Sagemaker

Question 1

Training options

Answer 1

• Built-in training algorithms
• Spark MLLib
• Custom Python Tensorflow/MXNet code
• PyTorch, Scikit-Learn, RLEstimator
• XGBoost, Hugging Face, Chainer
• Your own Docker image
• Algorithm purchased from AWS marketplace

Question 2

Model deployment Options

Answer 2

• Persistent endpoint for making predictions on demand
• Sagemaker Batch Transform
• Inference Pipeline for more complex processing
• Sagemaker Neo for edge devices
• Elastic Inference for accelerating deep learning models
• Automatic Scaling(Horizontal scaling of no. end points)
• Shadow Testing

Question 3

Word2Vec models

Answer 3

Parameter update models that learn word embeddings vectors by training models on word predictions

Continuous Bag of Words(CBOW) - predict target word given surrounding context words, trying to maximise posterior log prob of predicting the right word, in the process updated weights are used as embedding vectors

Skip-gram - opposite of cbow, predict context words given target words

More details here
https://arxiv.org/abs/1411.2738

Question 4

Blazing text

Answer 4

AWS implementation of Word2Vec & Text Classification algorithms. Optimized for Multi-core CPU

Not parallelizable

Modes
Single CPU: skip-gram, cbow, batch-skip-gram | supervised
Single GPU: skip-gram, cbow | supervised
Multi CPU: batch skip-gram

Input

Unsupervised:
text file with one training sentence per line, space-separated tokens

Supervised:
One sentence per line
The first “word” in the sentence is the string __label__<label></label>

Example:

__label__4 linux ready for prime time

__label__2 the cat jump over the lazy fox

Hyperparams:
Word2Vec:

• Mode(skip-gram, batch-skip-gram, cbow)
• Learning-rate
• Window-size
• Vector-dim
• Negative-samples

Text classification:
- Epochs
- Learning_rate
- Word_ngrams
- Vector_dim