fastcnn

fastcnn is another text classification CLI tool (inspired by fasttext).

Char-level CNN to classify
- supports binary/multi-class/multi-label classification
Command line tool (you don't need to write Python)
Not yet fast
Implemented with tensorflow.keras

Build, Install

We prepare Dockerfile and bin/fastcnn. I recommend this style rather than directly running.

make build  # build docker image
export PATH=$PWD/bin:$PATH  # add ./bin/fastcnn

If you need sudo privilege to run docker, please use properly.

Usage

Supervised training

The task type (binary, multi-class or multi-labels classification) are automatically detected. Some parameters also are adjusted.

# sample: binary classification
$ fastcnn supervised \
    ./samples/en_ja/input \
    --validate ./samples/en_ja/validate \
    --maxlen 20 --epochs 10 --lr 0.2

# sample: categorical (3 classes) classification
$ fastcnn supervised \
    ./samples/python_bash_coq/train \
    --validate ./samples/python_bash_coq/valid \
    --epochs 300 \
    --dim 8 \
    --kernel-size 3 \
    --lr 0.3

After fastcnn supervised, model parameter file {out}.h5 and a meta info file {out}.meta.yml are generated. Here, {out} is "out" by default (You can specify with -o).

test and predict

You can test the trained model by fastcnn test. And fastcnn predict shows the prediction.

$ fastcnn test out ./samples/python_bash_coq/test
Acc@1: 0.8883
__label__bash
- Recall: 0.9103
- Prec: 0.9595
- F1: 0.9342
__label__coq
- Recall: 0.7045
- Prec: 1.0000
- F1: 0.8267
__label__python
- Recall: 1.0000
- Prec: 0.7703
- F1: 0.8702

$ fastcnn predict out ./samples/python_bash_coq/test --show-data | head
__label__bash - discriminate.
__label__coq - done.
__label__coq - exact.
__label__python Example trans_eq_example :
__label__python From mathcomp Require Import all_ssreflect.
__label__python Import ListNotations.
__label__coq Proof.
__label__coq Qed.
__label__python Require Import Arith List Omega ZArith.
__label__coq S n = S m -> n = m.

more detail