(v2.2.3) How to use training loop with JSON data?

[abs-xyz]

This is regarding spaCy v2 (specifically 2.2.3). I need to change some hyperparameters.
I have seen the examples here. They use nlp.update() to set certain things. But the training data is in the form

("I like London and Berlin.", {"entities": [(7, 13, "LOC"), (18, 24, "LOC")]})

I have JSON format data (converted from CoNLL format using spacy convert). How can I set values of hyperparameters?

I saw this discussion. This shows how to change the dropout rate and batch sizes, but I can't find how to change the learning rate, etc.

(My main question is on how to change hyperparameters. One of the ways in some discussion was to use the training loop, but I am not sure on how to run it with the data format that I have.)

Edit:
If I do something like learn_rate = spacy.util.env_opt("learn_rate", 0.0001), will this learning rate be used while training with spacy train? (How do I confirm this? os.environ["SPACY_LEARN_RATE"] gives a KeyError.)
Also, doing something like !learn_rate=0.0001 python -m spacy train ... doesn't work (in a notebook).

[rmitsch]

The v2 documentation on training your models describes changing your hyperparameters. I'd recommend switching to v3 however, since configuring this is more convenient there.
Also, can you elaborate how the format of your training data relates to your question about hyperparameters?

[abs-xyz]

If I am not mistaken, there is no flag to change the learn_rate.
Does using set learn_rate=0.0001 on CMD change the learn rate? Is there some way to "confirm" this?

[abs-xyz]

Also, is the correct way os.environ["LEARN_RATE"]="0.0001" or os.environ["SPACY_LEARN_RATE"]="0.0001"? (I think the latter is, from here)

[abs-xyz]

Also, can you elaborate how the format of your training data relates to your question about hyperparameters?

Here it is told to use the nlp.update() function.
From the v2 docs,

for raw_text, entity_offsets in train_data:
    doc = nlp.make_doc(raw_text)
    gold = GoldParse(doc, entities=entity_offsets)
    nlp.update([doc], [gold], drop=0.5, sgd=optimizer)

My question: How do I use the JSON data in this loop (for train_data)? It seems to be of the format that I mentioned in my original post.

[rmitsch]

Before we get deeper into this, are you starting from scratch or do you already have a substantial code base in v2?