The company wants to predict the demand for a new product that will soon be launched, based on the sales history of similar products. This is a time series forecasting problem, which requires a machine learning algorithm that can learn from historical data and generate future predictions.

One of the most suitable solutions for this problem is to use the Amazon SageMaker DeepAR algorithm, which is a supervised learning algorithm for forecasting scalar time series using recurrent neural networks (RNN). DeepAR can handle multiple related time series, such as the sales of different products, and learn a global model that captures the common patterns and trends across the time series. DeepAR can also generate probabilistic forecasts that provide confidence intervals and quantify the uncertainty of the predictions.

DeepAR can outperform traditional forecasting methods, such as ARIMA, especially when the dataset contains hundreds or thousands of related time series. DeepAR can also use the trained model to forecast the demand for new products that are similar to the ones it has been trained on, by using the categorical features that encode the product attributes. For example, the company can use the product type, brand, flavor, size, and price as categorical features to group the products and learn the typical behavior for each group.

Therefore, the Machine Learning Specialist should apply the Amazon SageMaker DeepAR algorithm to forecast the demand for the new product, by using the sales history of the existing products as the training dataset, and the product attributes as the categorical features.

DeepAR Forecasting Algorithm - Amazon SageMaker

Now available in Amazon SageMaker: DeepAR algorithm for more accurate time series forecasting

 Early stopping is a technique that can be used to prevent overfitting and improve model performance on the validation set. Overfitting occurs when the model learns the training data too well and fails to generalize to new and unseen data. This can be seen in the graph, where the training loss keeps decreasing, but the validation loss starts to increase after some point. This means that the model is fitting the noise and patterns in the training data that are not relevant for the validation data. Early stopping is a way of stopping the training process before the model overfits the training data. It works by monitoring the validation loss and stopping the training when the validation loss stops decreasing or starts increasing. This way, the model is saved at the point where it has the best performance on the validation set. Early stopping can also save time and resources by reducing the number of epochs needed for training. References:

Early Stopping

How to Stop Training Deep Neural Networks At the Right Time Using Early Stopping

Amazon SageMaker script mode is a feature that allows users to use training scripts similar to those they would use outside SageMaker with SageMaker’s prebuilt containers for various frameworks such as TensorFlow. Script mode supports reading data from Amazon S3 buckets without requiring any changes to the training script. Therefore, option B is the best method of providing training data to Amazon SageMaker that would meet the business requirements with the least development overhead.

Option A is incorrect because using a local path of the data would not be scalable or reliable, as it would depend on the availability and capacity of the local storage. Moreover, using a local path of the data would not leverage the benefits of Amazon S3, such as durability, security, and performance. Option C is incorrect because rewriting the train.py script to convert TFRecords to protobuf would require additional development effort and complexity, as well as introduce potential errors and inconsistencies in the data format. Option D is incorrect because preparing the data in the format accepted by Amazon SageMaker would also require additional development effort and complexity, as well as involve using additional services such as AWS Glue or AWS Lambda, which would increase the cost and maintenance of the solution.

Bring your own model with Amazon SageMaker script mode

GitHub - aws-samples/amazon-sagemaker-script-mode

Deep Dive on TensorFlow training with Amazon SageMaker and Amazon S3

amazon-sagemaker-script-mode/generate_cifar10_tfrecords.py at master

SageMaker’s TrainingJobEarlyStoppingType set to AUTO stops training jobs early when they no longer improve the objective metric, significantly reducing tuning job durations.

From AWS documentation:

“You can use early stopping to terminate training jobs that are unlikely to result in a better objective metric than previously completed jobs. This can reduce tuning costs and duration.”

— AWS SageMaker Automatic Model Tuning documentation

The combination of steps that the ML engineer should take to train the model are to create a .lst file that contains a list of image files and corresponding class labels, upload the .lst file to Amazon S3, and initiate transfer learning by training the model using the images of less common species. This approach will allow the ML engineer to leverage the existing ImageNetV2 CNN model and fine-tune it with the new data using Pipe mode in SageMaker.

A .lst file is a text file that contains a list of image files and corresponding class labels, separated by tabs. The .lst file format is required for using the SageMaker image classification algorithm with Pipe mode. Pipe mode is a feature of SageMaker that enables streaming data directly from Amazon S3 to the training instances, without downloading the data first. Pipe mode can reduce the startup time, improve the I/O throughput, and enable training on large datasets that exceed the disk size limit. To use Pipe mode, the ML engineer needs to upload the .lst file to Amazon S3 and specify the S3 path as the input data channel for the training job1.

Transfer learning is a technique that enables reusing a pre-trained model for a new task by fine-tuning the model parameters with new data. Transfer learning can save time and computational resources, as well as improve the performance of the model, especially when the new task is similar to the original task. The SageMaker image classification algorithm supports transfer learning by allowing the ML engineer to specify the number of output classes and the number of layers to be retrained. The ML engineer can use the existing ImageNetV2 CNN model, which is trained on 1,000 classes of common objects, and fine-tune it with the new data of less common animal species, which is a similar task2.

The other options are either less effective or not supported by the SageMaker image classification algorithm. Using a ResNet model and initiating full training mode would require training the model from scratch, which would take more time and resources than transfer learning. Using an Inception model is not possible, as the SageMaker image classification algorithm only supports ResNet and ImageNetV2 models. Using an augmented manifest file in JSON Lines format is not compatible with Pipe mode, as Pipe mode only supports .lst files for image classification1.

1: Using Pipe input mode for Amazon SageMaker algorithms | AWS Machine Learning Blog

2: Image Classification Algorithm - Amazon SageMaker

Serverless inference in SageMaker is designed for workloads with intermittent traffic and unpredictable usage patterns, which aligns with the media company ' s periodic high-traffic windows. Because the payload is less than 4 MB (serverless inference supports payloads up to 4 MB), serverless inference is appropriate and provisioned concurrency ensures the endpoints are warm and ready during peak times to minimize latency.

From AWS documentation:

“Amazon SageMaker Serverless Inference is ideal for applications with intermittent or unpredictable traffic. You can optionally enable provisioned concurrency to ensure that your endpoints are always ready to process requests during anticipated peak traffic hours.”

— AWS SageMaker Serverless Inference documentation

Using real-time inference with auto scaling (A) incurs a higher cost as it requires always-on instances. Batch transform (D) is for offline, large-scale inferences, not low-latency real-time inference. Asynchronous inference (C) is typically used for large payloads ( > 4 MB) or long processing times.

According to the Amazon Kinesis Data Streams documentation, the maximum size of data blob (the data payload before Base64-encoding) per record is 1 MB. The maximum number of records that can be sent to a shard per second is 1,000. Therefore, the maximum throughput of a shard is 1 MB/sec for input and 2 MB/sec for output. In this case, the input throughput is 100 transactions per second * 100 KB per transaction = 10 MB/sec. Therefore, the minimum number of shards required is 10 MB/sec / 1 MB/sec = 10 shards. However, the question asks for the minimum number of shards in Kinesis Data Streams, not the minimum number of shards per stream. A Kinesis Data Streams account can have multiple streams, each with its own number of shards. Therefore, the minimum number of shards in Kinesis Data Streams is 1, which is the minimum number of shards per stream. References:

Amazon Kinesis Data Streams Terminology and Concepts

Amazon Kinesis Data Streams Limits

 To improve the test error for the image classifier, the Machine Learning Specialist should use the technique of increasing the training data by adding variation in rotation for training images. This technique is called data augmentation, which is a way of artificially expanding the size and diversity of the training dataset by applying various transformations to the original images, such as rotation, flipping, cropping, scaling, etc. Data augmentation can help the model learn more robust features that are invariant to the orientation, position, and size of the objects in the images. This can improve the generalization ability of the model and reduce the test error, especially for cases where the images are not well-aligned or have different perspectives1.

1: Image Augmentation - Amazon SageMaker

The Data Scientist is building a model to predict customer churn using a dataset of 100 continuous numerical features. The Marketing team wants to interpret the model and see the direct impact of relevant features on the model outcome. However, the Data Scientist observes that there is a wide gap between the training and validation set accuracy, which indicates that the model is overfitting the data and generalizing poorly to new data.

To improve the model performance and satisfy the Marketing team’s needs, the Data Scientist can use the following methods:

Add L1 regularization to the classifier: L1 regularization is a technique that adds a penalty term to the loss function of the logistic regression model, proportional to the sum of the absolute values of the coefficients. L1 regularization can help reduce overfitting by shrinking the coefficients of the less important features to zero, effectively performing feature selection. This can simplify the model and make it more interpretable, as well as improve the validation accuracy.

Perform recursive feature elimination: Recursive feature elimination (RFE) is a feature selection technique that involves training a model on a subset of the features, and then iteratively removing the least important features one by one until the desired number of features is reached. The idea behind RFE is to determine the contribution of each feature to the model by measuring how well the model performs when that feature is removed. The features that are most important to the model will have the greatest impact on performance when they are removed. RFE can help improve the model performance by eliminating the irrelevant or redundant features that may cause noise or multicollinearity in the data. RFE can also help the Marketing team understand the direct impact of the relevant features on the model outcome, as the remaining features will have the highest weights in the model.

Regularization for Logistic Regression

Recursive Feature Elimination