Amazon SageMaker – Bring your own Algorithm

In previous posts, we explored Amazon SageMaker’s AutoPilot, which was terrific, and we learned how to use your own algorithm with Docker, which was lovely but a bit of a fuzz. In this post, we will see how easy it is to bring your own algorithm using the script mode in SageMaker. Again we will use the new SageMaker Studio!

Let’s take a look at the resources we will use how much it will cost this tutorial:

  • AWS Level: 300
  • Total Cost: Let’s say 1$ (with no Endpoint deployed)
    • Amazon SageMaker Notebook: ml.system for 24 hours -> 0.00364 x 24 -> 0.087$ + ml.t3.medium for 12 hours -> 0.7$ = 0.8$
    • Amazon SageMaker Training instance:  225 seconds in ml.c5.2xlarge -> 0.476 x 0.07 = 0.03

You can find the code of this post here:


The only thing we need is a dataset. I chose the popular Fashion MNIST dataset.
We will download Fashion MNIST from Keras datasets as follows:

Let’s take a look at the data we just download:

Sample of Fashion-MNIST dataset

Pre-process dataset

We have to make some adjustments to our data for our model.

First, we normalize the images to 0-1. Fashion MNIST images are 28×28 arrays with values from 0 to 255, with 0 being color black and 255 being color white. We divide the array with 255 to bring its value to [0, 1] (line 4-5).

We also need to transform our data from 28×28 to 28x28x1 for our model (line 8-9)

Finally, we one-hot encode our labels (line 13-14)

Building the model

We will use Convolutional Neural Networks (CNNs) to solve that classification problem. Construct a relatively simple CNN as follows:

For this model, we used a pretty simple CNN network with three Conv2D layers, one 64 and two 32, both with kernel size 3×3 and activation function, our favorite ReLu. Below is a summary of the model.

Layer (type)                 Output Shape              Param #   
conv2d_28 (Conv2D)           (None, 26, 26, 64)        640       
conv2d_29 (Conv2D)           (None, 24, 24, 32)        18464     
conv2d_30 (Conv2D)           (None, 22, 22, 32)        9248      
flatten_12 (Flatten)         (None, 15488)             0         
dense_12 (Dense)             (None, 10)                154890    
Total params: 183,242
Trainable params: 183,242
Non-trainable params: 0

Setting up SageMaker

In the same Jupyter Notebook, we upload our data to S3 and “let SageMaker” train our model. First of all, we need to transform our data in a binary format; we use NumPy. After the transformation, we upload the files into SageMaker’s S3 bucket.

Train the model

To train the model, first, we need to bring the model code in an acceptable format by SageMaker. To do that we need to have in mind two things:

  1. SageMaker will get a python file as input, and it will run it with arguments
  2. The training of the model will be performed in another instance, so we need to get the data from S3


Given the above, we first need to accept the arguments that SageMaker uses in our script, and secondly, we need to read the data from S3. We create a new file named (you can use whatever name you want)

  • Line 26-40: Set up the argparse to accept specific arguments. Set learning rate, epochs, and much more through the command line, and the command line translates into arguments in SageMaker Estimator
  • Line 42-45: Parsing the arguments to variables
  • Line 47-49: reading the data from local. Wait… local? Yes, SageMaker copies the data that you have set in the parameters in a folder specified by you; in our case, args.train will hold our files fo training.
    • Line 12-21: reading the files with NumPy
  • Line 51-73: This is the model defined above
  • Line 75-77: Evaluating model
  • Line 79-80: Saving the model in args.model_dir

SageMaker Estimator

After setting the file, we need to set the Estimator for the script mode. We will use the built-in TensorFlow estimator from SageMaker to use the script mode.

  • Line 1: Is the directory to save the final model
  • Line 2: is the instance where we will train our model. Here you can find a list with all available instances in SageMaker, but for the GPU ones, you will have to increase your limit through AWS Support.
  • Line 3: are the hyperparameters/arguments for the script
  • Line 5-16: This is the Estimator that will spawn the instance to train the model
    • Line 6: This is our script we created earlier
    • Line 7: The type of instance to train the model
    • Line 8: The number of instances to train the model
    • Line 9: The folder to save our model
    • Line 11: The role of SageMaker needed to create everything. With this, you can run SageMaker from local also.
    • Line 12: The name of the training job that we will search in the Experiments section
    • Line 13: The version of Tensorflow. As for today, SageMaker supports up to 1.15 (I would love to see 2.0 with the new TripletSemiHardLoss add on)
    • Line 14: The python version
    • Line 15: Enable script mode
  • Line 18: Train the model with the generated inputs. Something we declared above, it is the train and the test location in S3

You can monitor the progress of your training job either in Jupyter Notebook or at the Experiment section. Remember when we used the name tf-fashion-mnist in the Estimator? Well, this is how we find the training job. Double-click “Unassigned trial components,” then find the name of the job, left-click and select “Open in trial details.” Select the section Metrics to see the evaluation score.


Experiment Section


Metrics of the training job

Furthermore, remember Line 76 on, where we printed the evaluation score of the model?

Test MSE : [0.2660780865192413, 0.9116]
  • Loss: 0.26
  • Accuracy: 91.1%

Well, that is pretty good, we can certainly do better, but for now, we are set. We can see that the monitoring of our model is a bit tricky; in future posts, I will take a closer look at the new debugger for SageMaker studio.

You can find the model in S3 and deploy it manually or by using the deploy endpoint command from SageMaker.


The script mode in SageMaker is definitely making the whole process of “Bring our own model” easier. You have everything set up out of the box, the only thing you have to do is to pass the arguments of the Estimator. In addition, SageMaker Studio makes the monitoring of the training job a bit easier, and the metrics are a very nice addition. On the other hand, there is limited support for the framework versions, for example, I would like to see Tensorflow 2.0 be available in the script mode. Furthermore, it was hard to set up the Local mode to train my model. The whole initialization process usually takes 2-4 minutes in order to start training your model, and I would like to test my model or run it locally first.

Well, that is it for today folks. I hope you find that article interesting. If you have any questions, suggestions or notices please let me know in the comment section below or at my Twitter account @siaterliskonsta. Make sure to check out my earlier posts about SageMaker AutoPilot and Serverless Data Lake. Until next time, take care!

6 thoughts on “Amazon SageMaker – Bring your own Algorithm”

  1. Hi, Great tutorial, I followed it but I am getting following error when I call
    …..sourcedir.tar.gz: An error occurred (AccessDenied) when calling the PutObject operation: Access Denied
    Can you please help me understand what could be the issue?

    1. Hello Karan, thank you for your kind words!

      Yes, it seems that SageMaker does not have access to the bucket. One thing that you can do is to print the role SageMaker uses, go to IAM, select that role and add a policy to access your bucket.

      To get SageMaker’s role:

      from sagemaker import get_execution_role

      role = get_execution_role()

      The policy you can use

      "Version": "2012-10-17",
      "Statement": [
      "Effect": "Allow",
      "Action": [
      "Resource": [

      If you are using a custom bucket to save the model, replace the with the name of your custom bucket.

      Let me know if that works for you.

  2. Hello, Thank you so much for detailed explanation.

    I am facing issue while deploying my model. I have my custom bucket and have passed output_path for artifacts while training model.

    But while deploying i am getting Access denied while creating bucket error.

    error message: Access Denied when calling the CreateBucket operation.

    Why it is trying to create bucket even if my estimator has role and output_path specified?

    It will be great if you able to help me to resolve my issue?

    Thank you so much!!!

  3. Hello! Sorry for the late reply!

    I have found this:

    output_path (str) – S3 location for saving the training result (model artifacts and output files). If not specified, results are stored to a default bucket. If the bucket with the specific name does not exist, the estimator creates the bucket during the fit() method execution. file:// urls are used for local mode. For example: ‘file://model/’ will save to the model folder in the current directory.

    It says that if the bucket does not exists then it tries to create the bucket instead of throwing an error.

    You could check to see if the bucket you have provided exists, and there are no spelling mistakes.

    Best, K

  4. Hello,
    Thank you for the tuto it help a lot !
    I had to modify the script (parser train, test, model to my path), and it doesn t work even if it is not an error.
    You have any idea, it would be really helpfull !

    2020-11-28 07:59:28 Uploading – Uploading generated training model
    2020-11-28 07:59:28 Failed – Training job failed

    I got this message:
    WARNING:tensorflow:Unresolved object in checkpoint: (root).optimizer.beta_1
    WARNING:tensorflow:Unresolved object in checkpoint: (root).optimizer.beta_2
    WARNING:tensorflow:Unresolved object in checkpoint: (root).optimizer.decay

    INFO:tensorflow:No assets to save.
    INFO:tensorflow:No assets to write.
    INFO:tensorflow:Signatures INCLUDED in export for Classify: None

    INFO:tensorflow:Signatures INCLUDED in export for Eval: None
    INFO:tensorflow:No assets to save.
    INFO:tensorflow:No assets to write.
    INFO:tensorflow:SavedModel written to: ./models/saved_model.pb

    have a good day

Leave a Comment

Your email address will not be published. Required fields are marked *