README.md

chainer-att-seq2seq

Use talk data of Facebook messenger and LINE for training data.

HOW TO USE

$ git clone $(This repository's URL)
$ cd chainer_att-seq2seq/

$ mv ~/Downloads/facebook-$(USER) ./raw/facebook
$ mkdir ./raw/line
$ mv ~/Downloads/\[LINE\]\ Chat\ with\ *.txt ./raw/line/

$ python parse.py
Parse facebook...
Parse line...
done.
$ python train.py
GPU: True
# Minibatch-size: 20
# embed_size: 100
# n_hidden: 100
# epoch: 200

Train
epoch: 1	tag: bigdata
	loss: 108549.05
	accuracy: 3658.35
	time: 0:04:13.800777
.
.
.
$ python decode.py
Interactive decode from ./result/30.npz
> お元気ですか？
元気です
>

Input data

data = [["query sentence", "response sentence"],
	["query sentence", "response sentence"],
	[..., ...], ...]

Important

MeCabの辞書は，mecab-ipadic-neologdを使用しました． MeCab自体のインストールは，mecab-ipadic-neologdを参考にしてください．

CPP

⚠️ -lboost_filesystemオプションは，昔の記事だと，-lboost-filesystemとなっている場合が多いですが，-lboost_filesystemが正しいです．

e.g. compile option

$ g++ -std=c++1z -O3 -mtune=native -march=native -I/usr/local/Cellar/boost/1.66.0 -lboost_filesystem -lboost_system `mecab-config --cflags` `mecab-config --libs` -o parse parse.cpp

速度検証

$ instruments -s
...
Known Templates:
"Activity Monitor"
"Allocations"
"Blank"
"Cocoa Layout"
"Core Animation"
"Core Data"
"Counters"
"Energy Log"
"File Activity"
"Leaks"
"Metal System Trace"
"Network"
"SceneKit"
"System Trace"
"System Usage"
"Time Profiler"
"Zombies"
$ instruments -t "Time Profiler" -l 10000 ./parse
$ open ./instrumentscli0.trace