allenai / lm-explorer

interactive explorer for language models

lm-explorer

interactive explorer for language models (currently only OpenAI GPT-2)

Running with Docker

# This creates a local directory where the model can be cached so you don't
# have to download it everytime you execute 'docker run'.
$ mkdir -p /$HOME/.pytorch_pretrained_bert
$ docker build -t lm-explorer:latest .
$ docker run -p 8000:8000 \
    -v /$HOME/.pytorch_pretrained_bert:/root/.pytorch_pretrained_bert \
    -v $(pwd):/local \
    lm-explorer:latest \
    python app.py --port 8000 --dev

Running without Docker

First create and activate a Python 3.6 (or later) virtual environment. Then install the requirements

$ pip install -r requirements.txt

and start the app

$ python app.py --port 8000 --dev

lm-explorer's Issues

Python 3.8 May Not Work?

Just an FYI, I had dependency issues installing the requirements with Python 3.8, but no problem with Python 3.6. It may be useful to change the documentation to require specifically Python 3.6 or maybe even Python 3.7?

CUDA support

Is there an option to have it run on CUDA ?

No access to Medium sized GPT-2 model

When the app is trying to download your dump from the 345M model of GPT-2 at https://storage.googleapis.com/allennlp/models/gpt2-345M-dump I get an Access denied response.
Is that wanted?

EDIT: I was able to convert the model from a tensorflow checkpoint manually with https://github.com/huggingface/pytorch-pretrained-BERT/issues/582

Empty page by runinng app.py

Following Running without Docker, I got empty page after runningpython app.py --port 8000 --dev and access to localhost:8000.

It seems ok, but I have no idea why I got empty page.

INFO:werkzeug:127.0.0.1 - - [15/Apr/2020 22:11:20] "GET / HTTP/1.1" 200 -

#20 was caused by version of styled-components, thus I fixed styled-components to latest version(5.1.0) but it failed.

static/app.js is not encoded as utf-8 when loaded

In app.py line 53 app_js = open("static/app.js") is not loaded with utf-8 encoding resulting in the hasher in line 54 complaining:

UnicodeDecodeError: 'charmap' codec can't decode byte 0x8f in position 9125: character maps to <undefined>

I solved it by adding the encoding: app_js = open("static/app.js", encoding='utf-8')

weird behavior with textarea resizing

if you shrink the input textarea horizontally, the column it's in keeps the same size.

then when you restore it to the original width, the column grows, forcing the word choices off the screen

Blank page upon running the application

I used pip install -r requirements.txt to install all requirements first and ran python app.py --port 8080 --dev to run the app.

However, I get a blank page on Port 8080.

And the terminal output looks like:

INFO:werkzeug:10.72.16.6 - - [15/Jan/2020 08:37:30] "GET / HTTP/1.1" 200 -
INFO:werkzeug:10.72.16.6 - - [15/Jan/2020 08:37:31] "GET /static/app.js?v=937222bbd3aed82c22ba8ad8e515bd89 HTTP/1.0" 200 -
INFO:werkzeug:10.72.16.6 - - [15/Jan/2020 08:37:32] "GET /static/favicon.ico HTTP/1.0" 200 -

What is the license for the code?

Reason: image not found

python app.py --port 8000 --dev

Traceback (most recent call last):
  File "app.py", line 13, in <module>
    import torch
  File "/Users/fulvio/anaconda3/lib/python3.7/site-packages/torch/__init__.py", line 79, in <module>
    from torch._C import *
ImportError: dlopen(/Users/fulvio/anaconda3/lib/python3.7/site-packages/torch/_C.cpython-37m-darwin.so, 9): Library not loaded: /usr/local/opt/libomp/lib/libomp.dylib
  Referenced from: /Users/fulvio/anaconda3/lib/python3.7/site-packages/torch/lib/libshm.dylib
  Reason: image not found

Empty webpage when run the model (from scratch)

I have cloned this repo to my machine and installed the dependencies as stated but when I run the model with the command python app.py --port 8000 --dev, the resulted web page is empty. Any idea?
I have tried to debug with the console:

styled-components.min.js:1 Uncaught TypeError: Cannot use 'in' operator to search for 'default' in undefined

Uncaught ReferenceError: styled is not defined

Example sentence

Hello,

Is it possible to predict the next or previous word in a sentence as the research claims ?

For example, like Google translate "apple", and produced the example sentences

Thanks