Skip to main content

The Continual Transfer Learning Benchmark

Project description

Continual Transfer Learning Benchmark

The CTrL benchmark aims to help research by studying the transfer behaviour of different models in the Lifelong Learning context.

For a quick presentation of the benchmark, please check out this blog post.

For more details, the paper is available on arxiv.

Creating a Stream:

The TaskGenerator class is at the center of the implementation of the CTrL Benchmark. It gives access to a high-level API allowing to seamlessly generate a wide variety of streams with a loose coupling between the different components such as the underlying dataset(s), the strategy to generate the tasks (split, incremental, mixture of datasets, ...) and the processing to apply to each task.

The 3 main components of a Task Generator are:

  • A pool of concepts to select from to generate the tasks. It can be a few classes, a full dataset or even a mixture of datasets.
  • A pool of transformation that can be modified or combined to apply specific processing to the data for each task.
  • A Strategy, describing how to combine the conecpts and trasnformation over time to generate an actual stream.

Each of these components can be created by hand or using our automatic TaskGenerator creation tool using yaml configuration files.

For examples simply executing

import ctrl
task_gen = ctrl.get_stream('s_minus')

will return the corresponding task generator that be used either directy to generate tasks on the fly:

t1 = task_gen.add_task()
t2 = task_gen.add_task()
t3 = task_gen.add_task()
...

or as an iterator:

for t in task_gen:
    print(t)

Available streams:

In the current version, only the streams of the CTrL benchmark are directly available, they can be obtained by passing the following name arguments in ctrl.get_stream:

  • S+: "s_plus"
  • S-: "s_minus"
  • Sin: "s_in"
  • Sout: "s_out"
  • Spl: "s_pl"
  • Slong: "s_long"

More documentation and details on the internal components will be progressively added.

See the CONTRIBUTING file for how to help out.

LICENSE

See the LICENSE file.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

ctrl-benchmark-0.0.4.tar.gz (33.6 kB view details)

Uploaded Source

Built Distribution

ctrl_benchmark-0.0.4-py3-none-any.whl (51.0 kB view details)

Uploaded Python 3

File details

Details for the file ctrl-benchmark-0.0.4.tar.gz.

File metadata

  • Download URL: ctrl-benchmark-0.0.4.tar.gz
  • Upload date:
  • Size: 33.6 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.4.2 importlib_metadata/0.22 pkginfo/1.7.1 requests/2.22.0 requests-toolbelt/0.9.1 tqdm/4.47.0 CPython/3.7.4

File hashes

Hashes for ctrl-benchmark-0.0.4.tar.gz
Algorithm Hash digest
SHA256 bdea1617448468662b9ad99dbc6975adb7af1503fd62688d7b621af6c52670a4
MD5 be39a21be5f7ab1e48e4129e740c1629
BLAKE2b-256 5882307fdbdf0f8b153ae4650561389aec4e51df1753f21f435d7fa0a2e20856

See more details on using hashes here.

File details

Details for the file ctrl_benchmark-0.0.4-py3-none-any.whl.

File metadata

  • Download URL: ctrl_benchmark-0.0.4-py3-none-any.whl
  • Upload date:
  • Size: 51.0 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.4.2 importlib_metadata/0.22 pkginfo/1.7.1 requests/2.22.0 requests-toolbelt/0.9.1 tqdm/4.47.0 CPython/3.7.4

File hashes

Hashes for ctrl_benchmark-0.0.4-py3-none-any.whl
Algorithm Hash digest
SHA256 7897040cd73bb4000f30c959da968eeb9a5376c76c7244a06512f34c485362ed
MD5 6292d3b9f1762cbf3dd63de12b6e6e56
BLAKE2b-256 a66df910a12d5656820e2d7a4e795defdac0dd36e0e63565ef4316ccc1fbc7ff

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page