Error Processing Input for GCP ML Engine Prediction - tensorflow

Question

I have a TensorFlow model on GCP ML Engine, however I have a problem with the JSON string below:
from googleapiclient import discovery
from oauth2client.client import GoogleCredentials
import json
credentials = GoogleCredentials.get_application_default()
api = discovery.build('ml', 'v1', credentials=credentials,
discoveryServiceUrl='https://storage.googleapis.com/cloud-ml/discovery/ml_v1_discovery.json')
request_data = {'instances':
[{
'inputs':{
'clump_thickness': 2,
'size_uniformity': 1,
'shape_uniformity': 1,
'marginal_adhesion': 1,
'epithelial_size': 2,
'bland_chromatin': 1,
'bare_nucleoli': 2,
'normal_nucleoli': 1,
'mitoses': 1
}
}]
}
parent = 'projects/%s/models/%s/versions/%s' % (PROJECT,
'breastCancer_optimized_06152018_2_2_a', 'v1')
response = api.projects().predict(body=request_data, name=parent).execute()
print(response)
I get the following error:
{'error': "Prediction failed: Error processing input: Expected string, got {u'epithelial_size': 2, u'marginal_adhesion': 1, u'clump_thickness': 2, u'size_uniformity': 1, u'shape_uniformity': 1, u'normal_nucleoli': 1, u'mitoses': 1, u'bland_chromatin': 1, u'bare_nucleoli': 2} of type 'dict' instead."}
I can't seem to format request_data properly. Does anyone see what is wrong?
original serving function:
clump_thickness = tf.feature_column.numeric_column("clump_thickness");
size_uniformity = tf.feature_column.numeric_column("size_uniformity");
shape_uniformity = tf.feature_column.numeric_column("shape_uniformity");
marginal_adhesion = tf.feature_column.numeric_column("marginal_adhesion");
epithelial_size = tf.feature_column.numeric_column("epithelial_size");
bare_nucleoli = tf.feature_column.numeric_column("bare_nucleoli");
bland_chromatin = tf.feature_column.numeric_column("bland_chromatin");
normal_nucleoli = tf.feature_column.numeric_column("normal_nucleoli");
mitoses = tf.feature_column.numeric_column("mitoses");
feature_columns = [clump_thickness, size_uniformity, shape_uniformity, marginal_adhesion, epithelial_size,
bare_nucleoli, bland_chromatin, normal_nucleoli, mitoses];
feature_spec = tf.feature_column.make_parse_example_spec(feature_columns);
export_input_fn = tf.estimator.export.build_parsing_serving_input_receiver_fn(feature_spec);
estimator.export_savedmodel(output_dir, export_input_fn, as_text=False)
Then I tried:
def serving_input_fn():
feature_placeholders = {
'clump_thickness' : tf.placeholder(tf.float32, [None]),
'size_uniformity' : tf.placeholder(tf.float32, [None]),
'shape_uniformity' : tf.placeholder(tf.float32, [None]),
'marginal_adhesion' : tf.placeholder(tf.float32, [None]),
'epithelial_size' : tf.placeholder(tf.float32, [None]),
'bare_nucleoli' : tf.placeholder(tf.float32, [None]),
'bland_chromatin' : tf.placeholder(tf.float32, [None]),
'normal_nucleoli' : tf.placeholder(tf.float32, [None]),
'mitoses' : tf.placeholder(tf.float32, [None]),
}
features = feature_placeholders # no transformation needed
return tf.estimator.export.ServingInputReceiver(features, feature_placeholders)
And in the train_and_eval function:
estimator.export_savedmodel(output_dir, serving_input_fn, as_text=False)
But now I get the following error:
{'error': "Prediction failed: Expected tensor name: inputs, got tensor name: [u'epithelial_size', u'marginal_adhesion', u'clump_thickness', u'size_uniformity', u'shape_uniformity', u'normal_nucleoli', u'mitoses', u'bland_chromatin', u'bare_nucleoli']."}
The estimator.export_savedmodel appears to create a model which requires a tensor input(in the request_data line).
When I use the model created with either serving function the following works fine:
predict_fn = tf.contrib.predictor.from_saved_model("gs://test-
203900/breastCancer_optimized_06182018/9/1529432417")
# Test inputs represented by Pandas DataFrame.
inputs = pd.DataFrame({
'clump_thickness': [2,5,4],
'size_uniformity': [1,10,8],
'shape_uniformity': [1,10,6],
'marginal_adhesion': [1,3,4],
'epithelial_size': [2,7,3],
'bland_chromatin': [1,3,4],
'bare_nucleoli': [2,8,10],
'normal_nucleoli': [1,10,6],
'mitoses': [1,2,1],
})
# Convert input data into serialized Example strings.
examples = []
for index, row in inputs.iterrows():
feature = {}
for col, value in row.iteritems():
feature[col] =
tf.train.Feature(float_list=tf.train.FloatList(value=[value]))
example = tf.train.Example(
features=tf.train.Features(
feature=feature
)
)
examples.append(example.SerializeToString())
# Make predictions.
predictions = predict_fn({'inputs': examples})

Answer 1

It depends on what your serving input function is. It appears from the error message that 'inputs' needs to be a string (maybe comma-separated?)
Try this:
saved_model_cli show --dir $MODEL_LOCATION --tag_set serve --signature_def serving_default
It will tell you what your serving input function is set to.
I suspect that what you want is for your serving input function to be:
def serving_input_fn():
feature_placeholders = {
'size_uniformity' : tf.placeholder(tf.float32, [None]),
'shape_uniformity' : tf.placeholder(tf.float32, [None])
}
features = feature_placeholders # no transformation needed
return tf.estimator.export.ServingInputReceiver(features, feature_placeholders)
and for your input format to be:
request_data = {'instances':
[
{
'size_uniformity': 1,
'shape_uniformity': 1
}
]
}