带有RaggedSensor输入的TensorFlow（2.4.0）模型训练问题

Question

我正在尝试训练一个TF模型，该模型将查看客户的购物车价值，并预测客户下一步将感兴趣的食品类别

只有5种食品类别：

>> all_food_categories
['pizza', 'side', 'dessert', 'drink', 'dip']

购物车中已经存在的食品类别是字符串类型的RaggedTensor

for i in generator(): is # generator returns a tuple, (features, target)
print(i)
break

(

{'categories_of_assumed_cart_flattened': <tf.RaggedTensor [[b'<blank>'], [b'pizza', b'dessert'], 
[b'pizza', b'side', b'side', b'dip', b'dip', b'dessert', b'dessert'],
[b'pizza', b'pizza'], [b'side'], [b'pizza', b'pizza', b'pizza', b'side', b'pizza'], [b'pizza'], [b'pizza', b'pizza', b'pizza', b'side'], [b'side', b'drink', b'side', b'dessert'], [b'side'], [b'pizza', b'side', b'side', b'dessert', b'dip', b'dip'], [b'pizza', b'pizza', b'pizza', b'pizza'], [b'pizza', b'drink', b'dip', b'dip'], [b'pizza', b'side', b'side'], [b'pizza', b'pizza', b'side'], [b'pizza', b'side'], [b'side'], [b'side', b'drink'], [b'pizza', b'side'], [b'pizza', b'side', b'side', b'dessert', b'drink'], [b'drink'], [b'drink'], [b'pizza', b'pizza', b'pizza'], [b'side', b'dip'], [b'side', b'side', b'side'], [b'<blank>'], [b'pizza', b'dip', b'dip', b'side'], [b'pizza', b'side', b'dessert'], [b'pizza'], [b'<blank>'], [b'pizza', b'pizza', b'side', b'side', b'dip', b'dip'], [b'pizza', b'pizza']]>, 'disposition_type': <tf.Tensor: shape=(32,), dtype=string, numpy=
    array([b'collection', b'delivery', b'collection', b'collection',
           b'delivery', b'delivery', b'collection', b'delivery', b'delivery',
           b'delivery', b'delivery', b'delivery', b'delivery', b'delivery',
           b'delivery', b'delivery', b'collection', b'delivery', b'delivery',
           b'delivery', b'delivery', b'collection', b'collection',
           b'delivery', b'delivery', b'delivery', b'delivery', b'delivery',
           b'delivery', b'delivery', b'delivery', b'collection'], dtype=object)>}, array([0, 0, 4, 1, 3, 3, 4, 1, 1, 1, 0, 4, 1, 2, 2, 1, 0, 0, 2, 0, 0, 0,
           4, 0, 0, 0, 1, 1, 1, 0, 1, 1]))

型号代码：

class NextItemCategory(tf.keras.Model):
    def __init__(self, vocab, mask_token = '', embed_dim=4, conv_kernels=[3,4,5], max_seq_len = 7):
        super(NextItemCategory, self).__init__()
        self.mask_token = mask_token
        self.max_seq_len = max_seq_len
        self.lookup = tf.keras.layers.experimental.preprocessing.StringLookup(vocabulary=vocab, mask_token=mask_token)
        self.embed = tf.keras.layers.Embedding(len(self.lookup.get_vocabulary()), 4)
        self.model_layers = [tf.keras.layers.Conv2D(filters=1, kernel_size=[ck_i, embed_dim], padding='same') for ck_i in conv_kernels]
        self.pool = tf.keras.layers.GlobalMaxPool2D()
        self.dense = tf.keras.layers.Dense(5, activation='softmax')


    def call(self, inputs):

        inp = inputs["categories_of_assumed_cart_flattened"]
        x = inp.to_tensor(default_value='', shape = (None, self.max_seq_len))
        x = self.lookup(x)
        x = self.embed(x)
        x = tf.expand_dims(input=x, axis=-1)
        z = []
        for layer in self.model_layers:
            y = layer(x)
            y = self.pool(y)
            z.append(y)
        z = tf.concat(z, axis=-1)
        z = self.dense(z)
        return z

cart_model = NextItemCategory(all_food_categories)
cart_model.compile(loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True), optimizer='adam', metrics=['accuracy'])

train_data_generator = generator()
cart_model.fit(train_data_generator, verbose=1)

模型正在构建中

Model: "next_item_category"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
string_lookup_20 (StringLook multiple                  0         
_________________________________________________________________
embedding_20 (Embedding)     multiple                  28        
_________________________________________________________________
conv2d_60 (Conv2D)           multiple                  13        
_________________________________________________________________
conv2d_61 (Conv2D)           multiple                  17        
_________________________________________________________________
conv2d_62 (Conv2D)           multiple                  21        
_________________________________________________________________
global_max_pooling2d_20 (Glo multiple                  0         
_________________________________________________________________
dense_20 (Dense)             multiple                  20        
=================================================================
Total params: 99
Trainable params: 99
Non-trainable params: 0

---

但作为fit()的一部分，我得到了这个错误：

AttributeError                            Traceback (most recent call last)
------
<ipython-input-110-5d3dc4a7c2fa> in call(self, inputs)
     12 
     13         inp = inputs["categories_of_assumed_cart_flattened"]
---> 14         x = inp.to_tensor(default_value='', shape = (None, self.max_seq_len))
     15         x = self.lookup(x)
     16         x = self.embed(x)

AttributeError: 'Tensor' object has no attribute 'to_tensor'

在卷积开始之前，我已经尝试将to_tensor()调用放在不同的代码位置，但是相同的错误仍然存在

从生成器输出可以清楚地看出inputs['categories_of_assumed_cart_flattened']始终是tf.RaggedTensor类型

我不知道问题出在哪里；非常感谢您的帮助！非常感谢