AWS Machine Learning Certification – Quiz #1

KNN can be used to impute missing values, but for numerical data.
There’s not much data, so dropping rows with missing values may have a detrimental effect.
Dropping the feature is not an option as we don’t know whether it’s importance to the prediction at this point.
There’s not much data – Deep learning may be useful to impute missing values but mean looks to be the most sensible answer here. Also watch out for Multiple Imputation by Chained Equations in these questions and is a strong choice for both categorical and numerical feature imputation.

https://datascienceplus.com/imputing-missing-data-with-r-mice-package/

feature_dim must be set to the number of input data features
k must be set to the nearest neighbours (integer)
predictor_type must be set to classifier or regressor
sample_size ust be set to the number of data points in the training data set
dimension_reduction_target must be set to a postive integer less than feature_dim

https://docs.aws.amazon.com/sagemaker/latest/dg/kNN_hyperparameters.html

Sklearn’s TfidfVectorizer uses non default idf(t) for both the default smoothed and non default idf. L2 normalisation is also used. The code example below shows the default, which if you know idf is typically ln(number of docs/docs with this term), you’d be expecting some zeros.

There are 2 documents (sentences) and 8 possible bigrams (listed below) with ‘a’ excluded because it’s shorter than 2 character minimum. So the resulting tf-idf is 2,8

https://scikit-learn.org/stable/modules/feature_extraction.html#text-feature-extraction

Example Code –

from sklearn.feature_extraction.text import TfidfVectorizer
corpus = [“product managers know about machine learning”,
“machine learning, a product owner essential”]
vectorizer_A = TfidfVectorizer(ngram_range = (2,2))
vectorizer_B = TfidfVectorizer(ngram_range = (2,2),smooth_idf=False, norm=None)
matrix = vectorizer_A.fit_transform(corpus)
print(“**list of bigrams**”)
for i, feature in enumerate(vectorizer_A.get_feature_names()):
print(i, feature)
matrix_A = vectorizer_A.fit_transform(corpus)
matrix_B = vectorizer_B.fit_transform(corpus)
print(“**bigram tf-idf with defaults values**” )
print(matrix_A)
print(“**bigram tf-idf with no smoothing or normalisation**” )
print(matrix_B)

**list of bigrams**0 about machine1 know about2 learning product3 machine learning4 managers know5 owner essential6 product managers7 product owner**bigram tf-idf with defaults values**  (0, 3)	0.33517574332792605  (0, 0)	0.47107781233161794  (0, 1)	0.47107781233161794  (0, 4)	0.47107781233161794  (0, 6)	0.47107781233161794  (1, 5)	0.534046329052269  (1, 7)	0.534046329052269  (1, 2)	0.534046329052269  (1, 3)	0.37997836159100784**bigram tf-idf with no smoothing or normalisation**  (0, 3)	1.0  (0, 0)	1.6931471805599454  (0, 1)	1.6931471805599454  (0, 4)	1.6931471805599454  (0, 6)	1.6931471805599454  (1, 5)	1.6931471805599454  (1, 7)	1.6931471805599454  (1, 2)	1.6931471805599454  (1, 3)	1.0

context_length is required, the number of time-points the model sees before making the prediction and should be about the same as prediction_length, which is also required. Epochs, the number of passes over the training data is required. You must also specify the time_frequency granularity.

https://docs.aws.amazon.com/sagemaker/latest/dg/deepar_hyperparameters.html

ROC/AUC are used for classifiers and not regression.
ROC curves to plot TPR vs FPR for different decision thresholds.
An AUC score of 1 indicates a perfect classifier, 0.8 or above good to excellent but 0.5 is only as good as a random guess.
If you have class imbalance, Precision Recall Curves are more helpful.
https://machinelearningmastery.com/roc-curves-and-precision-recall-curves-for-classification-in-python/

API Gateway is the best solution as all of the ETL and implementation specifics are behind the API Gateway and hidden from the client. The API gateway also makes it easier to manage and control access.

All are examples of transfer learning.

“Image classification in Amazon SageMaker can be run in two modes: full training and transfer learning. “https://docs.aws.amazon.com/sagemaker/latest/dg/IC-HowItWorks.html

“You also have a choice of backbones for the FCN, PSP, and DeepLabV3 algorithms: ResNet50 or ResNet101. These backbones include pretrained artifacts that were originally trained on the ImageNet classification task. You can fine-tune these backbones for segmentation using your own data. ” https://docs.aws.amazon.com/sagemaker/latest/dg/semantic-segmentation.html

“Object detection in Amazon SageMaker supports both VGG-16 and ResNet-50 as a base network for SSD. The algorithm can be trained in full training mode or in transfer learning mode. “https://docs.aws.amazon.com/sagemaker/latest/dg/algo-object-detection-tech-notes.html

“Amazon Rekognition Custom Labels builds off of Amazon Rekognition’s existing capabilities, which are already trained on tens of millions of images across many categories. Instead of thousands of images, you can upload a small set of training images (typically a few hundred images or less) that are specific to your use case. ” https://docs.aws.amazon.com/rekognition/latest/customlabels-dg/what-is.html

a) incorrect.  BlazingText does use an enhanced version of fast text, but for classification not word embeddings.
b) correct.  A distributed version of word2vec can be utilised, set evaluation=True to create a model artifact file (eval.json) and use vector.txt and TSNE to visualise.
c) incorrect. This is an unsupervised algorithm and the  validation channels is not supported. There is no such instance variable as model.spearman
d) Evalation = True is required to create the eval.json model attribute. The vector.txt file is created irrespective of hyper parameter settings. JSON lines manifest files are not supported by Blazing Text’s word2vec implementation.

For more information, see blazingtext_word2vec_text8.ipynb example notebook,  provided by AWS in your SageMaker Notebook instance (/home/ec2-user/sample-notebooks/introduction_to_amazon_algorithms)

a) correct. For training, raw text must be converted to space separated tokenised text with labels identified with __mylabel__. For pipe mode, there is no need to use RecordIO because JSON Lines manifest can be used.  Both train and validation channels are supported.
b) incorrect. The algorithm supports multi-label classification. Multiple label predictions can be returned by setting ‘k’ in the JSON request to the deployed endpoint.
c) incorrect. “It extends the FastText text classifier to leverage GPU acceleration using custom CUDA kernels. The model can be trained on more than a billion words in a couple of minutes using a multi-core CPU or a GPU”
d) incorrect. It does extend the FastText classifier, but it is supervised.
e) Accuracy on the validation data is used as a proxy to the quality of the algorithm . “The BlazingText Text Classification algorithm (supervised mode), also reports on a single metric during training: the validation:accuracy. When tuning the hyperparameter values for the text classification algorithm, use these metrics as the objective.”

https://docs.aws.amazon.com/sagemaker/latest/dg/blazingtext.html