tensorflow: vocabulary lookup with tf.estimator and keras

Question

I have the following dataset

username,itemname,value
"carl","socks",12.50
"john doe","shirts",30.00
...

I also have the following vocabulary lookup files

usernames.txt

carl
john doe
bob smith
...

itemnames.txt

socks
shirts
shoes
...

I will be receiving the strings at prediction time. There is no way around that. In order to make training similar I am using tf.contrib.lookup

import tf.contrib.lookup

user_lookup = tf.contrib.lookup.index_table_from_file(
    vocabulary_file='usernames.txt'
)

item_lookup = tf.contrib.lookup.index_table_from_file(
    vocabulary_file='itemnames.txt'
)    

Now I have the following model defined using the keras api

import tensorflow as tf

user_input = tf.keras.layers.Input(shape=(1,), dtype=tf.int32)
item_input = tf.keras.layers.Input(shape=(1,), dtype=tf.int32)

user_embedding = tf.keras.layers.Embedding(input_dim=num_users, output_dim=10)(user_input)
item_embedding = tf.keras.layers.Embedding(input_dim=num_items, output_dim=10)(item_input)

...
output = ...
model = tf.keras.Model([user_input, item_input], output)
model.compile(...)

I am using tf.estimator for training and prediction. So my first instinct is to do the following:

my_estimator = tf.keras.estimator.model_to_estimator(keras_model=model)

tf.tables_initializer()

def train_fn(dataset_iterator):
     (username, itemname), value = dataset_iterator.get_next()
     userid = user_lookup.lookup(username)
     itemid = item_lookup.lookup(itemname)
     return (username, itemname), value

my_train_spec = tf.estimator.TrainSpec(
   input_fn=train_fn(train_data)
)

my_eval_spec = tf.estimator.EvalSpec(
   input_fn=train_fn(validation_data)
)

tf.estimator.train_and_evaluate(
    estimator=my_estimator,
    train_spec=my_train_spec,
    eval_spec=my_eval_spec
)

When I run this I get the follow error:

ValueError: Tensor("Cast_2:0", shape=(), dtype=int32) must be from the same graph as Tensor("Item-Embedding-LMF/embeddings/Read/ReadVariableOp:0", shape=(429099, 10), dtype=float32, device=/job:ps/task:1).

Can anyone recommend a solution to this problem? Or maybe even a different approach to handling this lookup?

Answer 1

Mostly, no issue with the lookup. The all the variaables should be asscociated with the same graphs May be write the model in the scope eg

def model_fn: 
    with tf.variable_scope('my_model', reuse=tf.AUTO_REUSE):
    ....
    ..
     return estimator