如何使用Tensorflow估算器对保存的模型进行评分？

Question

All,

I built a customized model for binary image classification. I managed to successfully save model using tf estimator to .pb format. My jpg image files have images in various sizes, so I have a image transformation step to transform the images to 224x224. Here is how I define serving_input_fn():

def parse_function_test(filename):
    image_string=tf.read_file(filename)
    image=tf.image.decode_jpeg(image_string, channels=3)
    shape = tf.shape(image)
    init_width = shape[0]
    init_height = shape[1]
    max_size = 224
    resized_image = resize_image_keep_aspect(image,init_width,init_height,max_size)
    image_padded = tf.image.resize_image_with_crop_or_pad(resized_image,max_size,max_size)
    final_image_padded=tf.image.convert_image_dtype(image_padded,dtype=tf.float32)

    return final_image_padded


def serving_input_fn():
    serialized_tf_example=tf.placeholder(dtype=tf.string,shape=1,name='input_tensor')
    receiver_tensors={'inputs':serialized_tf_example}
    feature_spec ={'image/encoded':tf.FixedLenFeature([],dtype=tf.string)}
    features=tf.parse_example(serialized_tf_example,feature_spec)
    jpegs=features['image/encoded']
    images=tf.map_fn(parse_function_test,jpegs,dtype=tf.float32)

    return tf.estimator.export.ServingInputReceiver(images,receiver_tensors)

This is the scoring script. I tried to score on one image file.

exported_path='./1538070515'
testimg = './test.jpg'

filename = tf.constant([testimg])

image= tf.map_fn(parse_function_test,filename,dtype=tf.float32)
def _bytes_feature(value):
  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

def main():
    with tf.Session() as sess:
         value=sess.run(image)
         tf.saved_model.loader.load(sess, 
             [tf.saved_model.tag_constants.SERVING], exported_path)
         model_input=tf.train.Example(features=tf.train.Features(feature={'image/encoded':_bytes_feature(value.tostring())}))
         predictor= tf.contrib.predictor.from_saved_model(exported_path)
         input_tensor=tf.get_default_graph().get_tensor_by_name("input_tensor:0")
         print(input_tensor)
         model_input = model_input.SerializeToString()
         output_dict=predictor({'inputs':[model_input]})
         print("probability is",output_dict['probabilities'])

if __name__=="__main__":
    main()

Got error on output_dict=predictor({'inputs':[model_input]}) line says " NotFoundError: NewRandomAccessFile failed to Create/Open: "

Where did I do wrong? I am not sure if it was correct to transformation image in the scoring script first then make it a bytes_features... Or maybe I did something wrong in serving_input_fn().

Answer 1

I figured it out...

All the image transformation inside the map function were built to the output graph already. In the scoring script, just need to encode image name string to bytes, then use this as input.

model_input=tf.train.Example(features=tf.train.Features(feature={'image/encoded':_bytes_feature(testimg.encode())}))