How does TensorFlow SavedModel handle additional dependencies

Question

I am trying to export my TF model using the tf.saved_model.save() function. However, I'm using additional external libraries during preprocessing. A working code using the example of nltk looks like this:

import tensorflow as tf
import nltk
import tensorflow_datasets as tfds
from tensorflow.python.ops import gen_string_ops
from tensorflow.python.ops import string_ops

train_data, val_data, test_data = train_data, validation_data, test_data = tfds.load(
    name="imdb_reviews", 
    split=('train[:60%]', 'train[60%:]', 'test'),
    as_supervised=True)


def remove_stops(text):
    nltk.download('stopwords')
    stop_words = nltk.corpus.stopwords.words('english')
    for w in stop_words:
        text = string_ops.regex_replace(text, '\\b{}\\b'.format(w),'')  # remove all stopwords
    text = tf.strings.strip(text)  # remove trailing whitespaces
    return text

vectorize_layer = tf.keras.layers.experimental.preprocessing.TextVectorization( # this gives me a BOW representation of my data...
    max_tokens=1000,
    standardize=remove_stops,  # ...using my method to remove stopwords
    output_mode='count') 

train_feats = list(map(lambda x: x[0], train_data)) # bc PrefetchDatasets don't allow direct access
vectorize_layer.adapt(train_feats)  # BOW needs to be trained beforehand to get vector reps

# define some simple network
input_layer = tf.keras.Input(shape=(), name='input_text', dtype=tf.string)
model = tf.keras.Sequential()
model.add(input_layer)
model.add(vectorize_layer)
model.add(tf.keras.layers.Dense(units=64, activation='softmax')) 
model.add(tf.keras.layers.Dense(units=1, activation='sigmoid'))
model.compile(optimizer='adam',loss='binary_crossentropy', metrics = ['binary_accuracy'])
model.fit(train_data.shuffle(10000).batch(512),
             epochs=2,
             validation_data=val_data.batch(512))

tf.saved_model.save(model, './save_here')

As far as I've understood the docs , a savedModel includes the trained parameters and computations, but NOT my code. But when I load the same model in another script like so (ie without importing my external library at all)

import tensorflow as tf
loaded_model = tf.saved_model.load('./save_here')
loaded_model(tf.constant(['test me pls']))

I get my output without any errors

<tf.Tensor: shape=(1, 1), dtype=float32, numpy=array([[0.5086063]], dtype=float32)>

This is especially baffling to me as I am even using nltk to download a dataset of stopwords. How exactly do SavedModels deal with these external libraries during exportation?

Answer 1

Actually, you are not using NLTK libraries during inferencing from the saved model.

Your model only needs text as vectors to work, that are handled by the vectorize_layer in your code. if you want to preprocess your text before inferencing you should import libraries and preprocess it.