Issues running exported SavedModel

After using the collab to export a SavedModel. I attempted to inference it locally with the following code.

from __future__ import absolute_import, division, print_function, unicode_literals
import tensorflow as tf
import tensorflow_text
import numpy as np
physical_devices = tf.config.experimental.list_physical_devices('GPU')
if physical_devices:
  tf.config.experimental.set_memory_growth(physical_devices[0], True)

# %%
tf.compat.v1.enable_resource_variables()
path ="PATH_TO_SAVEDMODEL"
loaded = tf.saved_model.load(path, tags='serve')
print(list(loaded.signatures.keys()))
# %%

x = tf.constant(["translate English to German: this is a test.","translate English to German: this is a test."])
print(x)
# loaded.signatures['serving_default'](x)
print(loaded.signatures['serving_default'](x))

and got the following error:

InvalidArgumentError:  Could not parse example input, value: 'translate English to German: this is a test.'
	 [[{{node ParseSingleExample/ParseSingleExample}}]]
	 [[DatasetToSingleElement]] [Op:__inference_pruned_275843]

Function call stack:
pruned

for inference I am using:

ubuntu: 18.04 CUDA Version: 10.2 tensorflow: 2.1

I’m sure that I’m missing something simple, but any advice on how to solve this would be greatly appreciated.

[adarob]

check out https://github.com/google-research/text-to-text-transfer-transformer/pull/100

[adarob]

I believe it is expecting a serialized tf.train.Example proto, not a raw string.

You should try:

feature = {"input": tf.train.Feature(bytes_list=tf.train.BytesList(value=[b"test"]))}
input_str = tf.train.Example(features=tf.train.Features(feature=feature)).SerializeAsString()

Clearly this is not a great API. We should look into being able to pass strings directly.