ParlAI is a unified platform for training and evaluating dialog models across many tasks.
Currently, the following agents are implemented in this repository.

• RNNAgent by PyTorch
• RNNAgent by Chainer
• RNNAgent by TensorFlow
• AttentionAgent (seq2seq with Attention) by PyTorch
• MemN2NAgent (End-To-End Memory Networks) by Chainer
• SaveAgent (Save losses and attention weights)
• visualization.ipynb (visualize valid and test results, losses and attention weights)

I will implement it soon.

• EntNetAgent (Recurrent Entity Networks) by PyTorch
• RelNetAgent (Relation Networks) by PyTorch
• GeneralDictionaryAgent (Use sub-word, sentencepiece and character-level tokenizer)

I also wrote an article on ParlAI introduction in Japanese. Please see here.

Please download and install ParlAI first.

git clone https://github.com/facebookresearch/ParlAI.git ~/ParlAI
cd ~/ParlAI

pip install -r requirements.txt
sudo python setup.py develop

Then download and put this repository in ~/ParlAI/parlai/.

git clone https://github.com/ryonakamura/parlai_agents.git
mv parlai_agents ~/ParlAI/parlai/

bAbI is a pure text-based QA dataset [Weston, 2015]. There are 20 tasks, each corresponding to a particular type of reasoning, such as deduction, induction, or counting.

According to [Sukhbaatar, NIPS 2015], the mean test accuracy by LSTM in bAbI All 10k is 63.6%.
The following RNNAgent achieves similar mean test accuracy in any library.

Note that the correct labels in Task 19 (path finding) are two words, but RNN generates only one word.
The correct labels in Task 8 (lists/sets) also has two or more words, but the majority is one word.

The meaning of the arguments are as follows.

• -m: the model class name, should match parlai.parlai_agents.<directory_name>.<file_name>:<class_name>
• -t: ParlAI task(s), e.g. babi:Task1k:1 or babi,cbt
• -mf: model file name for loading and saving models
• -e: number of epochs
• -rnn: choose GRU or LSTM
• -bs: batch size for minibatch training schemes
• -hs: size of the hidden layers and embeddings
• -nl: number of hidden layers
• -lr: learning rate
• -dr: dropout rate
• -ltim: log every n secs
• -vtim: validation every n secs

I examined the simple benchmark of 1000 parleys (iterations) speed on single GPU.

in Joint Training

in Single Training (Task 1)

TensorFlow may be slow because it doesn't use truncated BPTT.

PyTorch is an easy and powerful way to implement ParlAI Agent.
When using GPU PyTorch is 1.5 ~ 2 times faster than Chainer.

cd ~/ParlAI
python examples/train_model.py -m parlai.parlai_agents.pytorch_rnn.pytorch_rnn:RNNAgent -t babi:Task1k:1 -mf './parlai/parlai_agents/pytorch_rnn/model_file/babi1' -e 20 -rnn GRU -bs 32 -hs 64 -nl 2 -lr 0.5 -dr 0.2 -ltim 2 -vtim 30

This implementation use multiple inheritance of ParlAI's Agent class and Chainer's Chain classe.
You can choose RNN from links.NStepGRU or links.NStepLSTM.

cd ~/ParlAI
python examples/train_model.py -m parlai.parlai_agents.chainer_rnn.chainer_rnn:RNNAgent -t babi:Task1k:1 -mf './parlai/parlai_agents/chainer_rnn/model_file/babi1' -e 20 -rnn GRU -bs 32 -hs 64 -nl 2 -lr 0.5 -dr 0.2 -ltim 2 -vtim 30

TensorFlow can handle variable length input using tf.nn.dynamic_rnn.
dynamic_rnn internally executes loop processing to unroll time series using control_flow_ops.while_loop (same as tf.while_loop).

If you want to run with only CPU on the GPU server, you can set the environment variable to CUDA_VISIBLE_DEVICES="".

cd ~/ParlAI
python examples/train_model.py -m parlai.parlai_agents.tensorflow_rnn.tensorflow_rnn:RNNAgent -t babi:Task1k:1 -mf './parlai/parlai_agents/tensorflow_rnn/model_file/babi1' -e 20 -rnn GRU -bs 32 -hs 64 -nl 2 -lr 0.5 -dr 0.2 -ltim 2 -vtim 30

The meaning of the additional arguments are as follows.

• -bi: if True, use a bidirectional encoder for the first layer
• -atte: if True, use an Luong's attention
• -cont: if True, use only a context vector without using decoder query for the final output
• -sf: select score function for attention from dot, general, concat
• -tf: teacher forcing ratio

cd ~/ParlAI
python examples/train_model.py -m parlai.parlai_agents.pytorch_attention.pytorch_attention:AttentionAgent -t babi:Task1k:1 -mf './parlai/parlai_agents/pytorch_attention/model_file/babi1' -e 20 -rnn GRU -bi True -atte True -sf general -tf 1. -bs 32 -hs 64 -nl 2 -lr 0.5 -dr 0.2 -ltim 2 -vtim 30

The meaning of the additional arguments are as follows.

• -ms: size of the memory (both key and value)
• -nl: number of memory layers (hops)
• -wt: select weight tying from Adjacent, Layer-wise, Nothing
• -pe: if True, use a Position Encoding for word embedding
• -te: if True, use a Temporal Encoding for sentence memorization
• -rn: if True, use a Random Noise to regularize TE
• -ls: if True, use a Linear Start (remove softmax for the memory layers)
• -opt: select optimizer from SGD, AdaGrad, Adam

cd ~/ParlAI
python examples/train_model.py -m parlai.parlai_agents.chainer_memn2n.chainer_memn2n:MemN2NAgent -t babi:Task1k:1 -mf './parlai/parlai_agents/chainer_memn2n/model_file/babi1' -e 100 -bs 32 -hs 20 -ms 50 -nl 3 -wt Adjacent -pe True -te True -rn True -ls False -opt Adam -lr 0.005 -ltim 2 -vtim 60 -vp -1

Benchmark results to compare this repo implementation, the author's original Matlab code on the bAbI tasks and the paper description.

Default Configuration: 3 Hops, Position Encoding (PE), Temporal Encoding (TE), Linear Start (LS), Random Noise (RN) and Adjacent Weight Tying.

• this repo: No weight tying, No Linear Start, 5 Hops, Hidden Size 128 and Adam (Annealing). It failed at linear start. (trained only once)
• matlab: check it
• paper: repeated 10 times with different random initializations, and picked the one with the lowest training error. check it

bAbI Task paper consider a task successfully passed if ≥ 95% accuracy is obtained.

In the best results of the paper, 14/20 tasks succeeded.
The best settings of this repo has succeeded in 15/20 tasks!

Save losses and attention weights.

The meaning of the arguments are as follows.

• -sltim, --save-loss-every-n-secs: second interval to save losses
• -sae, --save-attention-exs: maximum number of examples to save attention weights

If you have any questions or anything, please do not hesitate to contact me or post on my Github Issues page.

• @_Ryobot on Twitter
• nakamura.ryo.nm8[at]is.naist.jp

All codes in this repository are BSD-licensed.