April 1, 2019

1213 words 6 mins read



An open source library for deep learning end-to-end dialog systems and chatbots.

repo name deepmipt/DeepPavlov
repo link https://github.com/deepmipt/DeepPavlov
homepage https://deeppavlov.ai
language Python
size (curr.) 35292 kB
stars (curr.) 4037
created 2017-11-17
license Apache License 2.0

License Apache 2.0 Python 3.6, 3.7 Downloads

DeepPavlov is an open-source conversational AI library built on TensorFlow and Keras.

DeepPavlov is designed for

  • development of production ready chat-bots and complex conversational systems,
  • research in the area of NLP and, particularly, of dialog systems.

Please leave us your feedback on how we can improve the DeepPavlov framework.


Named Entity Recognition | Slot filling

Intent/Sentence Classification | Question Answering over Text (SQuAD)

Sentence Similarity/Ranking | TF-IDF Ranking

Morphological tagging | Automatic Spelling Correction

ELMo training and fine-tuning


Goal(Task)-oriented Bot | Seq2seq Goal-Oriented bot

Open Domain Questions Answering | eCommerce Bot

Frequently Asked Questions Answering | Pattern Matching


BERT embeddings for the Russian, Polish, Bulgarian, Czech, and informal English

ELMo embeddings for the Russian language

FastText embeddings for the Russian language

Auto ML

Tuning Models with Evolutionary Algorithm


REST API | Socket API | Yandex Alice

Telegram | Microsoft Bot Framework

Amazon Alexa | Amazon AWS


  1. We support Linux and Windows platforms, Python 3.6 and Python 3.7

    • Python 3.5 is not supported!
    • installation for Windows requires Git(for example, git) and Visual Studio 2015/2017 with C++ build tools installed!
  2. Create and activate a virtual environment:

    • Linux
    python -m venv env
    source ./env/bin/activate
    • Windows
    python -m venv env
  3. Install the package inside the environment:

    pip install deeppavlov


There is a bunch of great pre-trained NLP models in DeepPavlov. Each model is determined by its config file.

List of models is available on the doc page in the deeppavlov.configs (Python):

from deeppavlov import configs

When you’re decided on the model (+ config file), there are two ways to train, evaluate and infer it:

GPU requirements

To run supported DeepPavlov models on GPU you should have CUDA 10.0 installed on your host machine and TensorFlow with GPU support (tensorflow-gpu) installed in your python environment. Current supported TensorFlow version is 1.14.0. Run

pip install tensorflow-gpu==1.14.0

before installing model’s package requirements to install supported tensorflow-gpu version.

Before making choice of an interface, install model’s package requirements (CLI):

python -m deeppavlov install <config_path>
  • where <config_path> is path to the chosen model’s config file (e.g. deeppavlov/configs/ner/slotfill_dstc2.json) or just name without .json extension (e.g. slotfill_dstc2)

Command line interface (CLI)

To get predictions from a model interactively through CLI, run

python -m deeppavlov interact <config_path> [-d]
  • -d downloads required data – pretrained model files and embeddings (optional).

You can train it in the same simple way:

python -m deeppavlov train <config_path> [-d]

Dataset will be downloaded regardless of whether there was -d flag or not.

To train on your own data you need to modify dataset reader path in the train config doc. The data format is specified in the corresponding model doc page.

There are even more actions you can perform with configs:

python -m deeppavlov <action> <config_path> [-d]
  • <action> can be
    • download to download model’s data (same as -d),
    • train to train the model on the data specified in the config file,
    • evaluate to calculate metrics on the same dataset,
    • interact to interact via CLI,
    • riseapi to run a REST API server (see doc),
    • telegram to run as a Telegram bot (see doc),
    • msbot to run a Miscrosoft Bot Framework server (see doc),
    • predict to get prediction for samples from stdin or from <file_path> if -f <file_path> is specified.
  • <config_path> specifies path (or name) of model’s config file
  • -d downloads required data


To get predictions from a model interactively through Python, run

from deeppavlov import build_model

model = build_model(<config_path>, download=True)

# get predictions for 'input_text1', 'input_text2'
model(['input_text1', 'input_text2'])
  • where download=True downloads required data from web – pretrained model files and embeddings (optional),
  • <config_path> is path to the chosen model’s config file (e.g. "deeppavlov/configs/ner/ner_ontonotes_bert_mult.json") or deeppavlov.configs attribute (e.g. deeppavlov.configs.ner.ner_ontonotes_bert_mult without quotation marks).

You can train it in the same simple way:

from deeppavlov import train_model 

model = train_model(<config_path>, download=True)
  • download=True downloads pretrained model, therefore the pretrained model will be, first, loaded and then train (optional).

Dataset will be downloaded regardless of whether there was -d flag or not.

To train on your own data you need to modify dataset reader path in the train config doc. The data format is specified in the corresponding model doc page.

You can also calculate metrics on the dataset specified in your config file:

from deeppavlov import evaluate_model 

model = evaluate_model(<config_path>, download=True)

There are also available integrations with various messengers, see Telegram Bot doc page and others in the Integrations section for more info.

Breaking Changes

Breaking changes in version 0.7.0

Breaking changes in version 0.6.0

    • all models default endpoints were renamed to /model
    • by default model arguments names are taken from chainer.in configuration parameter instead of pre-set names from a settings file
    • swagger api endpoint moved from /apidocs to /docs
  • when using "max_proba": true in a proba2labels component for classification, it will return single label for every batch element instead of a list. One can set "top_n": 1 to get batches of single item lists as before

Breaking changes in version 0.5.0

  • dependencies have to be reinstalled for most pipeline configurations
  • models depending on tensorflow require CUDA 10.0 to run on GPU instead of CUDA 9.0
  • scikit-learn models have to be redownloaded or retrained

Breaking changes in version 0.4.0!

  • default target variable name for neural evolution was changed from MODELS_PATH to MODEL_PATH.

Breaking changes in version 0.3.0!

  • component option fit_on_batch in configuration files was removed and replaced with adaptive usage of the fit_on parameter.

Breaking changes in version 0.2.0!

  • utils module was moved from repository root in to deeppavlov module
  • ms_bot_framework_utils,server_utils, telegram utils modules was renamed to ms_bot_framework, server and telegram correspondingly
  • rename metric functions exact_match to squad_v2_em and squad_f1 to squad_v2_f1
  • replace dashes in configs name with underscores

Breaking changes in version 0.1.0!

  • As of version 0.1.0 all models, embeddings and other downloaded data for provided configurations are by default downloaded to the .deeppavlov directory in current user’s home directory. This can be changed on per-model basis by modifying a ROOT_PATH variable or related fields one by one in model’s configuration file.

  • In configuration files, for all features/models, dataset readers and iterators "name" and "class" fields are combined into the "class_name" field.

  • deeppavlov.core.commands.infer.build_model_from_config() was renamed to build_model and can be imported from the deeppavlov module directly.

  • The way arguments are passed to metrics functions during training and evaluation was changed and documented.


DeepPavlov is Apache 2.0 - licensed.

The Team

DeepPavlov is built and maintained by Neural Networks and Deep Learning Lab at MIPT within iPavlov project.

comments powered by Disqus