Save tensorflow model with StringLookup layer with encoded vocabulary

Question

I'm having some issues saving a trained TensorFlow model, where I have a StringLookup layer and I'm required to use TFRecods as input for training. A minimal example to reproduce the issue:

First I define the training data

vocabulary = [str(i) for i in range(100, 200)]
X_train = np.random.choice(vocabulary, size=(100,))
y_train = np.random.choice([0,1], size=(100,))

I save it in a file as tfrecords

def _int64_feature(value):
    return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))
def _string_feature(value):
    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[str(value).encode('utf-8')]))

with tf.io.TFRecordWriter('train.tfrecords') as writer:
    for i in range(len(X_train)):
        example = tf.train.Example(features=tf.train.Features(feature={
            'user_id': _string_feature(X_train[i]),
            'label': _int64_feature(y_train[i])
        }))
        writer.write(example.SerializeToString())

Then I use the tf.data API to be able to stream the data into training (the original data doesn't fit into memory)

data = tf.data.TFRecordDataset(['train.tfrecords'])
features = {
    'user_id': tf.io.FixedLenFeature([], tf.string),
    'label': tf.io.FixedLenFeature([], tf.int64)
} 
def parse(record):
    parsed = tf.io.parse_single_example(record, features)
    return (parsed['user_id'], parsed['label'])
data = data.map(parse)

The data looks like this:

print(list(data.take(5).as_numpy_iterator()))
[(b'166', 1), (b'144', 0), (b'148', 1), (b'180', 0), (b'192', 0)]

The strings of the original dataset were converted to bytes in the process. I have to pass this new vocabulary to the StringLookup contructor, as passing strings and training with bytes will throw an error

new_vocab = [w.encode('utf-8') for w in vocabulary]

inp = tf.keras.Input(shape=(1,), dtype=tf.string)
x = tf.keras.layers.StringLookup(vocabulary=new_vocab)(inp)
x = tf.keras.layers.Embedding(len(new_vocab)+1, 32)(x)
out = tf.keras.layers.Dense(1, activation='sigmoid')(x)
model = tf.keras.Model(inputs=[inp], outputs=[out])

model.compile(optimizer='adam', loss='BinaryCrossentropy')
model.fit(data.batch(10), epochs=5)

But when I try to save the model, I get an error because the vocabulary input to the StringLookup layer is encoded as bytes and can't be dumped into json

model.save('model/')
TypeError: ('Not JSON Serializable:', b'100')

I really don't know what to do, I read that TensorFlow recommends using encoded strings instead of normal strings but that doesn't allow to save the model. I also tried to preprocess the data decoding the strings before thay are fed to the model, but I wasn't able to do it without loading all the data into memory (using just tf.data operations)

Answer 1

Using your data and original vocabulary:

import tensorflow as tf
import numpy as np

vocabulary = [str(i) for i in range(100, 200)]
X_train = np.random.choice(vocabulary, size=(100,))
y_train = np.random.choice([0,1], size=(100,))
...
...
data = data.map(parse)

I ran your code (with an additional Flatten layer) and was able to save your model:

inp = tf.keras.Input(shape=(1,), dtype=tf.string)
x = tf.keras.layers.StringLookup(vocabulary=vocabulary)(inp)
x = tf.keras.layers.Embedding(len(vocabulary)+1, 32)(x)
x = tf.keras.layers.Flatten()(x)
out = tf.keras.layers.Dense(1, activation='sigmoid')(x)
model = tf.keras.Model(inputs=[inp], outputs=[out])

model.compile(optimizer='adam', loss='BinaryCrossentropy')
model.fit(data.batch(10), epochs=5)
model.save('model/')

Epoch 1/5
10/10 [==============================] - 1s 8ms/step - loss: 0.6949
Epoch 2/5
10/10 [==============================] - 0s 4ms/step - loss: 0.6864
Epoch 3/5
10/10 [==============================] - 0s 5ms/step - loss: 0.6787
Epoch 4/5
10/10 [==============================] - 0s 5ms/step - loss: 0.6707
Epoch 5/5
10/10 [==============================] - 0s 5ms/step - loss: 0.6620
INFO:tensorflow:Assets written to: model/assets

I do not see why you need new_vocab = [w.encode('utf-8') for w in vocabulary].

If you really need to use new_vocab, you can try setting it during training and afterwards setting vocabulary for saving your model, since the only difference is the encoding:

new_vocab = [w.encode('utf-8') for w in vocabulary]

lookup_layer = tf.keras.layers.StringLookup()
lookup_layer.adapt(new_vocab)
inp = tf.keras.Input(shape=(1,), dtype=tf.string)
x = lookup_layer(inp)
x = tf.keras.layers.Embedding(len(new_vocab)+1, 32)(x)
x = tf.keras.layers.Flatten()(x)
out = tf.keras.layers.Dense(1, activation='sigmoid')(x)
model = tf.keras.Model(inputs=[inp], outputs=[out])

model.compile(optimizer='adam', loss='BinaryCrossentropy')
model.fit(data.batch(10), epochs=5)
model.layers[1].adapt(vocabulary)

model.save('/model')

Admittingly, this is quite hacky.