dropout_rate)) return keras.Sequential(mlp_layers, name=name)④ 模型实现1：基线模型
为了对比效果，我们先简单使用MLP（多层前馈网络）进行建模，代码和注释如下。
def create_baseline_model( embedding_dims, num_mlp_blocks, mlp_hidden_units_factors, dropout_rate): # 创建输入. inputs = create_model_inputs() # 特征编码 encoded_categorical_feature_list, numerical_feature_list = encode_inputs( inputs, embedding_dims ) # 拼接所有特征 features = layers.concatenate( encoded_categorical_feature_list + numerical_feature_list ) # 前向计算 feedforward_units = [features.shape[-1]] # 构建全连接，并且添加跳跃连接(skip-connection) for layer_idx in range(num_mlp_blocks): features = create_mlp( hidden_units=feedforward_units, dropout_rate=dropout_rate, activation=keras.activations.gelu, normalization_layer=layers.LayerNormalization(epsilon=1e-6), name=f"feedforward_{layer_idx}", )(features) # MLP全连接的隐层结果 mlp_hidden_units = [ factor * features.shape[-1] for factor in mlp_hidden_units_factors ] # 最终的MLP网络 features = create_mlp( hidden_units=mlp_hidden_units, dropout_rate=dropout_rate, activation=keras.activations.selu, normalization_layer=layers.BatchNormalization(), name="MLP", )(features) # 添加sigmoid构建二分类器 outputs = layers.Dense(units=1, activation="sigmoid", name="sigmoid")(features) model = keras.Model(inputs=inputs, outputs=outputs) return model# 完整的模型baseline_model = create_baseline_model( embedding_dims=EMBEDDING_DIMS, num_mlp_blocks=NUM_MLP_BLOCKS, mlp_hidden_units_factors=MLP_HIDDEN_UNITS_FACTORS, dropout_rate=DROPOUT_RATE,)print("Total model weights:", baseline_model.count_params())keras.utils.plot_model(baseline_model, show_shapes=True, rankdir="LR")# Total model weights: 109629上述模型构建完成之后，我们通过plot_model操作，绘制出模型结构如下：
 |
接下来我们训练和评估一下基线模型：
history = run_experiment( model=baseline_model, train_data_file=train_data_file, test_data_file=test_data_file, num_epochs=NUM_EPOCHS, learning_rate=LEARNING_RATE, weight_decay=WEIGHT_DECAY, batch_size=BATCH_SIZE,)输出的训练过程日志如下：
Start training the model...Epoch 1/15123/123 [==============================] - 6s 25ms/step - loss: 110178.8203 - accuracy: 0.7478 - val_loss: 92703.0859 - val_accuracy: 0.7825Epoch 2/15123/123 [==============================] - 2s 14ms/step - loss: 90979.8125 - accuracy: 0.7675 - val_loss: 71798.9219 - val_accuracy: 0.8001Epoch 3/15123/123 [==============================] - 2s 14ms/step - loss: 77226.5547 - accuracy: 0.7902 - val_loss: 68581.0312 - val_accuracy: 0.8168Epoch 4/15123/123 [==============================] - 2s 14ms/step - loss: 72652.2422 - accuracy: 0.8004 - val_loss: 70084.0469 - val_accuracy: 0.7974Epoch 5/15123/123 [==============================] - 2s 14ms/step - loss: 71207.9375 - accuracy: 0.8033 - val_loss: 66552.1719 - val_accuracy: 0.8130Epoch 6/15123/123 [==============================] - 2s 14ms/step - loss: 69321.4375 - accuracy: 0.8091 - val_loss: 65837.0469 - val_accuracy: 0.8149Epoch 7/15123/123 [==============================] - 2s 14ms/step - loss: 68839.3359 - accuracy: 0.8099 - val_loss: 65613.0156 - val_accuracy: 0.8187Epoch 8/15123/123 [==============================] - 2s 14ms/step - loss: 68126.7344 - accuracy: 0.8124 - val_loss: 66155.8594 - val_accuracy: 0.8108Epoch 9/15123/123 [==============================] - 2s 14ms/step - loss: 67768.9844 - accuracy: 0.8147 - val_loss: 66705.8047 - val_accuracy: 0.8230Epoch 10/15123/123 [==============================] - 2s 14ms/step - loss: 67482.5859 - accuracy: 0.8151 - val_loss: 65668.3672 - val_accuracy: 0.8143Epoch 11/15123/123 [==============================] - 2s 14ms/step - loss: 66792.6875 - accuracy: 0.8181 - val_loss: 66536.3828 - val_accuracy: 0.8233Epoch 12/15123/123 [==============================] - 2s 14ms/step - loss: 65610.4531 - accuracy: 0.8229 - val_loss: 70377.7266 - val_accuracy: 0.8256Epoch 13/15123/123 [==============================] - 2s 14ms/step - loss: 63930.2500 - accuracy: 0.8282 - val_loss: 68294.8516 - val_accuracy: 0.8289Epoch 14/15123/123 [==============================] - 2s 14ms/step - loss: 63420.1562 - accuracy: 0.8323 - val_loss: 63050.5859 - val_accuracy: 0.8204Epoch 15/15123/123 [==============================] - 2s 14ms/step - loss: 62619.4531 - accuracy: 0.8345 - val_loss: 66933.7500 - val_accuracy: 0.8177Model training finishedValidation accuracy: 81.77%我们可以看到基线模型(全连接MLP网络)实现了约 82% 的验证准确度。
⑤ 模型实现2：TabTransformer
TabTransformer 架构的工作原理如下：
- 所有类别型特征都被编码为嵌入，使用相同的
embedding_dims。 - 将列嵌入（每个类别型特征的一个嵌入向量）添加类别型特征嵌入中。
- 嵌入的类别型特征被输入到一系列的 Transformer 块中。 每个 Transformer 块由一个多头自注意力层和一个前馈层组成。
- 最终 Transformer 层的输出， 与输入的数值型特征连接，并输入到最终的 MLP 块中。
- 尾部由一个
softmax结构完成分类。
def create_tabtransformer_classifier( num_transformer_blocks, num_heads, embedding_dims, mlp_hidden_units_factors, dropout_rate, use_column_embedding=False,): # 构建输入 inputs = create_model_inputs() # 编码特征 encoded_categorical_feature_list, numerical_feature_list = encode_inputs( inputs, embedding_dims ) # 堆叠类别型特征的embeddings，为输入Tansformer做准备 encoded_categorical_features = tf.stack(encoded_categorical_feature_list, axis=1) # 拼接数值型特征 numerical_features = layers.concatenate(numerical_feature_list) # embedding if use_column_embedding: num_columns = encoded_categorical_features.shape[1] column_embedding = layers.Embedding( input_dim=num_columns, output_dim=embedding_dims ) column_indices = tf.range(start=0, limit=num_columns, delta=1) encoded_categorical_features = encoded_categorical_features + column_embedding( column_indices ) # 构建Transformer块 for block_idx in range(num_transformer_blocks): # 多头自注意力 attention_output = layers.MultiHeadAttention( num_heads=num_heads, key_dim=embedding_dims, dropout=dropout_rate, name=f"multihead_attention_{block_idx}", )(encoded_categorical_features, encoded_categorical_features) # 第1个跳接/Skip connection x = layers.Add(name=f"skip_connection1_{block_idx}")( [attention_output, encoded_categorical_features] ) # 第1个层归一化/Layer normalization x = layers.LayerNormalization(name=f"layer_norm1_{block_idx}", epsilon=1e-6)(x) # 全连接层 feedforward_output = create_mlp( hidden_units=[embedding_dims], dropout_rate=dropout_rate, activation=keras.activations.gelu, normalization_layer=layers.LayerNormalization(epsilon=1e-6), name=f"feedforward_{block_idx}", )(x) # 第2个跳接/Skip connection x = layers.Add(name=f"skip_connection2_{block_idx}")([feedforward_output, x]) # 第2个层归一化/Layer normalization encoded_categorical_features = layers.LayerNormalization( name=f"layer_norm2_{block_idx}", epsilon=1e-6 )(x) # 展平embeddings categorical_features = layers.Flatten()(encoded_categorical_features) # 对数值型特征做层归一化 numerical_features = layers.LayerNormalization(epsilon=1e-6)(numerical_features) # 拼接作为最终MLP的输入 features = layers.concatenate([categorical_features, numerical_features]) # 计算MLP隐层单元 mlp_hidden_units = [ factor * features.shape[-1] for factor in mlp_hidden_units_factors ] # 构建最终的MLP. features = create_mlp( hidden_units=mlp_hidden_units, dropout_rate=dropout_rate, activation=keras.activations.selu, normalization_layer=layers.BatchNormalization(), name="MLP", )(features) # 添加sigmoid构建二分类 outputs = layers.Dense(units=1, activation="sigmoid", name="sigmoid")(features) model = keras.Model(inputs=inputs, outputs=outputs) return modeltabtransformer_model = create_tabtransformer_classifier( num_transformer_blocks=NUM_TRANSFORMER_BLOCKS, num_heads=NUM_HEADS, embedding_dims=EMBEDDING_DIMS, mlp_hidden_units_factors=MLP_HIDDEN_UNITS_FACTORS, dropout_rate=DROPOUT_RATE,)print("Total model weights:", tabtransformer_model.count_params())keras.utils.plot_model(tabtransformer_model, show_shapes=True, rankdir="LR")#Total model weights: 87479最终输出的模型结构示意图如下（因为模型结构较深，总体很长，点击放大）
 |
下面我们训练和评估一下TabTransformer 模型的效果：
history = run_experiment( model=tabtransformer_model, train_data_file=train_data_file, test_data_file=test_data_file, num_epochs=NUM_EPOCHS, learning_rate=LEARNING_RATE, weight_decay=WEIGHT_DECAY, batch_size=BATCH_SIZE,)Start training the model...Epoch 1/15123/123 [==============================] - 13s 61ms/step - loss: 82503.1641 - accuracy: 0.7944 - val_loss: 64260.2305 - val_accuracy: 0.8421Epoch 2/15123/123 [==============================] - 6s 51ms/step - loss: 68677.9375 - accuracy: 0.8251 - val_loss: 63819.8633 - val_accuracy: 0.8389Epoch 3/15123/123 [==============================] - 6s 51ms/step - loss: 66703.8984 - accuracy: 0.8301 - val_loss: 63052.8789 - val_accuracy: 0.8428Epoch 4/15123/123 [==============================] - 6s 51ms/step - loss: 65287.8672 - accuracy: 0.8342 - val_loss: 61593.1484 - val_accuracy: 0.8451Epoch 5/15123/123 [==============================] - 6s 52ms/step - loss: 63968.8594 - accuracy: 0.8379 - val_loss: 61385.4531 - val_accuracy: 0.8442Epoch 6/15123/123 [==============================] - 6s 51ms/step - loss: 63645.7812 - accuracy: 0.8394 - val_loss: 61332.3281 - val_accuracy: 0.8447Epoch 7/15123/123 [==============================] - 6s 51ms/step - loss: 62778.6055 - accuracy: 0.8412 - val_loss: 61342.5352 - val_accuracy: 0.8461Epoch 8/15123/123 [==============================] - 6s 51ms/step - loss: 62815.6992 - accuracy: 0.8398 - val_loss: 61220.8242 - val_accuracy: 0.8460Epoch 9/15123/123 [==============================] - 6s 52ms/step - loss: 62191.1016 - accuracy: 0.8416 - val_loss: 61055.9102 - val_accuracy: 0.8452Epoch 10/15123/123 [==============================] - 6s 51ms/step - loss: 61992.1602 - accuracy: 0.8439 - val_loss: 61251.8047 - val_accuracy: 0.8441Epoch 11/15123/123 [==============================] - 6s 50ms/step - loss: 61745.1289 - accuracy: 0.8429 - val_loss: 61364.7695 - val_accuracy: 0.8445Epoch 12/15123/123 [==============================] - 6s 51ms/step - loss: 61696.3477 - accuracy: 0.8445 - val_loss: 61074.3594 - val_accuracy: 0.8450Epoch 13/15123/123 [==============================] - 6s 51ms/step - loss: 61569.1719 - accuracy: 0.8436 - val_loss: 61844.9688 - val_accuracy: 0.8456Epoch 14/15123/123 [==============================] - 6s 51ms/step - loss: 61343.0898 - accuracy: 0.8445 - val_loss: 61702.8828 - val_accuracy: 0.8455Epoch 15/15123/123 [==============================] - 6s 51ms/step - loss: 61355.0547 - accuracy: 0.8504 - val_loss: 61272.2852 - val_accuracy: 0.8495Model training finishedValidation accuracy: 84.55%TabTransformer 模型实现了约 85% 的验证准确度，相比于直接使用全连接网络效果有一定的提升。
参考资料
- ? TabTransformer：https://arxiv.org/abs/2012.06678
- ? TensorFlow插件addons：https://www.tensorflow.org/addons/overview
- ?AI垂直领域工具库速查表 | TensorFlow2建模速查&应用速查：https://www.showmeai.tech/article-detail/109
