Facebook AI’s VizSeq is an intuitive tool that helps analyzing the performance of text generation dataset and models under. This tool can be deployed as a web application but can also be used in Google Colab.

Install the package

!pip install -q git+https://github.com/facebookresearch/vizseq

Given a source text and a target text, VizSeq let you quickly visualize statistics on the dataset, like: number of examples, sequence lengths, etc.

%matplotlib inline
vizseq.view_stats(sources, references)

VizSeq let you also visualize statistics on the ngrams on the dataset

vizseq.view_n_grams(sources)

VizSeq let you also visualize of metrics (e.g. BLUE) to analyze the performance of your model

vizseq.view_scores(actual_seq, predicted_seq, ['bleu', 'meteor'])

Another hady feature of VizSeq is the manual inspect of the dataset input/output examples of your model against true values:

vizseq.view_examples(input_seq, actual_seq, predicted_seq, ['bleu'], page_sz=2, page_no=12)

Here is notebook example that can be run on Colab - link