I am new to Tensorflow. I am trying to build and serve a model using Estimator on Google ML Engine. However, I am not sure how I can save the model for serving after trying a few ways.
I have successfully trained the model with acceptable accuracy. When I was trying to save the model for serving, I searched around and found a few ways to do so. However, I still ran into a number of problems...
I tried 3 ways of exporting based on suggestions made for a few other questions posted:
1) Getting a serialized example as input - I ran into an error "TypeError: Object of type bytes is not JSON serializable". Also, I couldn't find a good way to feed a serialized example for serving effectively. As I am using ML Engine for serving, it seems it would be easier to use a JSON input.
2) Getting a JSON as input with "basic" pre-processing - I was able to successfully export the model. After loading the model onto ML Engine, I tried making a few predictions. Although a prediction result was returned, I found that, no matter how I change the JSON inputs, the same result was returned. I looked at the validation results obtained during the training. The model should be able to return variety of results. I thought there is something wrong with the pre-processing within the serving function, so I tried the third way...
3) JSON input with the "same" pre-processing - I couldn't get my head around this, but I think it might be needed to do exactly the same pre-processing as how I process my data during model training. However, as the serving input function makes use of tf.placeholders, I have no idea how I could replicate the same pre-processing to make the exported model works...
(Please pardon my bad coding style...)
Training code:
col_names = ['featureA','featureB','featureC']
target_name = 'langIntel'
col_def = {}
col_def['featureA'] = {'type':'float','tfType':tf.float32,'len':'fixed'}
col_def['featureB'] = {'type':'int','tfType':tf.int64,'len':'fixed'}
col_def['featureC'] = {'type':'bytes','tfType':tf.string,'len':'var'}
def _float_feature(value):
if not isinstance(value, list): value = [value]
return tf.train.Feature(float_list=tf.train.FloatList(value=value))
def _int_feature(value):
if not isinstance(value, list): value = [value]
return tf.train.Feature(int64_list=tf.train.Int64List(value=value))
def _bytes_feature(value):
if not isinstance(value, list): value = [value]
return tf.train.Feature(
bytes_list=tf.train.BytesList(
value=[p.encode('utf-8') for p in value]
)
)
functDict = {'float':_float_feature,
'int':_int_feature,'bytes':_bytes_feature
}
training_targets = []
# Omitted validatin partition
with open('[JSON FILE PATH]') as jfile:
json_data_input = json.load(jfile)
random.shuffle(json_data_input)
with tf.python_io.TFRecordWriter('savefile1.tfrecord') as writer:
for item in json_data_input:
if item[target_name] > 0:
feature = {}
for col in col_names:
feature[col] = functDict[col_def[col]['type']](item[col])
training_targets.append(item[target_name])
example = tf.train.Example(
features=tf.train.Features(feature=feature)
)
writer.write(example.SerializeToString())
def _parse_function(example_proto):
example = {}
for col in col_names:
if col_def[col]['len'] == 'fixed':
example[col] = tf.FixedLenFeature([], col_def[col]['tfType'])
else:
example[col] = tf.VarLenFeature(col_def[col]['tfType'])
parsed_example = tf.parse_single_example(example_proto, example)
features = {}
for col in col_names:
features[col] = parsed_example[col]
labels = parsed_example.get(target_name)
return features, labels
def my_input_fn(batch_size=1,num_epochs=None):
dataset = tf.data.TFRecordDataset('savefile1.tfrecord')
dataset = dataset.map(_parse_function)
dataset = dataset.shuffle(10000)
dataset = dataset.repeat(num_epochs)
dataset = dataset.batch(batch_size)
iterator = dataset.make_one_shot_iterator()
features, labels = iterator.get_next()
return features, labels
allColumns = None
def train_model(
learning_rate,
n_trees,
n_batchespl,
batch_size):
periods = 10
vocab_list = ('vocab1', 'vocab2', 'vocab3')
featureA_bucket = tf.feature_column.bucketized_column(
tf.feature_column.numeric_column(
key="featureA",dtype=tf.int64
), [5,10,15]
)
featureB_bucket = tf.feature_column.bucketized_column(
tf.feature_column.numeric_column(
key="featureB",dtype=tf.float32
), [0.25,0.5,0.75]
)
featureC_cat = tf.feature_column.indicator_column(
tf.feature_column.categorical_column_with_vocabulary_list(
key="featureC",vocabulary_list=vocab_list,
num_oov_buckets=1
)
)
theColumns = [featureA_bucket,featureB_bucket,featureC_cat]
global allColumns
allColumns = theColumns
regressor = tf.estimator.BoostedTreesRegressor(
feature_columns=theColumns,
n_batches_per_layer=n_batchespl,
n_trees=n_trees,
learning_rate=learning_rate
)
training_input_fn = lambda: my_input_fn(batch_size=batch_size,num_epochs=5)
predict_input_fn = lambda: my_input_fn(num_epochs=1)
regressor.train(
input_fn=training_input_fn
)
# omitted evaluation part
return regressor
regressor = train_model(
learning_rate=0.05,
n_trees=100,
n_batchespl=50,
batch_size=20)
Export Trial 1:
def _serving_input_receiver_fn():
serialized_tf_example = tf.placeholder(dtype=tf.string, shape=None,
name='input_example_tensor'
)
receiver_tensors = {'examples': serialized_tf_example}
features = tf.parse_example(serialized_tf_example, feature_spec)
return tf.estimator.export.ServingInputReceiver(features,
receiver_tensors
)
servable_model_dir = "[OUT PATH]"
servable_model_path = regressor.export_savedmodel(servable_model_dir,
_serving_input_receiver_fn
)
Export Trial 2:
def serving_input_fn():
feature_placeholders = {
'featureA': tf.placeholder(tf.int64, [None]),
'featureB': tf.placeholder(tf.float32, [None]),
'featureC': tf.placeholder(tf.string, [None, None])
}
receiver_tensors = {'inputs': feature_placeholders}
feature_spec = tf.feature_column.make_parse_example_spec(allColumns)
features = tf.parse_example(feature_placeholders, feature_spec)
return tf.estimator.export.ServingInputReceiver(features,
feature_placeholders
)
servable_model_dir = "[OUT PATH]"
servable_model_path = regressor.export_savedmodel(
servable_model_dir, serving_input_fn
)
Export Trial 3:
def serving_input_fn():
feature_placeholders = {
'featureA': tf.placeholder(tf.int64, [None]),
'featureB': tf.placeholder(tf.float32, [None]),
'featureC': tf.placeholder(tf.string, [None, None])
}
def toBytes(t):
t = str(t)
return t.encode('utf-8')
tmpFeatures = {}
tmpFeatures['featureA'] = tf.train.Feature(
int64_list=feature_placeholders['featureA']
)
# TypeError: Parameter to MergeFrom() must be instance
# of same class: expected tensorflow.Int64List got Tensor.
tmpFeatures['featureB'] = tf.train.Feature(
float_list=feature_placeholders['featureB']
)
tmpFeatures['featureC'] = tf.train.Feature(
bytes_list=feature_placeholders['featureC']
)
tmpExample = tf.train.Example(
features=tf.train.Features(feature=tmpFeatures)
)
tmpExample_proto = tmpExample.SerializeToString()
example = {}
for key, tensor in feature_placeholders.items():
if col_def[key]['len'] == 'fixed':
example[key] = tf.FixedLenFeature(
[], col_def[key]['tfType']
)
else:
example[key] = tf.VarLenFeature(
col_def[key]['tfType']
)
parsed_example = tf.parse_single_example(
tmpExample_proto, example
)
features = {}
for key in tmpFeatures.keys():
features[key] = parsed_example[key]
return tf.estimator.export.ServingInputReceiver(
features, feature_placeholders
)
servable_model_dir = "[OUT PATH]"
servable_model_path = regressor.export_savedmodel(
servable_model_dir, serving_input_fn
)
How should the serving input function be structured in order for a JSON file to be inputted for prediction? Many thanks for any insights!
