facebookresearch / vizseq Goto Github PK
View Code? Open in Web Editor NEWAn Analysis Toolkit for Natural Language Generation (Translation, Captioning, Summarization, etc.)
Home Page: https://arxiv.org/abs/1909.05424
License: MIT License
An Analysis Toolkit for Natural Language Generation (Translation, Captioning, Summarization, etc.)
Home Page: https://arxiv.org/abs/1909.05424
License: MIT License
Bert Score (https://github.com/Tiiiger/bert_score) new version (0.3.1
) supports oop implementation. Current vizseq
uses the functional implementation which could be updated to oop implementation.
Currently, using Bert Score in a validation loop causes re-loading the model again and again. This can be avoided with oop implementation.
I see two solutions: (i) create a separate scorer named bert_score_oop
(ii) in the current implementation of bert_score
add argument whether to use oop implementation or not.
Are you willing to open a pull request?
Yes, I can send a pull request
I tried to apply BertScore on my data, but received this error :
TypeError: score() got an unexpected keyword argument 'bert'
In configuration, select BertScore
as metric.
Refresh the page
Stack trace/error message
Traceback (most recent call last):
File "/home/me/.venv/presum/lib/python3.6/site-packages/tornado/web.py", line 1590, in _execute
result = method(*self.path_args, **self.path_kwargs)
File "/home/me/workspace/vizseq/vizseq/server.py", line 103, in get
pd = wv.get_page_data()
File "/home/me/workspace/vizseq/vizseq/_view/web_view.py", line 158, in get_page_data
sorting_metric=self.sorting_metric, need_lang_tags=True
File "/home/me/workspace/vizseq/vizseq/_view/data_view.py", line 132, in get
for s in metrics
File "/home/me/workspace/vizseq/vizseq/_view/data_view.py", line 132, in <dictcomp>
for s in metrics
File "/home/me/workspace/vizseq/vizseq/_view/data_view.py", line 130, in <dictcomp>
for m, hh in cur_hypo.items()
File "/home/me/workspace/vizseq/vizseq/scorers/bert_score.py", line 28, in score
no_idf=True, verbose=self.verbose
TypeError: score() got an unexpected keyword argument 'bert'
Able to see BertScore
.
When having multiple references (attached muti_refs.zip), I cannot configure metric BLEU (but I can configure GLEU). I get a 500 internal server error.
This does not happen with a single reference (see attached single_ref.zip).
Executing vizseq.view_stats
breaks the layout of the Jupyter. The menu at the top obscures a majority of the screen and a blank area of ~60px appears at the top of the page.
** Minimal Code/Config snippet to reproduce **
jupyter notebook
When the first cell containing vizseq.view_stats
is finishes the layout changes and appears broken.
The display of tables and graphs by vizseq does not affect the layout of the Jupyter notebook.
Cause: The bootstrap.min.css and an inline stylesheet loaded by vizseq break the layout. The inline stylesheet is:
body {
padding-top: 60px; /* 60px to make the container go all the way to the bottom of the topbar */
}
The inline stylesheet is responsible for the blank bar at the top while Bootstrap breaks the menu's formatting.
To test this disable both stylesheets in the stylesheet editor included in the developer tools of a browser.
Audio segments from speech data in example speech translation task (IWSLT17 dev) are not correctly associated with reference data.
Only the first TED talk audio segments are correctly aligned to the reference... playing the audio segments related to any other talks (from # 3 / 10 / 887 ( 153 / 887 ) onwards on page 16 of the task using the defaults) results in the segments of the first TED talk audio being played rather than the segments specified in the task directory speech_translation_iwslt17_dev/src_0.zip/source.txt
Get the example speech task data (IWSLT7 dev)
$ bash get_example_data.sh speech_translation_iwslt17_dev
Start the server and navigate to :
http://127.0.0.1:5000/view?t=speech_translation_iwslt17_dev&m=&q=&p_sz=10&p_no=16&s=0&s_metric=
Play the audio segments: first two on this page will be correctly associated with reference text, from # 3 / 10 / 887 ( 153 / 887 ) onwards they are not.
vizseq.scorers.bleu.BLEUScorer
does not use Tokenizer13a
by default. When I look at the code, it looks like it should be used by default. sacrebleu
library uses the Tokenizer13a
by default as well.
Minimal Code/Config snippet to reproduce
import vizseq
scorer = vizseq.scorers.bleu.BLEUScorer()
print(scorer.score(["This is really nice."], [["That's really nice."]]))
# corpus_score = 31.947
scorer = vizseq.scorers.bleu.BLEUScorer(extra_args={'tokenizer': '13a'})
print(scorer.score(["This is really nice."], [["That's really nice."]]))
# corpus_score = 39.764
Stack trace/error message
The problem is here. Variable tokenizer
is set to string none
. When calling method get_default_args
(here), default value 13a
for parameter tokenize
is not used, because the string none
is passed.
vizseq.scorers.bleu.BLEUScorer
should use Tokenizer13a
by default.
[jw@cn05 ~]$ pip3 install vizseq
Defaulting to user installation because normal site-packages is not writeable
Collecting vizseq
Using cached vizseq-0.1.15-py3-none-any.whl (81 kB)
Collecting nltk>=3.5
Using cached nltk-3.5-py3-none-any.whl
Collecting sacrebleu>=1.4.13
Using cached sacrebleu-1.5.0-py3-none-any.whl (65 kB)
Collecting langid
Using cached langid-1.1.6.tar.gz (1.9 MB)
Requirement already satisfied: tqdm in ./.local/lib/python3.9/site-packages (from vizseq) (4.31.1)
Collecting google-cloud-translate
Using cached google_cloud_translate-3.0.2-py2.py3-none-any.whl (93 kB)
Collecting torch
Using cached torch-0.1.2.post2.tar.gz (128 kB)
Requirement already satisfied: numpy in ./.local/lib/python3.9/site-packages (from vizseq) (1.19.5)
Requirement already satisfied: jinja2 in ./.local/lib/python3.9/site-packages (from vizseq) (2.10.3)
Collecting soundfile
Using cached SoundFile-0.10.3.post1-py2.py3-none-any.whl (21 kB)
Requirement already satisfied: py-rouge in ./.local/lib/python3.9/site-packages (from vizseq) (1.1)
Requirement already satisfied: matplotlib in ./.local/lib/python3.9/site-packages (from vizseq) (3.3.2)
Requirement already satisfied: tornado in ./.local/lib/python3.9/site-packages (from vizseq) (6.1)
Requirement already satisfied: IPython in ./.local/lib/python3.9/site-packages (from vizseq) (7.18.1)
Collecting bert-score
Using cached bert_score-0.3.7-py3-none-any.whl (53 kB)
Requirement already satisfied: pandas in ./.local/lib/python3.9/site-packages (from vizseq) (1.1.4)
Collecting laserembeddings
Using cached laserembeddings-1.1.1-py3-none-any.whl (13 kB)
Requirement already satisfied: click in ./.local/lib/python3.9/site-packages (from nltk>=3.5->vizseq) (7.1.2)
Requirement already satisfied: regex in ./.local/lib/python3.9/site-packages (from nltk>=3.5->vizseq) (2020.11.13)
Requirement already satisfied: joblib in ./.local/lib/python3.9/site-packages (from nltk>=3.5->vizseq) (0.17.0)
Collecting portalocker
Using cached portalocker-2.2.1-py2.py3-none-any.whl (15 kB)
Collecting transformers>=3.0.0
Using cached transformers-4.3.3-py3-none-any.whl (1.9 MB)
Collecting bert-score
Using cached bert_score-0.3.6-py3-none-any.whl (53 kB)
Using cached bert_score-0.3.5-py3-none-any.whl (52 kB)
Using cached bert_score-0.3.4-py3-none-any.whl (52 kB)
Using cached bert_score-0.3.3-py3-none-any.whl (52 kB)
Using cached bert_score-0.3.2-py3-none-any.whl (52 kB)
Using cached bert_score-0.3.1-py3-none-any.whl (51 kB)
Using cached bert_score-0.3.0-py3-none-any.whl (48 kB)
Using cached bert_score-0.2.3-py3-none-any.whl (15 kB)
Using cached bert_score-0.2.2-py3-none-any.whl (14 kB)
Using cached bert_score-0.1.2-py3-none-any.whl (9.4 kB)
Using cached bert_score-0.1.1-py3-none-any.whl (9.4 kB)
Using cached bert_score-0.1.0-py3-none-any.whl (7.3 kB)
INFO: pip is looking at multiple versions of to determine which version is compatible with other requirements. This could take a while.
INFO: pip is looking at multiple versions of sacrebleu to determine which version is compatible with other requirements. This could take a while.
Collecting sacrebleu>=1.4.13
Using cached sacrebleu-1.4.14-py3-none-any.whl (64 kB)
Using cached sacrebleu-1.4.13-py3-none-any.whl (43 kB)
INFO: pip is looking at multiple versions of nltk to determine which version is compatible with other requirements. This could take a while.
INFO: pip is looking at multiple versions of vizseq to determine which version is compatible with other requirements. This could take a while.
Collecting vizseq
Using cached vizseq-0.1.14-py3-none-any.whl (81 kB)
Using cached vizseq-0.1.13-py3-none-any.whl (81 kB)
Using cached vizseq-0.1.12-py3-none-any.whl (81 kB)
Using cached vizseq-0.1.11-py3-none-any.whl (81 kB)
Using cached vizseq-0.1.10-py3-none-any.whl (80 kB)
Using cached vizseq-0.1.9-py3-none-any.whl (78 kB)
Requirement already satisfied: nltk in ./.local/lib/python3.9/site-packages (from vizseq) (3.4.5)
Collecting sacrebleu==1.4.7
Using cached sacrebleu-1.4.7-py3-none-any.whl (59 kB)
Requirement already satisfied: typing in ./.local/lib/python3.9/site-packages (from sacrebleu==1.4.7->vizseq) (3.7.4.3)
Collecting mecab-python3
Using cached mecab-python3-1.0.3.tar.gz (77 kB)
INFO: pip is looking at multiple versions of to determine which version is compatible with other requirements. This could take a while.
INFO: pip is looking at multiple versions of sacrebleu to determine which version is compatible with other requirements. This could take a while.
ERROR: Cannot install vizseq and vizseq==0.1.9 because these package versions have conflicting dependencies.
The conflict is caused by:
vizseq 0.1.9 depends on torch
bert-score 0.3.7 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.3.6 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.3.5 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.3.4 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.3.3 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.3.2 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.3.1 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.3.0 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.2.3 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.2.2 depends on torch>=1.0.0
vizseq 0.1.9 depends on torch
bert-score 0.1.2 depends on torch>=0.4.1
vizseq 0.1.9 depends on torch
bert-score 0.1.1 depends on torch>=0.4.1
vizseq 0.1.9 depends on torch
bert-score 0.1.0 depends on torch>=0.4.1
To fix this you could try to:
ERROR: ResolutionImpossible: for help visit https://pip.pypa.io/en/latest/user_guide/#fixing-conflicting-dependencies
[jw@cn05 ~]$
I am installing vizseq in a colab notebook and got an error:
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchaudio 2.3.0+cu121 requires torch==2.3.0, but you have torch 1.13.1 which is incompatible.
torchtext 0.18.0 requires torch>=2.3.0, but you have torch 1.13.1 which is incompatible.
torchvision 0.18.0+cu121 requires torch==2.3.0, but you have torch 1.13.1 which is incompatible.
How to solve this problem?
Thanks,
Hi :)
The tool returns a BLEU score for the machine translation and runs great in general, but I am not sure if the BLEU score represents the sentence level or corpus level? I haven't been able to gather anything conclusive from the sacreBLEU implementation, so I am hoping you can help me with this :)
Best regards,
Tobias
I just followed the installation steps and got this error.
** Minimal Code/Config snippet to reproduce **
** Stack trace/error message **
(base) diegomoussallem@Diegos-MBP examples % python -m vizseq.server --port 9001 --data-root examples/data
Traceback (most recent call last):
File "/opt/anaconda3/lib/python3.7/runpy.py", line 183, in _run_module_as_main
mod_name, mod_spec, code = _get_module_details(mod_name, _Error)
File "/opt/anaconda3/lib/python3.7/runpy.py", line 109, in _get_module_details
__import__(pkg_name)
File "/Users/diegomoussallem/Desktop/vizseq/vizseq/__init__.py", line 15, in <module>
from vizseq.ipynb import *
File "/Users/diegomoussallem/Desktop/vizseq/vizseq/ipynb/__init__.py", line 8, in <module>
from .core import (view_examples, view_n_grams, view_stats, view_scores,
File "/Users/diegomoussallem/Desktop/vizseq/vizseq/ipynb/core.py", line 15, in <module>
from vizseq._data import (VizSeqDataSources, PathOrPathsOrDictOfStrList,
File "/Users/diegomoussallem/Desktop/vizseq/vizseq/_data/__init__.py", line 14, in <module>
from .config_manager import VizSeqTaskConfigManager, VizSeqGlobalConfigManager
File "/Users/diegomoussallem/Desktop/vizseq/vizseq/_data/config_manager.py", line 13, in <module>
from .tokenizers import VizSeqTokenization
File "/Users/diegomoussallem/Desktop/vizseq/vizseq/_data/tokenizers.py", line 10, in <module>
from sacrebleu import tokenize_13a, tokenize_v14_international, tokenize_zh
ImportError: cannot import name 'tokenize_13a' from 'sacrebleu' (/opt/anaconda3/lib/python3.7/site-packages/sacrebleu/__init__.py)
Add any other context about the problem here.
When running the following for scores like Rouge and others like:
vizseq.view_scores(ref, hypo, ['metric that's not bleu'], tags=tag)
I am getting:
~/.local/lib/python3.6/site-packages/vizseq/scorers/__init__.py in _score_multiprocess_averaged(self, hypothesis, references, tags, sent_score_func)
170 for t in tag_set:
171 indices = [i for i, cur in enumerate(tags) if t in cur]
--> 172 group_scores[t] = np.mean([sent_scores[i] for i in indices])
173
174 return VizSeqScore.make(
~/.local/lib/python3.6/site-packages/vizseq/scorers/__init__.py in <listcomp>(.0)
170 for t in tag_set:
171 indices = [i for i, cur in enumerate(tags) if t in cur]
--> 172 group_scores[t] = np.mean([sent_scores[i] for i in indices])
173
174 return VizSeqScore.make(
TypeError: 'NoneType' object is not subscriptable
Any idea why that's going on? Using text data this is at least 3 tokens or more.
This works fine when running view_examples
.
In the official scorer example from https://facebookresearch.github.io/vizseq/docs/getting_started/scorer_example/, the second block confuse me.
Corpus-level BLEU: 67.945
Sentence-level BLEU: [75.984, 61.479]
Group BLEU: {'Test Group 2': 75.984, 'Test Group 1': 75.984}
I can see two generated sentences with corresponding reference sentences in the first block.
ref = [['This is a sample #1 reference.', 'This is a sample #2 reference.']]
hypo = ['This is a sample #1 prediction.', 'This is a sample #2 model prediction.']
tags = [['Test Group 1', 'Test Group 2']]
scores = scorer.score(hypo, ref, tags=tags)
print(f'Corpus-level BLEU: {scores.corpus_score}')
print(f'Sentence-level BLEU: {scores.sent_scores}')
print(f'Group BLEU: {scores.group_scores}')
The first sample belongs to Test Group 1
and the second sample belongs to Test Group 2
. If I'm not misunderstanding the use of the tag
, according to the Sentence-level BLEU
,the Group BLEU
should be {'Test Group 2': 61.479, 'Test Group 1': 75.984}
.
But the execution result is Group BLEU: {'Test Group 2': 75.984, 'Test Group 1': 75.984}
When trying to run the webapp with the example data, I have this error :
AttributeError: 'VizSeqLogger' object has no attribute 'set_console_mode'
Follow README instructions : download example data and run :
python -m vizseq.server --port 9001 --data-root ./examples/data
Stack trace/error message
Traceback (most recent call last):
File "/usr/lib/python3.6/runpy.py", line 193, in _run_module_as_main
"__main__", mod_spec)
File "/usr/lib/python3.6/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/home/me/workspace/vizseq/vizseq/server.py", line 14, in <module>
logger.set_console_mode(enable=True)
AttributeError: 'VizSeqLogger' object has no attribute 'set_console_mode'
The code run normally.
When trying to run the webapp with the example data, I have this error :
Uncaught exception GET
Follow README instructions : download example data and run :
python -m vizseq.server --port 9001 --data-root ./examples/data
The server starts fine, but when accessing the webapp at localhost:9001
, I can only see 500: Internal Server Error
.
Stack trace/error message
INFO - 11/04/19 10:36:39 - 0:00:00 - Application Started
You can navigate to http://localhost:9001
ERROR - 11/04/19 10:36:42 - 0:00:03 - Uncaught exception GET / (192.168.0.30)
HTTPServerRequest(protocol='http', host='192.168.0.231:9001', method='GET', uri='/', version='HTTP/1.1', remote_ip='192.168.0.30')
ERROR - 11/04/19 10:36:42 - 0:00:03 - 500 GET / (192.168.0.30) 1.18ms
The webapp run normally.
Hi there,
I've installed vizseq as described in the README and then tried to import Tokenizer13a. Unfortunately I got
ImportError: cannot import name 'Tokenizer13a'
error.
git clone https://github.com/facebookresearch/vizseq.git
cd vizseq
pip install --editable ./
> Successfully installed sacrebleu-2.0.0 vizseq-0.1.15
git clone https://github.com/facebookresearch/vizseq.git
cd vizseq
pip install --editable ./
python
>>> from sacrebleu.tokenizers import Tokenizer13a
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ImportError: cannot import name 'Tokenizer13a'
** Minimal Code/Config snippet to reproduce **
** Stack trace/error message **
>>> from sacrebleu.tokenizers import Tokenizer13a
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ImportError: cannot import name 'Tokenizer13a'
Add any other context about the problem here.
Hi
This repo is looking really cool and I would like to use the features it offers for my thesis :)
What are the plans/timeline for implementing win 10 support?
Best regards,
Tobias
A declarative, efficient, and flexible JavaScript library for building user interfaces.
๐ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. ๐๐๐
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google โค๏ธ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.