Added GPT2-Wildtalk module

[Toseban]

First version of Wildtalk using GPT2 (345M) :speaking_head: There are still some things to do:

• [x] Add some config options for server and gpt2-parameters :card_file_box:
• [x] Finetuning of gpt2-parameters :gear:
• [x] Integrate old wildtalk into server-client structure :arrow_right: Ideally plug and play for different models :electric_plug:
• [x] Currently pytorch-pretrained-bert is a GitHub requirement, because the latest version on pypi does not have the 345M model yet. :construction:
• [ ] Download models when building docker image :arrow_down:
• [ ] Add some tests! :chart:

[codecov-io]

Codecov Report

Merging #117 into velvet-revolution will decrease coverage by 7.55%. The diff coverage is 12.13%.

@@                  Coverage Diff                  @@
##           velvet-revolution     #117      +/-   ##
=====================================================
- Coverage              75.12%   67.56%   -7.56%     
=====================================================
  Files                     57       63       +6     
  Lines                   3509     3928     +419     
=====================================================
+ Hits                    2636     2654      +18     
- Misses                   873     1274     +401

Impacted Files	Coverage Δ
modules/ravestate/context.py	94.96% <ø> (ø)	:arrow_up:
modules/ravestate_akinator/__init__.py	44.44% <0%> (-1.46%)	:arrow_down:
modules/ravestate_ros1/__init__.py	0% <0%> (ø)
modules/ravestate_wildtalk/parlai_backend.py	0% <0%> (ø)
modules/ravestate_wildtalk/server.py	0% <0%> (ø)
modules/ravestate_telegramio/telegram_bot.py	41.84% <0%> (-0.46%)	:arrow_down:
modules/ravestate_roboyio/__init__.py	0% <0%> (ø)	:arrow_up:
modules/ravestate_wildtalk/convai_gpt_backend.py	0% <0%> (ø)
modules/ravestate_wildtalk/gpt2_backend.py	0% <0%> (ø)
modules/ravestate_ros1/ros1_properties.py	0% <0%> (ø)
... and 14 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 19adeaa...46bca78. Read the comment docs.

[josephbirkner]

[josephbirkner]

Problem: URL dependency breaks PyPI deployment. Use this handy snippet to download & install the model on first run (stolen from spacy):

def download_model(filename, user_pip_args=None):
    download_url = about.__download_url__ + "/" + filename
    pip_args = ["--no-cache-dir", "--no-deps"]
    if user_pip_args:
        pip_args.extend(user_pip_args)
    cmd = [sys.executable, "-m", "pip", "install"] + pip_args + [download_url]
    return subprocess.call(cmd, env=os.environ.copy())

[josephbirkner]

The server-client architecture with on-demand autostart is awesome. But I think we can skip integration of old Wildtalk. Nobody will use it (probably). Opinions @missxa?

[Toseban]

@josephbirkner The old Wildtalk is a more reliable in giving "normal responses" I think, so I would include it. Try the gpt2 Wildtalk for a few conversations and let me know what you think

[Toseban]

The old wildtalk module (roboy_parlai) :older_man: is now integrated as well as the conversational AI from https://github.com/huggingface/transfer-learning-conv-ai :speaking_head:

Models and their options can be set with ravestate configs :balloon: Server can also be started externally on its own :earth_africa: