cotk
is an open-source lightweight framework for model building and evaluation.
We provides standard dataset and evaluation suites in the domain of general language generation.
It easy to use and make you focus on designing your models!
Features included:
- Light-weight, easy to start. Don't bother your way to construct models.
- Predefined standard datasets, in the domain of language modeling, dialog generation and more.
- Predefined evaluation suites, test your model with multiple metrics in several lines.
- A dashboard to show experiments, compare your and others' models fairly.
- Long-term maintenance and consistent development.
This project is a part of dialtk
(Toolkits for Dialog System by Tsinghua University), you can follow dialtk or cotk on our home page.
Quick links
Index
- python 3
- numpy >= 1.13
- nltk >= 3.4
- tqdm >= 4.30
- checksumdir >= 1.1
- pytorch >= 1.0.0 (optional, used for calculation of perplexity)
- transformers (optional, used for tokenizer of pretrained models)
We support Unix, Windows, and macOS.
You can simply get the latest stable version from pip using
pip install cotk
- Clone the cotk repository
git clone https://github.com/thu-coai/cotk.git
- Install cotk via pip
cd cotk
pip install -e .
Let's skim through the whole package to find what you want.
Load common used dataset and do preprocessing:
- Download online resources or import from local path
- Split training set, development set and test set
- Construct vocabulary list
>>> import cotk.dataloader
>>> # automatically download online resources
>>> dataloader = cotk.dataloader.MSCOCO("resources://MSCOCO_small")
>>> # or download from a url
>>> dl_url = cotk.dataloader.MSCOCO("http://cotk-data.s3-ap-northeast-1.amazonaws.com/mscoco_small.zip#MSCOCO")
>>> # or import from local file
>>> dl_zip = cotk.dataloader.MSCOCO("./MSCOCO.zip#MSCOCO")
>>> print("Dataset is split into:", dataloader.key_name)
["train", "dev", "test"]
Inspect vocabulary list
>>> print("Vocabulary size:", dataloader.vocab_size)
Vocabulary size: 2588
>>> print("First 10 tokens in vocabulary:", dataloader.vocab_list[:10])
First 10 tokens in vocabulary: ['<pad>', '<unk>', '<go>', '<eos>', '.', 'a', 'A', 'on', 'of', 'in']
Convert between ids and strings
>>> print("Convert string to ids", \
... dataloader.convert_tokens_to_ids(["<go>", "hello", "world", "<eos>"]))
Convert string to string [2, 1379, 1897, 3]
>>> print("Convert ids to string", \
... dataloader.convert_ids_to_tokens([2, 1379, 1897, 3]))
Iterate over batches
>>> for data in dataloader.get_batch("train", batch_size=1):
... print(data)
{'sent':
array([[ 2, 181, 13, 26, 145, 177, 8, 22, 12, 5, 1, 1099, 4, 3]]),
# <go> This is an old photo of people and a <unk> wagon.
'sent_allvocabs':
array([[ 2, 181, 13, 26, 145, 177, 8, 22, 12, 5, 3755, 1099, 4, 3]]),
# <go> This is an old photo of people and a horse-drawn wagon.
'sent_length': array([14])}
......
Or using while
if you like
>>> dataloader.restart("train", batch_size=1):
>>> while True:
... data = dataloader.get_next_batch("train")
... if data is None: break
... print(data)
note: If you want to know more about Dataloader
, please refer to docs of dataloader.
We found there are different versions of the same metric in different papers,
which leads to unfair comparison between models. For example, whether considering
unk
, calculating the mean of NLL across sentences or tokens in
perplexity
may introduce huge differences.
We provide a unified implementation for metrics, where hashvalue
is provided for
checking whether the same data is used. The metric object receives mini-batches.
>>> import cotk.metric
>>> metric = cotk.metric.SelfBleuCorpusMetric(dataloader, gen_key="gen")
>>> metric.forward({
... "gen":
... [[2, 181, 13, 26, 145, 177, 8, 22, 12, 5, 3755, 1099, 4, 3],
... [2, 46, 145, 500, 1764, 207, 11, 5, 93, 7, 31, 4, 3]]
... })
>>> print(metric.close())
{'self-bleu': 0.02206768072402293,
'self-bleu hashvalue': 'c206893c2272af489147b80df306ee703e71d9eb178f6bb06c73cb935f474452'}
You can merge multiple metrics together by cotk.metric.MetricChain.
>>> metric = cotk.metric.MetricChain()
>>> metric.add_metric(cotk.metric.SelfBleuCorpusMetric(dataloader, gen_key="gen"))
>>> metric.add_metric(cotk.metric.FwBwBleuCorpusMetric(dataloader, reference_test_list=dataloader.get_all_batch()['sent_allvocabs'], gen_key="gen"))
>>> metric.forward({
... "gen":
... [[2, 181, 13, 26, 145, 177, 8, 22, 12, 5, 3755, 1099, 4, 3],
... [2, 46, 145, 500, 1764, 207, 11, 5, 93, 7, 31, 4, 3]]
... })
>>> print(metric.close())
{'self-bleu': 0.02206768072402293,
'self-bleu hashvalue': 'c206893c2272af489147b80df306ee703e71d9eb178f6bb06c73cb935f474452',
'fw-bleu': 0.3831004349785445, 'bw-bleu': 0.025958979254273006, 'fw-bw-bleu': 0.04862323612604027,
'fw-bw-bleu hashvalue': '530d449a096671d13705e514be13c7ecffafd80deb7519aa7792950a5468549e'}
We also provide recommended metrics for selected dataloader.
>>> metric = dataloader.get_inference_metric(gen_key="gen")
>>> metric.forward({
... "gen":
... [[2, 181, 13, 26, 145, 177, 8, 22, 12, 5, 3755, 1099, 4, 3],
... [2, 46, 145, 500, 1764, 207, 11, 5, 93, 7, 31, 4, 3]]
... })
>>> print(metric.close())
{'self-bleu': 0.02206768072402293,
'self-bleu hashvalue': 'c206893c2272af489147b80df306ee703e71d9eb178f6bb06c73cb935f474452',
'fw-bleu': 0.3831004349785445, 'bw-bleu': 0.025958979254273006, 'fw-bw-bleu': 0.04862323612604027,
'fw-bw-bleu hashvalue': '530d449a096671d13705e514be13c7ecffafd80deb7519aa7792950a5468549e',
'gen': [
['<go>', 'This', 'is', 'an', 'old', 'photo', 'of', 'people', 'and', 'a', 'horse-drawn', 'wagon', '.'],
['<go>', 'An', 'old', 'stone', 'castle', 'tower', 'with', 'a', 'clock', 'on', 'it', '.']
]}
note: If you want to know more about metrics, please refer to docs of metrics.
We provide an online dashboard to manage your experiments.
Here we provide an simple example:
- Initialize a git repository in your command line.
git init
- Write your model with an entry function in
main.py
.
import cotk.dataloader
import json
def run():
dataloader = cotk.dataloader.MSCOCO("resources://MSCOCO_small")
metric = dataloader.get_inference_metric()
metric.forward({
"gen":
[[2, 181, 13, 26, 145, 177, 8, 22, 12, 5, 3755, 1099, 4, 3],
[2, 46, 145, 500, 1764, 207, 11, 5, 93, 7, 31, 4, 3]]
})
json.dump(metric.close(), open("result.json", 'w'))
note: The only requirement of your model is to output a file named result.json
,
you can do whatever you want (even don't load data using cotk
).
- Commit your changes and set upstream branch in your command line.
git add -A
git commit -a -m "init"
git remote add origin master https://github.com/USERNAME/REPONAME.git
git push origin -u master
note: In current version, we only support github for identifying your repository and commit. However, you can use private repositories or do not push your commit to the repository. That means the others cannot access your code or reproduce your results.
- Type
cotk run
to run your model and upload to cotk dashboard.
cotk
will automatically collect your git repository, username (of the dashboard), commit
and result.json
to the cotk dashboard. You can manage your experiments or share results
with others on the dashboard.
If you don't want to use the cotk dashboard, you can also directly upload your model to github. Follow the instructions at Fast Model Reproduction.
note: The reproducibility should be maintained by the author. We only make sure all the inputs
are the same, but differences can be introduced by different random seeds, devices or other
affects. Before you upload, run cotk run --only-run
several times and check whether the results
are the same.
You can download models in dashboard and try to reproduce their results.
cotk download ID
The ID
comes from dashboard id.
cotk
will download the codes from dashboard and tell you how to run the models.
INFO: Fetching USERNAME/REPO/COMMIT
13386B [00:00, 54414.25B/s]
INFO: Codes from USERNAME/REPO/COMMIT fetched.
INFO: Model running cmd written in run_model.sh
Model running cmd: cd ./PATH && cotk run --only-run --entry main
Type cotk run --only-run
will reproduce the same experiments.
You can also directly download your model from github. Follow the instructions at Fast Model Reproduction. For example:
cotk download thu-coai/seq2seq-pytorch/master
We have provided some baselines for the classical tasks, see Model Zoo in docs for details.
You can also use cotk download thu-coai/MODEL_NAME/master
to get the codes.
You are welcome to create an issue if you want to request a feature, report a bug or ask a general question.
We welcome contributions from community.
- If you want to make a big change, we recommend first creating an issue with your design.
- Small contributions can be directly made by a pull request.
- If you like make contributions for our library, see issues to find what we need.
cotk
is maintained and developed by Tsinghua university conversational AI group (THU-coai). Check our main pages (In Chinese).
Apache License 2.0