Sampler configuration

There are many queries to ask about in triplet embedding tasks. Most of these queries aren’t useful; chances are most queries will have obvious answers and won’t improve the embedding much.

Choosing the most useful queries to improve the embedding is the task of “active machine learning” aka “adaptive sampling algorithms.” These algorithms use all previous responses collected to determine the next query that will help improve the embedding the most.

If quality embeddings are desired, the benchmarks at “Active sampling” are likely of interest, as are the FAQs on “When should I use random/active sampling?” and “What active samplers are recommended?” However, quality embeddings may not always be of interest: sometimes, the goal is not to generate a good embedding, but rather to see how well the crowd agrees with each other, or to make sure participants don’t influence each other (and each response is independent of other responses). Here’s a rule of thumb:


Want an unbiased estimate of what the crowd thinks? Don’t specify samplers, or use the default Random: {}, which will rely on Random

Want to generate a better embedding? Worried about the cost of collecting responses? Use ARR: {}, which will rely on ARR.

Want to measure how well the crowd agrees with one another? Use Validation: {}, which will rely on Validation,

Now, let’s go over how to configure different samplers and an example.

File structure

Part of a init.yaml configuration file might look like this:

# file: init.yaml
  ARR: {random_state: 42}
  Random: {}
  probs: {"ARR": 85, "Random": 15}
  samplers_per_user: 0  # use probability above

This will create a versions of ARR with random_state=42 would be created alongside the default version of Random. When a query is generated, it will be generated from ARR 85% of the time and from Random the rest of the time. Generally, the keys in samplers and sampling follow these general rules:

  • samplers: controls how one specific sampler behaves (i.e., sampler initialization).

    • The type of sampler is specified by each key (e.g., key ARR corresponds to ARR), and any arguments are initialization parameters for that object.

  • sampling: controls samplers interact. For example:

    • the probs key controls how frequently each sampler in samplers is used

    • the common key controls sending initialization arguments to every sampler.

    • the samplers_per_user key controls how many samplers are seen by any one user who visits Salmon.

A good default configuration is mentioned in the FAQ “What active samplers are recommended?” The defaults and complete details are in Config.

Multiple samplers of the same name

If two samplers with different purposes want to be used, the key class is used to specify which type of sampler should be used:

# file: init.yaml
    class: Random
    class: Random

This will generate queries from these two samplers with equal probability:

from salmon.triplets.samplers import Random

Initialization arguments

Arguments inside each key are passed to the sampler. For example,

# file: init.yaml
    random_state: 42
    module: CKL
    d: 3
    scorer: "uncertainty"

would create an instance of ARR:

from salmon.triplets.samplers import ARR
ARR(random_state=42, module="CKL", d=3, scorer="uncertainty")

Note that the argument are documented in ARR. Some argument are arguments that ARR directly uses (like module), and other are passed to Adaptive as mentioned in the docstring of ARR.

If you have multiple arguments for every sampler, you can specify that with the common key:

# file: init.yaml
    module: TSTE
    module: CKL
    d: 3
    random_state: 42

This would initialize these samplers:

from salmon.triplets.samplers import ARR
ARR(module="TSTE", d=3, random_state=42)
ARR(module="CKL",  d=3, random_state=42)

The documentation for ARR is available at ARR.


Let’s start out with a simple init.yaml file, one suited for random sampling.

targets: ["obj1", "obj2", "foo", "bar", "foobar!"]
  Random: {}
  Validation: {"n_queries": 10}

By default, samplers defaults to Random: {}. We have to customize the samplers key use adaptive sampling algorithms:

targets: ["obj1", "obj2", "foo", "bar", "foobar!"]
  ARR: {}
  Random: {}
  Validation: {"n_queries": 10}
  probs: {"ARR": 70, "Random": 20, "Validation": 10}

When ARR is specified as a key for samplers, salmon.triplets.samplers.ARR is used for the sampling method. Customization is possible by passing different keyword arguments to ARR. For example, this could be a configuration:

targets: ["obj1", "obj2", "foo", "bar", "foobar!"]
  Random: {}
    module: "TSTE"

Validation sampler

Note: generating validation queries will likely require two uploads of your experiment and resetting Salmon

The indices for Validation are indices of the target list, which is available at http://[url]:8421/config. This code generates the list:

>>> import yaml
>>> from pathlib import Path
>>> # config.yaml from [1], copy/pasted into text file named "config.yaml"
>>> # [1]:http://[url]:8421/config
>>> config = yaml.safe_load(Path("config.yaml").open())
>>> # Items will be selected from this list
>>> config["targets"]
['soccer', 'skiing', 'curling', 'skating', 'hockey']

In this case, if you wanted to ask the query “is skating more similar to hockey or curling?”, you would specify:

Validation(..., queries=[(2, 3, 4)])

Let’s check that these are the right indices:

>>> targets[2]
>>> targets[3]
>>> targets[4]

If the YAML configuration file, indices are specified as below:

      - [2, 3, 4]
      - [1, 0, 3]
  # (if asking the above query and
  # a query with head "skiing" and feet "soccer" and "skating".

Sampling detail

Let’s say you want to collect data from three samplers: an active sampler for training, a random sampler for testing and a validation sampler to see measure each participant’s attention (e.g., “are they blindly clicking answers?”).

That sampling is possible through this partial config:

  ARR: {}  # generating the embedding
  Random: {}  # testing the embedding
  Validation: {}  # measure of human quality

  probs: {ARR: 80, Random: 20, Validation: 0}

    1: {sampler: Validation, query: [0, 1, 2]}
    10: {sampler: Validation, query: [0, 1, 2]}

targets: [zero, one, two, three, four]
# targets are indexed by Python. Each target above is a textual representation
# of the index. For index 0 in sampling.details.query will
# show the user ``targets[0] == "zero"``.

   max_queries: 10

With this init.yaml, the crowdsourcing participant will see the same query at the beginning and end (both the 1st and 10th query they see). It will have head "zero" and feet "one" and "two".

More detail on the target indexing is in Validation sampler and Validation. Another example is in Sampling.