# Tabular
## Format specification
Tabular dataset generally refers to table data with multiple rows and columns.
`.csv` files are the most common format, and OpenML uses `.arff` as the official format.
Datumaro only supports tabular data in `.csv` format where the first row is a header with unique column names.
It's because other formats can be converted to `.csv` easily as shown below.
```python
# convert '.arff' to '.csv'
from scipy.io.arff import loadarff
import pandas as pd
data = loadarff("dataset.arff")
df = pd.DataFrame(data[0])
categorical = [col for col in df.columns if df[col].dtype=="O"]
df[categorical] = df[categorical_columns].apply(lambda x: x.str.decode('utf8'))
df.to_csv("arff.csv", index=False)
# convert '.parquet', '.feather', '.hdf5', '.pickle' to '.csv'.
pd.read_parquet("dataset.parquet").to_csv('parquet.csv', index=False)
pd.read_feather("dataset.feather").to_csv('feather.csv', index=False)
pd.read_hdf("dataset.hdf5").to_csv('hdf5.csv', index=False)
pd.read_pickle("dataset.pickle").to_csv('pickle.csv', index=False)
# convert '.jay' to '.csv'
import datatable as dt
data = dt.fread("dataset.jay")
data.to_csv("jay.csv")
```
A tabular dataset can be one of the following:
- a single file with a `.csv` extension
- a directory contains `.csv` files (supports only 1 depth).
```
dataset/
├── aaa.csv
├── ...
└── zzz.csv
```
Supported annotation types:
- `Tabular`
## Convert tabular dataset
A tabular dataset can be converted in the following way:
```bash
datum convert --input-format tabular --input-path \
--output-format --output-dir