Data Inputs

Input Types

• Sources: Source-side inputs for text generation tasks.
• References: Target-side references for text generation tasks.
• Hypothesis: Model predictions. One sentence per example per model.
• Tags: Per-example tags for example grouping.

Data Sources

VizSeq accepts data inputs in various ways.

Python Variables

VizSeq Jupyter notebook interface accepts Python list of lists of strings as inputs. The inner list is for multiple inputs of the same type, for example:

ref_1 = ['This is ref #1 for example #1.', 'This is ref #1 for example #2.']
ref_2 = ['This is ref #2 for example #1.', 'This is ref #2 for example #2.']
ref_3 = ['This is ref #3 for example #1.', 'This is ref #3 for example #2.']
references = [ref_1, ref_2, ref_3]

Plain text or ZIP file paths

VizSeq also accepts file paths. For example for Jupyter notebook:

references = ['ref_1.txt', 'ref_2.txt', 'ref_3.txt']

For the web App interface, it expects data to be organized in the following folder structure:

[data_root]/[task_name]/src_*.txt
[data_root]/[task_name]/src_*.zip
[data_root]/[task_name]/ref_*.txt
[data_root]/[task_name]/pred_*.txt
[data_root]/[task_name]/tag_*.txt

where

• src_*.txt: A text source, one sentence per line.
• src_*.zip: A packed source, with a source.txt in it, which provides either the sentences or the image/audio/video filenames per line.
• ref_*.txt: A text reference, one sentence per line.
• pred_*.txt: A text model prediction, one sentence per line.
• tag_*.txt: (Optional) Example tags, one phrase per line.

File Formats

Text

• .txt
• .txt in .zip

Image (packed in .zip)

• .jpg
• .png
• .gif
• .bmp

Audio (packed in .zip)

• .mp3
• .wav

Video (packed in .zip)

• .mp4
• .webm