Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Tensor Parser

Build Status

A package for constructing sparse tensors from CSV-like data sources. This package constructs mappings from columns in the CSV file(s) to contiguous indices in the tensor and merges duplicate non-zeros (i.e., duplicate rows).


tensor_parser is written in Python3. Its dependencies are:

  • python >= 3.5
  • python-dateutil
  • csvsorter from GitHub

To install external dependencies, you can simply use pip:

$ pip install -r requirements.txt

To run, use the entry point:

$ ./scripts/ --help

CSV Files

We support CSV files stored in text, compressed gzip (.gz), or compressed bzip2 (.bz2) formats.

By default, we attempt to auto-detect the header and delimiter of the CSV file via Python's supplied CSV parsing library. The --query option will query the detected CSV metadata and print to STDOUT:

$ ./build/ traffic.csv.gz out.tns --query
Found delimiter: ","
Found fields:
['Date Of Stop', 'Time Of Stop', 'Latitude', 'Longitude', 'Description']

Note that out.tns is not touched when querying a CSV file, though it is required as a positional argument.

Any number of CSV files can be provided for output, so long as the fields used to construct the sparse tensor are found in each file.

If no header is detected, a default of ["1", "2", ...] is used.

If you wish to use something other than the detected delimiter or field names, they can be modified with --field-sep= and --has-header=<yes,no>.

Tensor Files

Two types of files are created:

  • Sparse tensor (.tns): the actual tensor data. Each line is a list of    indices and a value. For example, 1 1 1 1.38 would be one non-zero in a third-order tensor.
  • Mode mappings (.map): map the tensor indices to the original values in the source data. Line i is the source data that was mapped to index i in the tensor.

For more information on file formats, see FROSTT.

Tensor Construction

Mode selection

Columns of the CSV file (referred to as "fields") are selected using the --field= flag (abbreviated -f). If the CSV file has a header, the supplied parameter must match a field in the header (but is not case sensitive). Otherwise, the columns are referenced by number and one-indexed. If the field has spaces in the name, simply enclose it in quotes: --field="time of day".

Tensor values

A field of the CSV file can be selected to be used as the values of the tensor using the --vals= flag. If no field is selected to use as the tensor values, 1 is used and the resulting tensor is one of count data.

Mode Types

A critical step when constructing a sparse tensor is to select the datatype of the CSV columns. When the CSV is parsed, the fields are read and sorted as strings. Thus, values with the same string representation are mapped to the same index in the tensor mode. In practice, however, columns often should be treated as integers, floats, dates, or other types.

In addition to affecting the ordering of the resulting indices, the type of a column affects the mapping of CSV entries to unique indices. For example, one may wish to round floats such that 1.38 and 1.38111 map to the same index, or to map dates Aug 20 and August 20 to the same index.

We provide several types which can be specified with the --type= flag:

  • str => String (default)
  • int => Integer
  • float => Floating-point number
  • roundf-X => Floating-point numbers rounded to X decimal places
  • date => A datetime object that encodes year, month, day, hour, minute, second, and millisecond
  • year => A year (integer extracted from date)
  • month => A month (integer in range [0,11] extracted from date)
  • day => A day (integer in range [0,30] extracted from date)
  • hour => A hour (integer in range [0,23] extracted from date)
  • min => A min (integer in range [0,60] extracted from date)
  • sec => A sec (integer in range [0,60] extracted from date)

Smart date matching is provided by the dateutil package. For example, "Aug 20" and "08/20/92" will map to the same index if the type is either month or day. However, the package maps to the current year if none is specified, and thus they will map to different indices if the type is year.

You can specify multiple fields in the same --type instance. For example: --type=userid,itemid,int would treat the fields userid and itemid both as integers.

Advanced mode types

A "type" in our context is any object which supports:

  • construction: type("X") should return some representation of "X" (note that a string is always the parameter)
  • comparison: __le__() is required to sort indices. If no comparison is possible, be sure to disable sorting of the mode with --no-sort
  • printing: __str__() is required to construct .map files

The specification of a type is as simple as providing a function which maps a string to some object. Conveniently, most builtin types already support this interface via their constructors. Functions such as int() and float() work well.

Many types can be specified with a short anonymous function. If the specified type is not found in the list of builtin types (above), then it is treated as source code and specifies a custom type. For example,

--type=latitude,"lambda x : round(float(x),3)"

would round all latitudes to three decimal places and is equivalent to the builtin type roundf-3. Note that all type functions take a single parameter which will be an str object.

Pruning Tensor Entries

A type function can return None to omit a non-zero from the tensor. For example,

--type=age,"lambda x : int(x) if int(x) > 16 else None"

would omit any non-zeros whose age is less than 16, and treat the remaining ages as integers.

Handling Duplicates

By default, duplicate non-zero values are removed and their values are summed. This behavior can be changed with --merge=, which takes one of the following options:

  • none (do not remove duplicate non-zeros)
  • sum
  • min
  • max
  • avg
  • count (use the number of duplicates)

Note that merging duplicates requires the tensor to be sorted. A disk-based sort is provided by a fork of the csvsorter library.


Suppose you have the following CSV file:

$ zcat traffic.csv.gz
Date Of Stop,Time Of Stop,Latitude,Longitude,Description

We want to keep the dates, hour of violation, lower-case description, and round the geolocations to three decimal places:

$ ./scripts/ traffic.csv.gz traffic.tns \
    -f "date of stop" --type="date of stop",date \
    -f "time of stop" --type="time of stop",hour \
    -f latitude -f longitude --type=latitude,longitude,roundf-3 \
    -f description --type=description,"lambda x : x.lower()"

The resulting tensor is built:

$ cat
2013-01-01 00:00:00

$ cat

$ cat

$ cat

$ cat

$ cat traffic.tns
1 1 1 2 3 1
1 2 2 3 1 1
1 3 3 1 2 1


This project uses the builtin unittest library provided by Python. You can run all unit tests via:

$ python3 -m unittest


A package for constructing sparse tensors from CSV-like data sources.







No releases published


No packages published