Prediction failed: contents must be scalar - tensorflow

Question

I have successfully trained, exported and uploaded my 'retrained_graph.pb' to ML Engine. My export script is as follows:
import tensorflow as tf
from tensorflow.python.saved_model import signature_constants
from tensorflow.python.saved_model import tag_constants
from tensorflow.python.saved_model import builder as saved_model_builder
input_graph = 'retrained_graph.pb'
saved_model_dir = 'my_model'
with tf.Graph().as_default() as graph:
# Read in the export graph
with tf.gfile.FastGFile(input_graph, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, name='')
# Define SavedModel Signature (inputs and outputs)
in_image = graph.get_tensor_by_name('DecodeJpeg/contents:0')
inputs = {'image_bytes': tf.saved_model.utils.build_tensor_info(in_image)}
out_classes = graph.get_tensor_by_name('final_result:0')
outputs = {'prediction_bytes': tf.saved_model.utils.build_tensor_info(out_classes)}
signature = tf.saved_model.signature_def_utils.build_signature_def(
inputs=inputs,
outputs=outputs,
method_name='tensorflow/serving/predict'
)
with tf.Session(graph=graph) as sess:
# Save out the SavedModel.
b = saved_model_builder.SavedModelBuilder(saved_model_dir)
b.add_meta_graph_and_variables(sess,
[tf.saved_model.tag_constants.SERVING],
signature_def_map={'serving_default': signature})
b.save()
I build my prediction Json using the following:
# Copy the image to local disk.
gsutil cp gs://cloud-ml-data/img/flower_photos/tulips/4520577328_a94c11e806_n.jpg flower.jpg
# Create request message in json format.
python -c 'import base64, sys, json; img = base64.b64encode(open(sys.argv[1], "rb").read()); print json.dumps({"image_bytes": {"b64": img}}) ' flower.jpg &> request.json
# Call prediction service API to get classifications
gcloud ml-engine predict --model ${MODEL_NAME} --json-instances request.json
However this fails with the response:
{
"error": "Prediction failed: Error during model execution: AbortionError(code=StatusCode.INVALID_ARGUMENT, details=\"contents must be scalar, got shape [1]\n\t [[Node: Deco
deJpeg = DecodeJpeg[_output_shapes=[[?,?,3]], acceptable_fraction=1, channels=3, dct_method=\"\", fancy_upscaling=true, ratio=1, try_recover_truncated=false, _device=\"/job:l
ocalhost/replica:0/task:0/device:CPU:0\"](_arg_DecodeJpeg/contents_0_0)]]\")"
}
Any help appreciated, I'm so close I can taste it :D

Answer 1

Why do you have this line:
in_image = graph.get_tensor_by_name('DecodeJpeg/contents:0')?
inputs = {'image_bytes': tf.saved_model.utils.build_tensor_info(in_image)}
The shape here is scalar. Can you make sure you create input with shape [None]
https://github.com/GoogleCloudPlatform/cloudml-samples/blob/master/flowers/trainer/model.py#L364

Answer 2

The server will decode and batch all the inputs. So the input to your graph is essentially [base64_decode("xxx")] where you actually want to feed base64_decode("xxx") since the op takes a string type tensor. Server-side assumes the shape of input as [None, ] i.e. the first dimension can be anything for batching. In your case, [None]. You might want to create a tensor of that shape and then feed that into the op.