I've trained a model using the Getting Started tutorial from Google Cloud ML Engine as a reference. I could manage to deploy and serve this model on Google Cloud ML without problems.
Now I'm trying to serve it using Tensorflow Serving but I'm getting the message error below:
2017-03-17 19:20:17.064146: W tensorflow_serving/sources/storage_path/file_system_storage_path_source.cc:204] No versions of servable default found under base path /serving/tf_models/extrato/output/
The command line call I'm using to start de Tensorflow serving is:
root@df98954689a1:/serving# bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server --port=9000 --model_base_path=/serving/tf_models/extrato/output/
The content of the output folder is:
root@df98954689a1:/serving# ls -la tf_models/extrato/output
total 119740
drwxr-xr-x 4 root root 4096 Mar 17 17:02 .
drwxr-xr-x 3 root root 4096 Mar 17 17:02 ..
-rw-r--r-- 1 root root 184 Mar 17 17:02 checkpoint
drwxr-xr-x 2 root root 4096 Mar 17 17:02 eval
-rw-r--r-- 1 root root 96390060 Mar 17 17:02 events.out.tfevents.1489705843.elio-MS-7A66
drwxr-xr-x 3 root root 4096 Mar 17 17:02 export
-rw-r--r-- 1 root root 1362798 Mar 17 17:02 graph.pbtxt
-rw-r--r-- 1 root root 7633781 Mar 17 17:02 model.ckpt-1000001.data-00000-of-00001
-rw-r--r-- 1 root root 1975 Mar 17 17:02 model.ckpt-1000001.index
-rw-r--r-- 1 root root 637623 Mar 17 17:02 model.ckpt-1000001.meta
-rw-r--r-- 1 root root 7633781 Mar 17 17:02 model.ckpt-2.data-00000-of-00001
-rw-r--r-- 1 root root 1975 Mar 17 17:02 model.ckpt-2.index
-rw-r--r-- 1 root root 637623 Mar 17 17:02 model.ckpt-2.meta
-rw-r--r-- 1 root root 7633781 Mar 17 17:02 model.ckpt-566170.data-00000-of-00001
-rw-r--r-- 1 root root 1975 Mar 17 17:02 model.ckpt-566170.index
-rw-r--r-- 1 root root 637623 Mar 17 17:02 model.ckpt-566170.meta
UPDATE: I tried to use the frozen model (.pb file and variables folder) which are indeed the folder I use to deploy the model on Google Cloud ML Engine, but got the same error message.
These files are located inside the folder below:
root@d4f1c917b59d:/serving# ls -la tf_models/extrato/output/export/Servo/1489706933289/
total 356
drwxr-xr-x 3 root root 4096 Mar 17 17:02 .
drwxr-xr-x 3 root root 4096 Mar 17 17:02 ..
-rw-r--r-- 1 root root 348848 Mar 17 17:02 saved_model.pb
drwxr-xr-x 2 root root 4096 Mar 17 17:02 variables
The code I've used to train and export the model is:
import argparse
import model
import tensorflow as tf
from tensorflow.contrib.learn.python.learn import learn_runner
from tensorflow.contrib.learn.python.learn.utils import (
saved_model_export_utils)
def generate_experiment_fn(train_files,
eval_files,
num_epochs=None,
train_batch_size=40,
eval_batch_size=40,
embedding_size=8,
first_layer_size=100,
num_layers=4,
scale_factor=0.7,
**experiment_args):
"""Create an experiment function given hyperparameters.
See command line help text for description of args.
Returns:
A function (output_dir) -> Experiment where output_dir is a string
representing the location of summaries, checkpoints, and exports.
this function is used by learn_runner to create an Experiment which
executes model code provided in the form of an Estimator and
input functions.
All listed arguments in the outer function are used to create an
Estimator, and input functions (training, evaluation, serving).
Unlisted args are passed through to Experiment.
"""
# Check verbose logging flag
verbose_logging = experiment_args.pop('verbose_logging')
model.set_verbose_logging(verbose_logging)
def _experiment_fn(output_dir):
# num_epochs can control duration if train_steps isn't
# passed to Experiment
train_input = model.generate_input_fn(
train_files,
num_epochs=num_epochs,
batch_size=train_batch_size,
)
# Don't shuffle evaluation data
eval_input = model.generate_input_fn(
eval_files,
batch_size=eval_batch_size,
shuffle=False
)
return tf.contrib.learn.Experiment(
model.build_estimator(
output_dir,
embedding_size=embedding_size,
# Construct layers sizes with exponetial decay
hidden_units=[
max(2, int(first_layer_size * scale_factor**i))
for i in range(num_layers)
]
),
train_input_fn=train_input,
eval_input_fn=eval_input,
# export strategies control the prediction graph structure
# of exported binaries.
export_strategies=[saved_model_export_utils.make_export_strategy(
model.serving_input_fn,
default_output_alternative_key=None,
exports_to_keep=1
)],
**experiment_args
)
return _experiment_fn
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# Input Arguments
parser.add_argument(
'--train-files',
help='GCS or local paths to training data',
nargs='+',
required=True
)
parser.add_argument(
'--num-epochs',
help="""\
Maximum number of training data epochs on which to train.
If both --max-steps and --num-epochs are specified,
the training job will run for --max-steps or --num-epochs,
whichever occurs first. If unspecified will run for --max-steps.\
""",
type=int,
)
parser.add_argument(
'--train-batch-size',
help='Batch size for training steps',
type=int,
default=40
)
parser.add_argument(
'--eval-batch-size',
help='Batch size for evaluation steps',
type=int,
default=40
)
parser.add_argument(
'--train-steps',
help="""\
Steps to run the training job for. If --num-epochs is not specified,
this must be. Otherwise the training job will run indefinitely.\
""",
type=int
)
parser.add_argument(
'--eval-steps',
help='Number of steps to run evalution for at each checkpoint',
default=100,
type=int
)
parser.add_argument(
'--eval-files',
help='GCS or local paths to evaluation data',
nargs='+',
required=True
)
# Training arguments
parser.add_argument(
'--embedding-size',
help='Number of embedding dimensions for categorical columns',
default=8,
type=int
)
parser.add_argument(
'--first-layer-size',
help='Number of nodes in the first layer of the DNN',
default=100,
type=int
)
parser.add_argument(
'--num-layers',
help='Number of layers in the DNN',
default=4,
type=int
)
parser.add_argument(
'--scale-factor',
help='How quickly should the size of the layers in the DNN decay',
default=0.7,
type=float
)
parser.add_argument(
'--job-dir',
help='GCS location to write checkpoints and export models',
required=True
)
# Argument to turn on all logging
parser.add_argument(
'--verbose-logging',
default=False,
type=bool,
help='Switch to turn on or off verbose logging and warnings'
)
# Experiment arguments
parser.add_argument(
'--eval-delay-secs',
help='How long to wait before running first evaluation',
default=10,
type=int
)
parser.add_argument(
'--min-eval-frequency',
help='Minimum number of training steps between evaluations',
default=1,
type=int
)
args = parser.parse_args()
arguments = args.__dict__
job_dir = arguments.pop('job_dir')
print('Starting Census: Please lauch tensorboard to see results: tensorboard --logdir=$MODEL_DIR')
# Run the training job
# learn_runner pulls configuration information from environment
# variables using tf.learn.RunConfig and uses this configuration
# to conditionally execute Experiment, or param server code
learn_runner.run(generate_experiment_fn(**arguments), job_dir)
Does anyone have any tip about what I'm doing wrong?
Best Regards!
TensorFlow serving expects you to point to a base directory which includes a version subdirectory. In your case, "Servo" is the directory you want to point to and "1489706933289" is the directory for the version.
The following should work:
bazel-bin/tensorflow_serving/model_servers/tensorflow_model_server \
--port=9000 \
--model_base_path=/serving/tf_models/extrato/output/Servo
(Note the addition of "Servo" to the base path, and the absence of "1489706933289")
Note that in CloudML you directly deploy a version, so you'll want to point to a subdirectory on GCS akin to gs://my_bucket/.../tf_models/extrato/output/Servo/1489706933289