How to keep count of annotations done by a person ?


(Abhishek Dandona) #1

I have a team of annotators that is marking text for NER. I have shared a single IP address for them to work on. So at the moment, there is no way to find how much an individual has.

Can anyone tell me how to find out annotations done per person? And how can I avoid repetition ?

(Ines Montani) #2

If you care about who did the annotations, you ideally want to be starting multiple sessions of Prodigy, e.g. on separate ports, one for each annotator. You can do this manually, or programmatically in Python, a Bash script, etc. Each annotator can then receive their own dedicated dataset, and you’ll be able to compare their annotations to others, and keep track of who did what.

If you search for “multiple annotators”, you’ll find various strategies for implementing a workflow multiple consumers and one annotation task queue. Streams are just Python generators, so you could implement a check that looks at the _task_hash(the unique identifier of an annotation question), and only yields out the question if it’s not already present in any of the other datasets.

You might also be interested in Prodigy Scale, an extension we’ve been working on that will let you manage annotation teams and build large datasets:

In the meantime, you should also check out this open-source multi-user extension, developed by a fellow Prodigy user:

(Abhishek Dandona) #3

Is there a tutorial or a video on how to achieve this ?

(Abhishek Dandona) #4

I am running this code but the I am getting NO_LABEL every time. I saw a similar post here on NO_LABEL, so I tried using ner.manual but, I do not have the functionality of counting number of annotations done per annotator (each annotator has their own link). There is no on_load , update or on_exit option.

import prodigy
import spacy
from multiprocessing import Process
from time import sleep
from import batch_train
import atexit
from pathlib import Path
import datetime as dt

from prodigy.components import printers
from prodigy.components.loaders import get_stream
from prodigy.components.preprocess import add_tokens
from prodigy.core import recipe, recipe_args
from prodigy.util import TASK_HASH_ATTR, log, get_labels
from datetime import datetime
from collections import Counter

# It's all going to be run by coder name.

# Config:
# - add list of coders
# - ?? add port per coder?
# - base file name for files
# - recipe, db, model, output

def mark_custom(dataset, spacy_model, source=None, view_id=None, label=None, api=None,
         loader=None, memorize=False, exclude=None):
    Click through pre-prepared examples, with no model in the loop.
    log('RECIPE: Starting recipe mark', locals())
    nlp = spacy.load(spacy_model)
    log("RECIPE: Loaded model {}".format(spacy_model))
    stream = get_stream(source, api, loader)
    stream = list(add_tokens(nlp, stream))
    labels = get_labels(label, nlp)
    counts = Counter()
    memory = {}
    def fill_memory(ctrl):
        if memorize:
            examples = ctrl.db.get_dataset(dataset)
            log("RECIPE: Add {} examples from dataset '{}' to memory"
                .format(len(examples), dataset))
            for eg in examples:
                memory[eg[TASK_HASH_ATTR]] = eg['answer']

    def ask_questions(stream,nlp=nlp):
        for eg in stream:
            eg['time_loaded'] =
            if TASK_HASH_ATTR in eg and eg[TASK_HASH_ATTR] in memory:
                answer = memory[eg[TASK_HASH_ATTR]]
                counts[answer] += 1
                if label:
                    eg['label'] = label
                yield eg

    def recv_answers(answers):
        for eg in answers:
            counts[eg['answer']] += 1
            memory[eg[TASK_HASH_ATTR]] = eg['answer']
            eg['time_returned'] =

    def print_results(ctrl):

    def get_progress(session=0, total=0, loss=0):
        progress = len(counts) / len(stream)
        return progress
    return {
        'view_id': view_id,
        'dataset': dataset,
        'stream': ask_questions(stream),
        'exclude': exclude,
        'update': recv_answers,
        'on_load': fill_memory,
        'on_exit': print_results,
        'config': {'label': labels}

class MultiProdigy:
    def __init__(self,
        coder_list = [{"name" : "X", "port" : 9010},
                      {"name" : "Y", "port" : 9011}
        self.coder_list = coder_list
        self.processes = []
        self.spacy_model = '/shopin-data/mohit_pandey/PRODIGY/MODEL/ner_iter2_model_batch_train_1'
        self.label = "Details,Length,Lining,Fabric,Neckline,Occasion,Pattern,Personality,Support"

    def serve(self, coder, port):
        base = "/shopin-data/mohit_pandey/PRODIGY/DATA/data_"
        filename = "{0}{1}.jsonl".format(base, coder)
        prodigy.serve('mark_custom',       # recipe
                      "ner_iteration_2_manual_tagged",  # db
                      self.spacy_model, # model
                      filename, # input file
                      "ner_manual", # view ID
                      self.label, # labels
                      None, # api
                      None, # loader
                      True, # memorize
                      "ner_iteration_2_manual_tagged", # exclude
                      port=port)  # port

    def make_prodigies(self):
        for coder_info in enumerate(self.coder_list):
            coder_info = coder_info[1] # wut
            thread =  Process(target=self.serve, args = (coder_info['name'], coder_info['port']))

    def start_prodigies(self):
        print("Starting Prodigy processes...")
        for p in self.processes:

    def kill_prodigies(self):
        print("Killing Prodigy threads")
        for i in self.processes:
            except AttributeError:
                print("Process {0} doesn't exist?".format(i))
        self.processes = []

if __name__ == "__main__":
    mp = MultiProdigy()
    while True:
    #    if > mp.retrain_time:
    #        print("Retraining model and scheduling next retraining for tomorrow")
    #        mp.make_retrain_time() # bump to tomorrow
    #        mp.train_and_restart()

Is there any way I can use ner.manual and get a count of annotations done per annotator and also avoid repetition of questions ?

(Ines Montani) #5

This usually happens if the label list isn’t available via the config. In manual annotation mode, the full label set that’s displayed above the card is a separate config option. Try changing {'label': labels} to {'labels': labels}.

(Abhishek Dandona) #6

Awesome! It is working beautifully. Thanks so much ! :smile:

Just another question, When exactly is on_exit executed in this case ?
When I see that all instances of prodigy have No tasks available, even then the on_exit method doesn’t execute. I have to Ctrl+c every time.

Isn’t the server supposed to shutdown on its own after tasks are completed ?

(Ines Montani) #7

The on_exit method is executed when the server is shut down – e.g. when you hit ctrl + c, or otherwise kill the process.

No, the server will keep running until you shut it down – otherwise, you wouldn’t be able to save your progress, or reload, or perform any other actions anymore. Some annotation projects may work with infinite streams that keep refreshing themselves, so shutting down prematurely may be problematic here.

Annotation goals may also vary from project to project – that’s also why Prodigy itself doesn’t really have a concept of “taks being completed”. In some cases, a task is complete if the progress reaches X%, in other cases you do want to annotate each example once, so the task is complete when all examples are annotated.

(Andy Halterman) #8

I never fully documented the multiuser_prodigy code and it’ll be superseded by Prodigy Scale, but I can quickly walk through what it’s doing.

  1. it draws the tasks to be annotated from a MongoDB that gets filled up from a JSONL loaded in with this code
  2. There’s some logic that governs how examples are pulled from the Mongo (for example, I wanted three annotators to see each document).
  3. The coder identity is hackily stored in the Prodigy output here and here.
  4. Information about what each annotator has done is also written back to the Mongo, both to prevent an annotator from seeing the same example twice, and also so summaries of what each coder has done can be written out and displayed as a report.

(Abhishek Dandona) #9

Thanks for sharing the code, I made changes to the code and made it usable to me. I really like the report maker, it is a good idea.

Thanks, @ines @andy