How to pass deep learning model data to map function in Spark

问题

I have a very simple use-case where I am reading large number of images as rdd from s3 using sc.binaryFiles method. Once this RDD is created I am passing the content inside the rdd to the vgg16 feature extractor function. So, in this I will need the model data using which the feature extraction will be done, so I am putting the model data into broadcast variable and then accesing the value in each map function. Below is the code:-

s3_files_rdd = sc.binaryFiles(RESOLVED_IMAGE_PATH)

s3_files_rdd.persist()

model_data = initVGG16()
broadcast_model = sc.broadcast(model_data)

features_rdd = s3_files_rdd.mapPartitions(extract_features_)

response_rdd = features_rdd.map(lambda x: (x[0], write_to_s3(x, OUTPUT, FORMAT_NAME)))

extract_features_ method:-

def extract_features_(xs):
    model_data = initVGG16()
    for k, v in xs:
        yield k, extract_features2(model_data,v)

extract_features method:-

from keras.preprocessing import image
from keras.applications.vgg16 import VGG16
from keras.models import Model
from io import BytesIO
from keras.applications.vgg16 import preprocess_input
def extract_features(model,obj):
    try:
        print('executing vgg16 feature extractor...')
        img = image.load_img(BytesIO(obj), target_size=(224, 224,3))
        img_data = image.img_to_array(img)
        img_data = np.expand_dims(img_data, axis=0)
        img_data = preprocess_input(img_data)
        vgg16_feature = model.predict(img_data)[0]
        print('++++++++++++++++++++++++++++',vgg16_feature.shape)
        return vgg16_feature
    except Exception as e:
        print('Error......{}'.format(e.args))
        return []

write to s3 method:-

def write_to_s3(rdd, output_path, format_name):
    file_path = rdd[0]
    file_name_without_ext = get_file_name_without_ext(file_name)
    bucket_name = output_path.split('/', 1)[0]

    final_path = 'deepak' + '/' + file_name_without_ext + '.' + format_name

    LOGGER.info("Saving to S3....")
    cci = cc.get_interface(bucket_name, ACCESS_KEY=os.environ.get("AWS_ACCESS_KEY_ID"),
                           SECRET_KEY=os.environ.get("AWS_SECRET_ACCESS_KEY"), endpoint_url='https://s3.amazonaws.com')
    response = cci.upload_npy_array(final_path, rdd[1])
    return response

Inside the write_to_s3 method I am getting the RDD, extracting the key name to be saved and bucket. then using a library called cottoncandy to drectly save the RDD content which is numpy array in my case instead of saving any intermediate file.

I am getting below error :-

127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/cloudpickle.py", line 600, in save_reduce
    save(state)
  File "/usr/lib64/python2.7/pickle.py", line 286, in save
    f(self, obj) # Call unbound method with explicit self
  File "/usr/lib64/python2.7/pickle.py", line 655, in save_dict
    self._batch_setitems(obj.iteritems())
  File "/usr/lib64/python2.7/pickle.py", line 687, in _batch_setitems
    save(v)
  File "/usr/lib64/python2.7/pickle.py", line 306, in save
    rv = reduce(self.proto)
TypeError: can't pickle thread.lock objects
Traceback (most recent call last):
  File "one_file5.py", line 98, in <module>
    run()
  File "one_file5.py", line 89, in run
    LOGGER.info('features_rdd rdd created,...... %s',features_rdd.count())    
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/rdd.py", line 1041, in count
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/rdd.py", line 1032, in sum
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/rdd.py", line 906, in fold
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/rdd.py", line 809, in collect
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/rdd.py", line 2455, in _jrdd
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/rdd.py", line 2388, in _wrap_function
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/rdd.py", line 2374, in _prepare_for_python_RDD
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/serializers.py", line 464, in dumps
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/cloudpickle.py", line 704, in dumps
  File "/mnt/yarn/usercache/hadoop/appcache/application_1541576150127_0010/container_1541576150127_0010_01_000001/pyspark.zip/pyspark/cloudpickle.py", line 162, in dump
pickle.PicklingError: Could not serialize object: TypeError: can't pickle thread.lock objects.

When I am commenting out the the code part of features_rdd, then the program runs fine which means something is not proper in the features_rdd part. Not sure what I am doing wrong here.

I am running the program in AWS EMR, with 4 executors. executor core 7 executor RAM 8GB Spark version 2.2.1

回答1:

Replace your current code with mapPartitions:

def extract_features_(xs):
    model_data = initVGG16()
    for k, v in xs:
        yield k, extract_features(model_data, v)

features_rdd = s3_files_rdd.mapPartitions(extract_features_)

来源：https://stackoverflow.com/questions/53187570/how-to-pass-deep-learning-model-data-to-map-function-in-spark

标签

apache-spark

keras

pyspark

deep-learning