AWS Machine Learning Blog

Build your own object classification model in SageMaker and import it to DeepLens

We are excited to launch a new feature for AWS DeepLens that allows you to import models trained using Amazon SageMaker directly into the AWS DeepLens console with one click. This feature is available as of AWS DeepLens software version 1.2.3. You can update your AWS DeepLens software by re-booting your device or by using the command sudo apt-get install awscam on the Ubuntu terminal. For this tutorial, you need the MXNet version 0.12. You can update the MXNet version by using the command sudo pip3 install mxnet==0.12.1. You can access the Ubuntu terminal via SSH or uHDMI.

To demonstrate the capability, we will walk you through building a model to classify common objects. This object classification model is based on Caltech-256 dataset and is trained using ResNet network. Through this walk through tutorial, you will build an object classifier that can identify 256 commonly found objects.

To build you own model, you first need to identify a dataset. You can bring your own dataset or use an existing one. In this tutorial, we show you how to build an object detection model in Amazon SageMaker using Caltech-256 image classification dataset.

In order to follow through with this tutorial, ensure that your DeepLens software version is updated to version 1.2.3 and above and MXNet version 0.12.

To build this model in Amazon SageMaker, Visit Amazon SageMaker console (

Create notebook instance. Provide the name for your notebook instance and select an instance type (for example ml.t2.medium). Choose to create a new role or use an existing role. Choose Create notebook instance.

Once your notebook instance is created, open the notebook instance you just created.

You will see the Jupyter notebook hosted in your instance.

Create a new notebook by choosing New and conda_mxnet_p36 kernel.

Let’s start by importing the necessary packages. Importing boto3 SDK for Python allows you to access Amazon services like S3. get_execution_role will allow Amazon SageMaker to assume the role created during instance creation and accesses resources on your behalf.

import boto3
from sagemaker import get_execution_role
role = get_execution_role()

Next we define a bucket which hosts the dataset that will be used. In this example, the dataset is Caltech- 256. Create a bucket in your S3. The name for your bucket must contain the prefix ‘deeplens’. In this example, the bucket is ‘deeplens-imageclassification’.


Next we define the containers. Containers are docker containers and the training job defined in this notebook will run in the container for your region.

containers = {'us-west-2': '',
              'us-east-1': '',
              'us-east-2': '',
              'eu-west-1': ''}
training_image = containers[boto3.Session().region_name]

Next let’s import the dataset and upload it to your S3 bucket. We will download the train and validation sets for Caltech-256 and upload it to the S3 bucket created earlier.

import os 
import urllib.request
import boto3

def download(url):
    filename = url.split("/")[-1]
    if not os.path.exists(filename):
        urllib.request.urlretrieve(url, filename)

def upload_to_s3(channel, file):
    s3 = boto3.resource('s3')
    data = open(file, "rb")
    key = channel + '/' + file
    s3.Bucket(bucket).put_object(Key=key, Body=data)
# caltech-256
upload_to_s3('train', 'caltech-256-60-train.rec')
upload_to_s3('validation', 'caltech-256-60-val.rec')

Next let’s define the network that we will use to train the dataset. For this tutorial, we will use ResNet network. ResNet is the default image classification model in Amazon SageMaker. In this step, you can customize the hyper parameters of the network to train your dataset.

num_layers lets you define the network depth. ResNet supports multiple network depths. For example: 18, 34, 50, 101, 152, 200 etc. For this example we choose the network depth as 50.

Next we need to specify the input image dimensions. The dataset that we used in this example has the dimensions 224 x 224 and has 3 color channels: RGB.

Next we specify the number of training samples in the training set. For Caltecg-256, the number of training samples are 15420.

Next, we specify the number of output classes for the model. In this example, the number of output classes for Caltech-256 is 257.

Batch size refers to the number of training examples utilized in one iteration. You can customize this number based on the computation resources available to you. Epoch is when the entire dataset is processed by the network once. Learning rate determines how fast the weights or coefficients of your network change. You can customize batch size, epochs and learning rates. You can refer to the definitions here:

# The algorithm supports multiple network depth (number of layers). They are 18, 34, 50, 101, 152 and 200
# For this training, we will use 18 layers
num_layers = "50" 
# we need to specify the input image shape for the training data
image_shape = "3,224,224"
# we also need to specify the number of training samples in the training set
# for caltech it is 15420
num_training_samples = "15420"
# specify the number of output classes
num_classes = "257"
# batch size for training
mini_batch_size =  "128"
# number of epochs
epochs = "300"
# learning rate
learning_rate = "0.1"
#optimizer ='Adam'
checkpoint_frequency = "300"

To train a model in Amazon SageMaker, you create a training job. The training job includes the following information:

  • The URL of the Amazon Simple Storage Service (Amazon S3) bucket where you’ve stored the training data.
  • The compute resources that you want Amazon SageMaker to use for model training. Compute resources are ML compute instances that are managed by Amazon SageMaker.
  • The URL of the S3 bucket where you want to store the output of the job.
  • The Amazon Elastic Container Registry path where the training code is stored.

In this sample, we pass the default image classifier (ResNet) built in Amazon SageMaker. The checkpoint_frequency determines the frequency by which model files are stored during training. Since we only need the final model file for deeplens, it is set equal to the number of epochs.

Please make a note of job_name_prefix, S3OutputPath, InstanceType, InstanceCount.

import time
import boto3
from time import gmtime, strftime

s3 = boto3.client('s3')
# create unique job name 
job_name_prefix = 'DEMO-imageclassification'
timestamp = time.strftime('-%Y-%m-%d-%H-%M-%S', time.gmtime())
job_name = job_name_prefix + timestamp
training_params = \
    # specify the training docker image
    "AlgorithmSpecification": {
        "TrainingImage": training_image,
        "TrainingInputMode": "File"
    "RoleArn": role,
    "OutputDataConfig": {
        "S3OutputPath": 's3://{}/{}/output'.format(bucket, job_name_prefix)
    "ResourceConfig": {
        "InstanceCount": 1,
        "InstanceType": "ml.p2.8xlarge",
        "VolumeSizeInGB": 50
    "TrainingJobName": job_name,
    "HyperParameters": {
        "image_shape": image_shape,
        "num_layers": str(num_layers),
        "num_training_samples": str(num_training_samples),
        "num_classes": str(num_classes),
        "mini_batch_size": str(mini_batch_size),
        "epochs": str(epochs),
        "learning_rate": str(learning_rate),
        "lr_scheduler_step": str(lr_scheduler_step),
        "lr_scheduler_factor": str(lr_scheduler_factor),
        "augmentation_type": str(augmentation_type),
        "checkpoint_frequency": str(checkpoint_frequency),
        "augmentation_type" : str(augmentation_type)
    "StoppingCondition": {
        "MaxRuntimeInSeconds": 360000
#Training data should be inside a subdirectory called "train"
#Validation data should be inside a subdirectory called "validation"
#The algorithm currently only supports fullyreplicated model (where data is copied onto each machine)
    "InputDataConfig": [
            "ChannelName": "train",
            "DataSource": {
                "S3DataSource": {
                    "S3DataType": "S3Prefix",
                    "S3Uri": 's3://{}/train/'.format(bucket),
                    "S3DataDistributionType": "FullyReplicated"
            "ContentType": "application/x-recordio",
            "CompressionType": "None"
            "ChannelName": "validation",
            "DataSource": {
                "S3DataSource": {
                    "S3DataType": "S3Prefix",
                    "S3Uri": 's3://{}/validation/'.format(bucket),
                    "S3DataDistributionType": "FullyReplicated"
            "ContentType": "application/x-recordio",
            "CompressionType": "None"
print('Training job name: {}'.format(job_name))
print('\nInput Data Location: {}'.format(training_params['InputDataConfig'][0]['DataSource']['S3DataSource']))

In the next step, you can check the status of the Job in CloudWatch.

# create the Amazon SageMaker training job
sagemaker = boto3.client(service_name='sagemaker')

# confirm that the training job has started
status = sagemaker.describe_training_job(TrainingJobName=job_name)['TrainingJobStatus']
print('Training job current status: {}'.format(status))

    # wait for the job to finish and report the ending status
    training_info = sagemaker.describe_training_job(TrainingJobName=job_name)
    status = training_info['TrainingJobStatus']
    print("Training job ended with status: " + status)
    print('Training failed to start')
     # if exception is raised, that means it has failed
    message = sagemaker.describe_training_job(TrainingJobName=job_name)['FailureReason']
    print('Training failed with the following error: {}'.format(message))

To check the status, go to SageMaker dashboard and choose Jobs. Select the Job you have defined and scroll down to the details page on Job to “monitor” section. You will see a link to logs which will open CloudWatch.

Once you run the notebook, it will create a model which can be directly imported into AWS DeepLens as a project. Once the training is complete, your model is ready to be imported in to AWS DeepLens.

Ensure that the mxnet version on your AWS DeepLens is 0.12. In case you need to upgrade, you can type the following code in your Ubuntu terminal.

sudo pip3 install mxnet==0.12.1

Now Log into AWS DeepLens Console (

Create new project

Choose – Create a new blank project

Name project – e.g. imageclassification

Select Add Model – this will open new page, “Import model to AWS Deeplens”

Select Amazon SageMaker trained model, in the Model setting, Amazon SageMaker training job ID drop down, select the imageclassification model you selected. In Model name choose model name e.g. imageclassification, keep description as image classification.

Go back to import model screen, select the imageclassification model you imported earlier, click Add model. Once model is added, you need to add a lambda function by choosing Add function.

To create a AWS DeepLens lambda function, you can follow the blog post: Dive deep into AWS DeepLens Lambda functions and the new model optimizer.

To provide an easy reference, we have provided the instructions for the lambda function for image classification below.

To create an inference Lambda function, use the AWS Lambda console and follow the steps below:

  1. Choose Create function. You customize this function to run inference for your deep learning models.
  2. Choose Blueprints
  3. Search for the greengrass-hello-world blueprint.
  4. Give your Lambda function the same name as your model e.g. imageclassification_lambda.
  5. Choose an existing IAM role: AWSDeepLensLambdaRole. You must have created this role as part of the registration process.
  6. Choose Create function.
  7. In Function code, make sure the handler is greengrassHelloWorld.function_handler.
  8. In the GreengrassHello file, remove all of the code. You will write the code for inference Lambda function in this file.

    Replace existing code with code below

    # Copyright Amazon AWS DeepLens, 2017
    import os
    import greengrasssdk
    from threading import Timer
    import time
    import awscam
    import cv2
    import mo
    from threading import Thread
    # Creating a greengrass core sdk client
    client = greengrasssdk.client('iot-data')
    # The information exchanged between IoT and clould has 
    # a topic and a message body.
    # This is the topic that this code uses to send messages to cloud
    iotTopic = '$aws/things/{}/infer'.format(os.environ['AWS_IOT_THING_NAME'])
    jpeg = None
    Write_To_FIFO = True
    class FIFO_Thread(Thread):
        def __init__(self):
            ''' Constructor. '''
        def run(self):
            fifo_path = "/tmp/results.mjpeg"
            if not os.path.exists(fifo_path):
            f = open(fifo_path,'w')
            client.publish(topic=iotTopic, payload="Opened Pipe")
            while Write_To_FIFO:
                except IOError as e:
    def greengrass_infinite_infer_run():
            input_width = 224
            input_height = 224
            model_name = "image-classification"
            error, model_path = mo.optimize(model_name,input_width,input_height, aux_inputs={'--epoch': 300})
            # The aux_inputs is equal to the number of epochs and in this case, it is 300
            # Load model to GPU (use {"GPU": 0} for CPU)
            mcfg = {"GPU": 1}
            model = awscam.Model(model_path, mcfg)
            client.publish(topic=iotTopic, payload="Model loaded")
            model_type = "classification"
            with open('caltech256_labels.txt', 'r') as f:
    	        labels = [l.rstrip() for l in f]
            topk = 5
            results_thread = FIFO_Thread()
            # Send a starting message to IoT console
            client.publish(topic=iotTopic, payload="Inference is starting")
            doInfer = True
            while doInfer:
                # Get a frame from the video stream
                ret, frame = awscam.getLastFrame()
                # Raise an exception if failing to get a frame
                if ret == False:
                    raise Exception("Failed to get frame from the stream")
                # Resize frame to fit model input requirement
                frameResize = cv2.resize(frame, (input_width, input_height))
                # Run model inference on the resized frame
                inferOutput = model.doInference(frameResize)
                # Output inference result to the fifo file so it can be viewed with mplayer
                parsed_results = model.parseResult(model_type, inferOutput)
                top_k = parsed_results[model_type][0:topk]
                msg = '{'
                prob_num = 0 
                for obj in top_k:
                    if prob_num == topk-1: 
                        msg += '"{}": {:.2f}'.format(labels[obj["label"]], obj["prob"]*100)
                        msg += '"{}": {:.2f},'.format(labels[obj["label"]], obj["prob"]*100)
                prob_num += 1
                msg += "}"  
                client.publish(topic=iotTopic, payload = msg)
    cv2.putText(frame, labels[top_k[0]["label"]], (0,22), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 165, 20), 4)
                global jpeg
                ret,jpeg = cv2.imencode('.jpg', frame)
        except Exception as e:
            msg = "myModel Lambda failed: " + str(e)
            client.publish(topic=iotTopic, payload=msg)
        # Asynchronously schedule this function to be run again in 15 seconds
        Timer(15, greengrass_infinite_infer_run).start()
    # Execute the function above
    # This is a dummy handler and will not be invoked
    # Instead the code above will be executed in an infinite loop for our example
    def function_handler(event, context):
  1. To add the text file to your lambda function: In the Function code block, choose File. Then choose New File, add following code, then save file as caltech256_labels.txt

    #Insert caltech256_labels.txt
    0 ak47
    1 american-flag
    2 backpack
    3 baseball-bat
    4 baseball-glove
    5 basketball-hoop
    6 bat
    7 bathtub
    8 bear
    9 beer-mug
    10 billiards
    11 binoculars
    12 birdbath
    13 blimp
    14 bonsai-101
    15 boom-box
    16 bowling-ball
    17 bowling-pin
    18 boxing-glove
    19 brain-101
    20 breadmaker
    21 buddha-101
    22 bulldozer
    23 butterfly
    24 cactus
    25 cake
    26 calculator
    27 camel
    28 cannon
    29 canoe
    30 car-tire
    31 cartman
    32 cd
    33 centipede
    34 cereal-box
    35 chandelier-101
    36 chess-board
    37 chimp
    38 chopsticks
    39 cockroach
    40 coffee-mug
    41 coffin
    42 coin
    43 comet
    44 computer-keyboard
    45 computer-monitor
    46 computer-mouse
    47 conch
    48 cormorant
    49 covered-wagon
    50 cowboy-hat
    51 crab-101
    52 desk-globe
    53 diamond-ring
    54 dice
    55 dog
    56 dolphin-101
    57 doorknob
    58 drinking-straw
    59 duck
    60 dumb-bell
    61 eiffel-tower
    62 electric-guitar-101
    63 elephant-101
    64 elk
    65 ewer-101
    66 eyeglasses
    67 fern
    68 fighter-jet
    69 fire-extinguisher
    70 fire-hydrant
    71 fire-truck
    72 fireworks
    73 flashlight
    74 floppy-disk
    75 football-helmet
    76 french-horn
    77 fried-egg
    78 frisbee
    79 frog
    80 frying-pan
    81 galaxy
    82 gas-pump
    83 giraffe
    84 goat
    85 golden-gate-bridge
    86 goldfish
    87 golf-ball
    88 goose
    89 gorilla
    90 grand-piano-101
    91 grapes
    92 grasshopper
    93 guitar-pick
    94 hamburger
    95 hammock
    96 harmonica
    97 harp
    98 harpsichord
    99 hawksbill-101
    100 head-phones
    101 helicopter-101
    102 hibiscus
    103 homer-simpson
    104 horse
    105 horseshoe-crab
    106 hot-air-balloon
    107 hot-dog
    108 hot-tub
    109 hourglass
    110 house-fly
    111 human-skeleton
    112 hummingbird
    113 ibis-101
    114 ice-cream-cone
    115 iguana
    116 ipod
    117 iris
    118 jesus-christ
    119 joy-stick
    120 kangaroo-101
    121 kayak
    122 ketch-101
    123 killer-whale
    124 knife
    125 ladder
    126 laptop-101
    127 lathe
    128 leopards-101
    129 license-plate
    130 lightbulb
    131 light-house
    132 lightning
    133 llama-101
    134 mailbox
    135 mandolin
    136 mars
    137 mattress
    138 megaphone
    139 menorah-101
    140 microscope
    141 microwave
    142 minaret
    143 minotaur
    144 motorbikes-101
    145 mountain-bike
    146 mushroom
    147 mussels
    148 necktie
    149 octopus
    150 ostrich
    151 owl
    152 palm-pilot
    153 palm-tree
    154 paperclip
    155 paper-shredder
    156 pci-card
    157 penguin
    158 people
    159 pez-dispenser
    160 photocopier
    161 picnic-table
    162 playing-card
    163 porcupine
    164 pram
    165 praying-mantis
    166 pyramid
    167 raccoon
    168 radio-telescope
    169 rainbow
    170 refrigerator
    171 revolver-101
    172 rifle
    173 rotary-phone
    174 roulette-wheel
    175 saddle
    176 saturn
    177 school-bus
    178 scorpion-101
    179 screwdriver
    180 segway
    181 self-propelled-lawn-mower
    182 sextant
    183 sheet-music
    184 skateboard
    185 skunk
    186 skyscraper
    187 smokestack
    188 snail
    189 snake
    190 sneaker
    191 snowmobile
    192 soccer-ball
    193 socks
    194 soda-can
    195 spaghetti
    196 speed-boat
    197 spider
    198 spoon
    199 stained-glass
    200 starfish-101
    201 steering-wheel
    202 stirrups
    203 sunflower-101
    204 superman
    205 sushi
    206 swan
    207 swiss-army-knife
    208 sword
    209 syringe
    210 tambourine
    211 teapot
    212 teddy-bear
    213 teepee
    214 telephone-box
    215 tennis-ball
    216 tennis-court
    217 tennis-racket
    218 theodolite
    219 toaster
    220 tomato
    221 tombstone
    222 top-hat
    223 touring-bike
    224 tower-pisa
    225 traffic-light
    226 treadmill
    227 triceratops
    228 tricycle
    229 trilobite-101
    230 tripod
    231 t-shirt
    232 tuning-fork
    233 tweezer
    234 umbrella-101
    235 unicorn
    236 vcr
    237 video-projector
    238 washing-machine
    239 watch-101
    240 waterfall
    241 watermelon
    242 welding-mask
    243 wheelbarrow
    244 windmill
    245 wine-bottle
    246 xylophone
    247 yarmulke
    248 yo-yo
    249 zebra
    250 airplanes-101
    251 car-side-101
    252 faces-easy-101
    253 greyhound
    254 tennis-shoes
    255 toad
    256 clutter

  1. Save the lambda function

    Now deploy the lambda function by selecting Actions dropdown button. And then select Publish new version
  1. This will pop up new box. You can keep version description blank, and choose Publish. This will publish the lambda function.
  1. Once done, add the lambda function to the project and choose Create new project to finish the project creation.
    You will see your project created in the Projects list.


Once the project is created, select the project and choose Deploy to device. Choose your target AWS DeepLens device. Choose Review.

Now you are ready to deploy your own object detection model. Choose Deploy.

Congratulations! You have built your own object classification model based on a dataset and deployed it to AWS DeepLens for inference.

You can also refer to the notebook for training the dataset in Amazon SageMaker on Github.

About the Authors

Mahendra Bairagi is a Senior Solutions Architect Specialists in IoT, helping customers build intelligence everywhere. He has extensive experience on AWS IoT and AI specific services, along with expertise in AWS analytics, mobile and serverless services. Prior to joining Amazon Web Services, He had long tenure as entrepreneur, IT leader, enterprise architect and software developer at AWS partner organization and fortune 500 corporations. In spare time he coaches junior Olympics archery development team, build bots, RC planes and help animal shelters.



Jyothi Nookula is a Senior Product Manager for AWS DeepLens. She loves to build products that delight her customers. In her spare time, she loves to paint and host charity fund raisers for her art exhibitions.