Tensorflow處理變長特征

處理流程：

1. 變長特征分割成變長數組
2. 變長數據填充成規則數組，組成n * m的矩陣 (keras.preprocessing.sequence.pad_sequences)
3. 每一行數據進行embedding，結果可以按權重求平均、直接求平均、求最大值 得到 n*1結果矩陣

第3步求平均可以用tf.nn.lookup_embedding_sparse 來做，也可以在Embedding之后再加一層MaxPooling2D或者AVGPooling2D。

參考：

• https://www.machinecurve.com/index.php/2020/03/03/classifying-imdb-sentiment-with-keras-and-embeddings-dropout-conv1d/
  構建池化層時類似：

# Define the Keras model
model = Sequential()
model.add(Embedding(num_distinct_words, embedding_output_dims, input_length=max_sequence_length))
model.add(Dropout(0.50))
model.add(Conv1D(filters=32, kernel_size=2, padding='same', activation='relu'))
model.add(Dropout(0.50))
model.add(MaxPooling1D(pool_size=2))
model.add(Flatten())
model.add(Dropout(0.50))
model.add(Dense(1, activation='sigmoid'))

• https://zhuanlan.zhihu.com/p/141443615
  知乎的這個使用了lookup_embedding_sparse，自帶根據權重求平均。

• https://zhuanlan.zhihu.com/p/94212544
  lookup_sparse 計算方法

處理變長特征代碼實現參考

# -*- encoding: utf-8 -*-

import pandas as pd
from sklearn import preprocessing
from tensorflow import keras
from tensorflow.keras.models import Model, save_model
from tensorflow.python.keras.preprocessing.sequence import pad_sequences
from tensorflow.keras.layers import Input
import tensorflow as tf


class VarLenColumnEmbedding(tf.keras.layers.Embedding):
    """
    變長列做Embedding
    """
    def __init__(self, pooling_strategy='max', dropout_rate=0.,  **kwargs):
        if pooling_strategy not in ['mean', 'max']:
            raise ValueError("Param strategy should is one of mean, max")
        self.pooling_strategy = pooling_strategy
        self.dropout_rate = dropout_rate  # 支持dropout
        super(VarLenColumnEmbedding, self).__init__(**kwargs)

    def build(self, input_shape):
        super(VarLenColumnEmbedding, self).build(input_shape)  # Be sure to call this somewhere!
        height = input_shape[1]
        if self.pooling_strategy == "mean":
            self._pooling_layer = tf.keras.layers.AveragePooling2D(pool_size=(height, 1))
        else:
            self._pooling_layer = tf.keras.layers.MaxPooling2D(pool_size=(height, 1))

        if self.dropout_rate > 0:
            self._dropout = tf.keras.layers.SpatialDropout1D(self.dropout_rate)
        else:
            self._dropout = None

        self.built = True

    def call(self, inputs):
        # 1. do embedding
        embedding_output = super(VarLenColumnEmbedding, self).call(inputs)

        # 2. add dropout
        if self._dropout is not None:
            dropout_output = self._dropout(embedding_output)
        else:
            dropout_output = embedding_output

        # 3. expand dim for polling
        inputs_4d = tf.expand_dims(dropout_output, 3)  # add channels dim

        # 4. polling
        tensor_pooling = self._pooling_layer(inputs_4d)

        # 5. format output
        return tf.squeeze(tensor_pooling, 3)

    def compute_mask(self, inputs, mask):
        return None

    def get_config(self, ):
        config = {'pooling_strategy': self.pooling_strategy}
        base_config = super(VarLenColumnEmbedding, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))


data = pd.read_csv("movielens_sample.txt")

series_genres = data.pop('genres').map(lambda _: _.split("|"))
gs_list = []
gs_max_len = 0
for gs in series_genres:
    if len(gs) > gs_max_len:
        gs_max_len = len(gs)
    gs_list.extend(gs)

gs_list = list(set(gs_list))
gs_n_unique = len(gs_list)

lb = preprocessing.LabelEncoder()
lb.fit(gs_list)

y = data['rating'].values

data_geners = pad_sequences(series_genres.map(lambda _: lb.transform(_).tolist()), maxlen=gs_max_len, padding='post', )
feature_names = ['user_id', 'movie_id', 'timestamp', 'title', 'gender', 'age', 'occupation', 'zip']

for col in feature_names:
    data[col] = preprocessing.LabelEncoder().fit_transform(data[col])

movie_id_uniques = len(set(data['movie_id'].tolist()))

input_movie_genres = Input(shape=(gs_max_len,), name='movie_genres_input')
layer_embedding_genres = VarLenColumnEmbedding(pooling_strategy='max', input_dim=gs_n_unique, output_dim=4)(input_movie_genres)

dense1 = keras.layers.Dense(512, activation='relu')(layer_embedding_genres)
output = keras.layers.Dense(6, activation='softmax')(dense1)

model = Model(inputs=[input_movie_genres], outputs=output)
model.run_eagerly = True

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])
model.summary()

model.fit(data_geners, y, epochs=10)

網絡結構：

使用到的數據：
movielens_sample.zip