The present repo contains the code accompanying the blog post π¦ How to build a State-of-the-Art Conversational AI with Transfer Learning.
This code is a clean and commented code base with training and testing scripts that can be used to train a dialog agent leveraging transfer Learning from an OpenAI GPT and GPT-2 Transformer language model.
This codebase can be used to reproduce the results of HuggingFace's participation to NeurIPS 2018 dialog competition ConvAI2 which was state-of-the-art on the automatic metrics. The 3k+ lines of competition code was distilled in about 250 lines of training code with distributed & FP16 options to form the present repository.
This model can be trained in about one hour on a 8 V100 cloud instance (currently costs about $25) and a pre-trained model is also made available.
To install and use the training and inference scripts please clone the repo and install the requirements:
git clone https://github.com/huggingface/transfer-learning-conv-ai
cd transfer-learning-conv-ai
pip install -r requirements.txt
python -m spacy download en
To install using docker please build the self-contained image:
docker build -t convai .
Note: Make sure your Docker setup allocates enough memory to building the container. Building with the default of 1.75GB will fail due to large Pytorch wheel.
You can then enter the image
ip-192-168-22-157:transfer-learning-conv-ai loretoparisi$ docker run --rm -it convai bash
root@91e241bb823e:/# ls
Dockerfile README.md boot dev home lib media models proc root sbin sys train.py utils.py
LICENCE bin convai_evaluation.py etc interact.py lib64 mnt opt requirements.txt run srv tmp usr var
You can then run the interact.py
script on the pretrained model:
python3 interact.py --model models/
We make a pretrained and fine-tuned model available on our S3 here. The easiest way to download and use this model is just to run the interact.py
script to talk with the model. Without any argument, this script will automatically download and cache our model.
The training script can be used in single GPU or multi GPU settings:
python ./train.py # Single GPU training
python -m torch.distributed.launch --nproc_per_node=8 ./train.py # Training on 8 GPUs
The training script accept several arguments to tweak the training:
Argument | Type | Default value | Description |
---|---|---|---|
dataset_path | str |
"" |
Path or url of the dataset. If empty download from S3. |
dataset_cache | str |
'./dataset_cache.bin' |
Path or url of the dataset cache |
model | str |
"openai-gpt" |
Path, url or short name of the model |
num_candidates | int |
2 |
Number of candidates for training |
max_history | int |
2 |
Number of previous exchanges to keep in history |
train_batch_size | int |
4 |
Batch size for training |
valid_batch_size | int |
4 |
Batch size for validation |
gradient_accumulation_steps | int |
8 |
Accumulate gradients on several steps |
lr | float |
6.25e-5 |
Learning rate |
lm_coef | float |
1.0 |
LM loss coefficient |
mc_coef | float |
1.0 |
Multiple-choice loss coefficient |
max_norm | float |
1.0 |
Clipping gradient norm |
n_epochs | int |
3 |
Number of training epochs |
personality_permutations | int |
1 |
Number of permutations of personality sentences |
device | str |
"cuda" if torch.cuda.is_available() else "cpu" |
Device (cuda or cpu) |
fp16 | str |
"" |
Set to O0, O1, O2 or O3 for fp16 training (see apex documentation) |
local_rank | int |
-1 |
Local rank for distributed training (-1: not distributed) |
Here is how to reproduce our results on a server with 8 V100 GPUs (adapt number of nodes and batch sizes to your configuration):
python -m torch.distributed.launch --nproc_per_node=8 ./train.py --gradient_accumulation_steps=4 --lm_coef=2.0 --max_history=2 --n_epochs=1 --num_candidates=4 --personality_permutations=2 --train_batch_size=2 --valid_batch_size=2
This model should give a Hits@1 over 79, perplexity of 20.5 and F1 of 16.5 using the convai2 evaluation script (see below).
These numbers are slightly lower than the number we obtained in the ConvAI2 competition. Here is what you can tweak to reach the same results:
- in the ConvAI2 competition we also used tweaked position emebddings so that the history of the dialog always start at with the same embeddings. This is easy to add with pytorch-transformers and should improve the hits@1 metric.
- in the ConvAI2 competition we used a beam search decoder. While the results are better in term of f1 metric, our feeling is that the human experience is less compelling with beam search versus the nucleus sampling detector which is provided in the present repository.
The training script saves all the experiments and checkpoints in a sub-folder named with the timestamp of the experiment in the ./runs
folder of the repository base folder.
You can then use the interactive script to interact with the model simply by pointing to this folder.
Here is an example command line to run the interactive script:
python ./interact.py --model_checkpoint ./data/Apr17_13-31-38_thunder/ # run the interactive script with a training checkpoint
python ./interact.py # run the interactive script with the finetuned model on our S3
The fine-tuned model will gives FINAL Hits@1: 0.715
The interactive script accept a few arguments to tweak the decoding algorithm:
Argument | Type | Default value | Description |
---|---|---|---|
dataset_path | str |
"" |
Path or url of the dataset. If empty download from S3. |
dataset_cache | str |
'./dataset_cache.bin' |
Path or url of the dataset cache |
model | str |
"openai-gpt" |
Path, url or short name of the model |
max_history | int |
2 |
Number of previous utterances to keep in history |
device | str |
cuda if torch.cuda.is_available() else cpu |
Device (cuda or cpu) |
no_sample | action store_true |
Set to use greedy decoding instead of sampling | |
max_length | int |
20 |
Maximum length of the output utterances |
min_length | int |
1 |
Minimum length of the output utterances |
seed | int |
42 |
Seed |
temperature | int |
0.7 |
Sampling softmax temperature |
top_k | int |
0 |
Filter top-k tokens before sampling (<=0 : no filtering) |
top_p | float |
0.9 |
Nucleus filtering (top-p) before sampling (<=0.0 : no filtering) |
To run the evaluation scripts of the ConvAI2 challenge, you first need to install ParlAI
in the repo base folder like this:
git clone https://github.com/facebookresearch/ParlAI.git
cd ParlAI
python setup.py develop
You can then run the evaluation script from ParlAI
base folder:
cd ParlAI
python ../convai_evaluation.py --eval_type hits@1 # to download and evaluate our fine-tuned model on hits@1 metric
python ../convai_evaluation.py --eval_type hits@1 --model_checkpoint ./data/Apr17_13-31-38_thunder/ # to evaluate a training checkpoint on hits@1 metric
The evaluation script accept a few arguments to select the evaluation metric and tweak the decoding algorithm:
Argument | Type | Default value | Description |
---|---|---|---|
eval_type | str |
"hits@1" |
Evaluate the model on hits@1 , ppl or f1 metric on the ConvAI2 validation dataset |
model | str |
"openai-gpt" |
Path, url or short name of the model |
max_history | int |
2 |
Number of previous utterances to keep in history |
device | str |
cuda if torch.cuda.is_available() else cpu |
Device (cuda or cpu) |
no_sample | action store_true |
Set to use greedy decoding instead of sampling | |
max_length | int |
20 |
Maximum length of the output utterances |
min_length | int |
1 |
Minimum length of the output utterances |
seed | int |
42 |
Seed |
temperature | int |
0.7 |
Sampling softmax temperature |
top_k | int |
0 |
Filter top-k tokens before sampling (<=0 : no filtering) |
top_p | float |
0.9 |
Nucleus filtering (top-p) before sampling (<=0.0 : no filtering) |
see example_entry.py
, and the comment at the top.
If you use this code in your research, you can cite our NeurIPS CAI workshop paper:
@article{DBLP:journals/corr/abs-1901-08149,
author = {Thomas Wolf and
Victor Sanh and
Julien Chaumond and
Clement Delangue},
title = {TransferTransfo: {A} Transfer Learning Approach for Neural Network
Based Conversational Agents},
journal = {CoRR},
volume = {abs/1901.08149},
year = {2019},
url = {http://arxiv.org/abs/1901.08149},
archivePrefix = {arXiv},
eprint = {1901.08149},
timestamp = {Sat, 02 Feb 2019 16:56:00 +0100},
biburl = {https://dblp.org/rec/bib/journals/corr/abs-1901-08149},
bibsource = {dblp computer science bibliography, https://dblp.org}
}