infomax recommendations poor when using Universal Sentence Encoder 4

[mammykins]

In a previous PR we adjusted the code in helper_embedding.py

Note the commented out larger transformer model.

# DAN model, lighter A stands for averaging; download and unzip
# https://tfhub.dev/google/universal-sentence-encoder/4
model = hub.Module(os.path.join(os.getenv('DIR_DATA_EXTERNAL'), 'universal-sentence-encoder_4'))
# Transformer model, more performant, runs on GPU, if available
# model = hub.load('data/external/universal-sentence-encoder-large_5')

Both @avisionh and I noticed the recommendations from the downstream model in 04_annoy_recommend_content.py are a bit rubbish!

This contrasts to those we saw produced by @whojammyflip when using his GPU and the larger model.

The main difference is the universal sentence encoder model version and size.

At a minimum we should document this in the script or change the default behaviour and recommend use on a GPU (on the cloud). This would require additional work.