问题
I followed the codelab TensorFlow For Poets for transfer learning using inception_v3. It generates retrained_graph.pb and retrained_labels.txt files, which can used to make predictions locally (running label_image.py).
Then, I wanted to deploy this model to Cloud ML Engine, so that I could make online predictions. For that, I had to export the retrained_graph.pb to SavedModel format. I managed to do it by following the indications in this answer from Google's @rhaertel80 and this python file from the Flowers Cloud ML Engine Tutorial. Here is my code:
import tensorflow as tf
from tensorflow.contrib import layers
from tensorflow.python.saved_model import builder as saved_model_builder
from tensorflow.python.saved_model import signature_constants
from tensorflow.python.saved_model import signature_def_utils
from tensorflow.python.saved_model import tag_constants
from tensorflow.python.saved_model import utils as saved_model_utils
export_dir = '../tf_files/saved7'
retrained_graph = '../tf_files/retrained_graph2.pb'
label_count = 5
def build_signature(inputs, outputs):
signature_inputs = { key: saved_model_utils.build_tensor_info(tensor) for key, tensor in inputs.items() }
signature_outputs = { key: saved_model_utils.build_tensor_info(tensor) for key, tensor in outputs.items() }
signature_def = signature_def_utils.build_signature_def(
signature_inputs,
signature_outputs,
signature_constants.PREDICT_METHOD_NAME
)
return signature_def
class GraphReferences(object):
def __init__(self):
self.examples = None
self.train = None
self.global_step = None
self.metric_updates = []
self.metric_values = []
self.keys = None
self.predictions = []
self.input_jpeg = None
class Model(object):
def __init__(self, label_count):
self.label_count = label_count
def build_image_str_tensor(self):
image_str_tensor = tf.placeholder(tf.string, shape=[None])
def decode_and_resize(image_str_tensor):
return image_str_tensor
image = tf.map_fn(
decode_and_resize,
image_str_tensor,
back_prop=False,
dtype=tf.string
)
return image_str_tensor
def build_prediction_graph(self, g):
tensors = GraphReferences()
tensors.examples = tf.placeholder(tf.string, name='input', shape=(None,))
tensors.input_jpeg = self.build_image_str_tensor()
keys_placeholder = tf.placeholder(tf.string, shape=[None])
inputs = {
'key': keys_placeholder,
'image_bytes': tensors.input_jpeg
}
keys = tf.identity(keys_placeholder)
outputs = {
'key': keys,
'prediction': g.get_tensor_by_name('final_result:0')
}
return inputs, outputs
def export(self, output_dir):
with tf.Session(graph=tf.Graph()) as sess:
with tf.gfile.GFile(retrained_graph, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, name="")
g = tf.get_default_graph()
inputs, outputs = self.build_prediction_graph(g)
signature_def = build_signature(inputs=inputs, outputs=outputs)
signature_def_map = {
signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature_def
}
builder = saved_model_builder.SavedModelBuilder(output_dir)
builder.add_meta_graph_and_variables(
sess,
tags=[tag_constants.SERVING],
signature_def_map=signature_def_map
)
builder.save()
model = Model(label_count)
model.export(export_dir)
This code generates a saved_model.pb file, which I then used to create the Cloud ML Engine model. I can get predictions from this model using gcloud ml-engine predict --model my_model_name --json-instances request.json, where the contents of request.json are:
{ "key": "0", "image_bytes": { "b64": "jpeg_image_base64_encoded" } }
However, no matter which jpeg I encode in the request, I always get the exact same wrong predictions:
Prediction output
I guess the problem is in the way the CloudML Prediction API passes the base64 encoded image bytes to the input tensor "DecodeJpeg/contents:0" of inception_v3 ("build_image_str_tensor()" method in the previous code). Any clue on how can I solve this issue and have my locally retrained model serving correct predictions on Cloud ML Engine?
(Just to make it clear, the problem is not in retrained_graph.pb, as it makes correct predictions when I run it locally; nor is it in request.json, because the same request file worked without problems when following the Flowers Cloud ML Engine Tutorial pointed above.)
回答1:
First, a general warning. The TensorFlow for Poets codelab was not written in a way that is very amenable to production serving (partly manifested by the workarounds you are having to implement). You would normally export a prediction-specific graph that doesn't contain all of the extra training ops. So while we can try and hack something together that works, extra work may be needed to productionize this graph.
The approach of your code appears to be to import one graph, add some placeholders, and then export the result. This is generally fine. However, in the code shown in the question, you are adding input placeholders without actually connecting them to anything in the imported graph. You end up with a graph containing multiple disconnected subgraphs, something like (excuse the crude diagram):
image_str_tensor [input=image_bytes] -> <nothing>
keys_placeholder [input=key] -> identity [output=key]
inception_subgraph -> final_graph [output=prediction]
By inception_subgraph I mean all of the ops that you are importing.
So image_bytes is effectively a no-op and is ignored; key gets passed through; and prediction contains the result of running the inception_subgraph; since it's not using the input you are passing, it's returning the same result everytime (though I admit I actually expected an error here).
To address this problem, we would need to connect the placeholder you've created to the one that already exists in inception_subgraph to create a graph more or less like this:
image_str_tensor [input=image_bytes] -> inception_subgraph -> final_graph [output=prediction]
keys_placeholder [input=key] -> identity [output=key]
Note that image_str_tensor is going to be a batch of images, as required by the prediction service, but the inception graph's input is actually a single image. In the interest of simplicity, we're going to address this in a hacky way: we'll assume we'll be sending images one-by-one. If we ever send more than one image per request, we'll get errors. Also, batch prediction will never work.
The main change you need is the import statement, which connects the placeholder we've added to the existing input in the graph (you'll also see the code for changing the shape of the input):
Putting it all together, we get something like:
import tensorflow as tf
from tensorflow.contrib import layers
from tensorflow.python.saved_model import builder as saved_model_builder
from tensorflow.python.saved_model import signature_constants
from tensorflow.python.saved_model import signature_def_utils
from tensorflow.python.saved_model import tag_constants
from tensorflow.python.saved_model import utils as saved_model_utils
export_dir = '../tf_files/saved7'
retrained_graph = '../tf_files/retrained_graph2.pb'
label_count = 5
class Model(object):
def __init__(self, label_count):
self.label_count = label_count
def build_prediction_graph(self, g):
inputs = {
'key': keys_placeholder,
'image_bytes': tensors.input_jpeg
}
keys = tf.identity(keys_placeholder)
outputs = {
'key': keys,
'prediction': g.get_tensor_by_name('final_result:0')
}
return inputs, outputs
def export(self, output_dir):
with tf.Session(graph=tf.Graph()) as sess:
# This will be our input that accepts a batch of inputs
image_bytes = tf.placeholder(tf.string, name='input', shape=(None,))
# Force it to be a single input; will raise an error if we send a batch.
coerced = tf.squeeze(image_bytes)
# When we import the graph, we'll connect `coerced` to `DecodeJPGInput:0`
input_map = {'DecodeJPGInput:0': coerced}
with tf.gfile.GFile(retrained_graph, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, input_map=input_map, name="")
keys_placeholder = tf.placeholder(tf.string, shape=[None])
inputs = {'image_bytes': image_bytes, 'key': keys_placeholder}
keys = tf.identity(keys_placeholder)
outputs = {
'key': keys,
'prediction': tf.get_default_graph().get_tensor_by_name('final_result:0')}
}
tf.simple_save(sess, output_dir, inputs, outputs)
model = Model(label_count)
model.export(export_dir)
回答2:
I believe that your error is quite simple to solve:
{ "key": "0", "image_bytes": { "b64": "jpeg_image_base64_encoded" } }
You used " to specify what, I believe, is a string. By doing that, your program is reading jpeg_image_base64_encoded instead of the actual value of the variable.
That's why you get always the same prediction.
回答3:
For anyone working on deploying TensorFlow image-based models on Google Cloud ML, in particular trying to get the base64 encoding working for images (as discussed in this question), I'd recommend also having a look at the following repo that I put together. I spent a lot of time working through the deployment process and was only able to find partial information across the web and on stack overflow. This repo has a full working version of deploying a TensorFlow tf.keras model onto google cloud ML and I think it will be of help to people who are facing the same challenges I faced. Here's the github link:
https://github.com/mhwilder/tf-keras-gcloud-deployment.
The repo covers the following topics:
- Training a fully convolutional tf.keras model locally (mostly just to have a model for testing the next parts)
- Example code for exporting models that work with the Cloud ML Engine
- Three model versions that accept different JSON input types (1. An image converted to a simple list string, 2. An image converted to a base64 encoded string, and 3. A URL that points to an image in a Google Storage bucket)
- Instructions and references for general Google Cloud Platform setup
- Code for preparing the input JSON files for the 3 different input types
- Google Cloud ML model and version creation instructions from the console
- Examples using the Google Cloud SDK to call predict on the models
来源：https://stackoverflow.com/questions/47558050/retrained-inception-v3-model-deployed-in-cloud-ml-engine-always-outputs-the-same