The Language Interpretability Tool (LIT): Interactive Exploration and Analysis of NLP Models

• With these challenges in mind, we built and open-sourced the Language Interpretability Tool (LIT), an interactive platform for NLP model understanding. LIT builds upon the lessons learned from the What-If Tool with greatly expanded capabilities, which cover a wide range of NLP tasks including sequence generation, span labeling, classification and regression, along with customizable and extensible visualizations and model analysis. (GitHub)

From the Washington Post this morning

Book

• Read and annotate Michelle’s outline, and add something about attention. That’s also the core of my response to Antonio
• More cults
• 2:00 Meeting
• Thinking about how design must address American Gnosticism, and the danger and opportunities of online “research”, and also how things like maps and diversity injection can potentially make profound impacts

GOES

• Update test code to use least squares/quaternion technique
• look into celluloid package for animating pyplot
• 3:00 Meeting

GPT-2 Agents

• I’m ingesting data from phase 12!
• Generate an Arabic dataset with less meta info for sim to run as a baseline
• Try to get embeddings from GPT? This might work:
  • github.com/huggingface/transformers/issues/1458
  • I think this is working?

import tensorflow as tf
# pip install git+https://github.com/huggingface/transformers.git
from transformers import TFGPT2LMHeadModel, GPT2Tokenizer

# options are (from https://huggingface.co/transformers/pretrained_models.html)
# 'gpt2' : 12-layer, 768-hidden, 12-heads, 117M parameters. # OpenAI GPT-2 English model
# 'gpt2-medium' : 24-layer, 1024-hidden, 16-heads, 345M parameters. # OpenAI’s Medium-sized GPT-2 English model
# 'gpt2-large' : 36-layer, 1280-hidden, 20-heads, 774M parameters. # OpenAI’s Large-sized GPT-2 English model
# 'gpt2-xl' : 48-layer, 1600-hidden, 25-heads, 1558M parameters.. # OpenAI’s XL-sized GPT-2 English model
tokenizer = GPT2Tokenizer.from_pretrained("../models/chess_model")

# add the EOS token as PAD token to avoid warnings
# model = TFGPT2LMHeadModel.from_pretrained("../models/gpt2-medium", pad_token_id=tokenizer.eos_token_id)
model = TFGPT2LMHeadModel.from_pretrained("../models/chess_model", pad_token_id=tokenizer.eos_token_id, from_pt=True)

wte = model.transformer.wte
wpe = model.transformer.wpe
word_embeddings:tf.Variable = wte.weight  # Word Token Embeddings
print("\nword_embeddings.shape = {}".format(word_embeddings.shape))
terms = ['black', 'white', 'king', 'queen', 'rook', 'bishop', 'knight', 'pawn']
for term in terms:
    text_index_list = tokenizer.encode(term)
    print("\nindex for {} = {}".format(term, text_index_list))
    for ti in text_index_list:
        vec = word_embeddings[ti]
        print("{}[{}] = {}...{}".format(term, ti, vec[:3], vec[-3:]))

• It gives the following results:

word_embeddings.shape = (50257, 768)
 index for black = [13424]
 black[13424] = [ 0.1466832  -0.03205131  0.13073246]…[ 0.03556942  0.2691384  -0.15679955]
 index for white = [11186]
 white[11186] = [ 0.01213744 -0.08717686  0.09657521]…[-0.01646501  0.05803612 -0.14158668]
 index for king = [3364]
 king[3364] = [ 0.07679952 -0.36437798  0.04769149]…[-0.2532825   0.11794613 -0.22853516]
 index for queen = [4188, 268]
 queen[4188] = [ 0.01280739 -0.12996083  0.10692213]…[0.03401601 0.01343785 0.30656403]
 queen[268] = [-0.17423214 -0.14485645  0.04941033]…[-0.16350408 -0.10608979 -0.03318951]
 index for rook = [305, 482]
 rook[305] = [ 0.08708595 -0.13490516  0.17987011]…[-0.17060643  0.07456072  0.04632077]
 rook[482] = [-0.07434712 -0.01915449  0.04398194]…[ 0.02418434 -0.06441653  0.26534158]
 index for bishop = [27832]
 bishop[27832] = [-0.05137009 -0.11024677  0.0080909 ]…[-0.02372078  0.00158158 -0.08555448]
 index for knight = [74, 3847]
 knight[74] = [ 0.10828184 -0.20851855  0.2618368 ]…[0.10234124 0.1747297  0.15052234]
 knight[3847] = [-0.15940899 -0.14975397  0.13490209]…[ 0.01935775  0.056772   -0.08009521]
 index for pawn = [79, 3832]
 pawn[79] = [-0.02358418 -0.18336709  0.08343078]…[ 0.23536623  0.06735501 -0.13106383]
 pawn[3832] = [0.12719391 0.05303555 0.12345099]…[-0.15112995  0.14558738 -0.05049708]

• Which are stable from run to run. Note that there can be multiple embeddings for the same term
• Monday, I’ll try to project into 2D using scikit-learn’s manifold learning

GPT-2 Agents

• Looks like we are getting close to ingesting all the new data
• Had a meeting with Ashwag last night (Note – we need to move the time), and the lack of ‘story-ness’ in the training set is really coming out in the model. The meta information works perfectly, but it’s wrapped around stochastic tweets, since there is no threading. I think there needs to be some topic structure in the meta information that allows similar topics to be grouped sequentially in the training set.
• 3:30 Meeting

GOES

• 9:30 meeting
• Update code with new limits on how small a step can be. Done, but I’m still having normal problems. It could be because I’m normalizing the contributions?

• Switching to a least-squares approach done?!

transform = lambda x: unpad(np.dot(pad(x), A))

print("\nTarget:")
print(secondary)
print("\nResult:")
print(transform(primary))
print("\nMax error: \n{}".format(np.abs(secondary - transform(primary)).max()))
print("\nA = \n{}".format(A))

Ap = A[:3, :3]
print("\nrotation matrix = \n{}".format(Ap))


print("getting quaternion")
q = Quaternion(matrix=Ap)
print("got quaternion")
print("Axis = {}".format(q.get_axis()))
print("Degrees = {}".format(q.degrees))

• Understanding BERT — Word Embeddings

Book

• More cults. Tying together Jonestown and Moby-Dick seems to be working better that what I was doing before

GPT-2 Agents

• Good meeting last night. The next action items are to
  • Figure out how to build embeddings from generated tweets
  • Finish updating database
  • Create new training set
  • Train small and medium models using current and compact meta frame (something like [[month, location, retweets]], or some similar easy regex)
• Need to register for Emerging Techniques forum: mors.org/Events/Workshops/Emerging-Techniques-Forum

GOES

• Need to fix the angle rollover in vehicle (and reference?) frames. I don’t think that it will fix anything though. I just don’t get why the satellite drifts after 70-ish degrees:

• There is something not right in the normal calculation?

• I think the problem is going to be here. Need to dig into this in a scratch file:

def get_angle(self, v1:Tuple, v2:Tuple) -> [float, float]:
    dot = np.dot(v1, v2)
    v1norm = np.linalg.norm(np.array(v1))
    v2norm = np.linalg.norm(np.array(v2))
    v1n_v2n = v1norm*v2norm
    normdot = np.divide(dot, v1n_v2n)
    # r_angle = np.arccos(np.clip(dot, -1.0, 1.0))
    r_angle = np.arccos(normdot)
    d_angle = math.degrees(r_angle)
    # print("rads = {}, degrees = {}".format(r_angle, d_angle))
    return r_angle, d_angle

• The small angle steps seem to also be causing a problem:

• Need to finish my telecommute form

Book

• Writing cults section.

GOES

• Got my panda3D text positioning working. I create a node and then attach the text to it. I think that’s not needed, that all you have to do is use render rather than the a2* options. Here’s how it works:

for n in self.tuple_list:
    ls:LineSegs = self.lp.get_LineSeg(n)
    node = TextNode(n)
    node.setText(n)
    tnp:NodePath = self.render.attachNewNode(node)
    #tnp.set_pos(-1, ypos, 1)
    tnp.setScale(0.1)
    self.text_node_dict[n] = tnp

• I then access the node paths through the dict

Book

• Spent a good chunk of the morning discussing the concept of dominance hierarchies and how they affect networks with Antonio
• Need to write some, dammit!

GOES

• My abstract has been accepted at the Military Operations Research Symposium’s (MORS) 4 day workshop in November!
• More Replayer. Working on text nodes. Done! It looks good, and is pointing out some REALLY ODD THINGS . I mean, the reaction wheel axis are not staying with the vehicle frame…

• 10:00 meeting with Vadim
• The (well, a) problem was that the reaction wheel vectors weren’t being reset each time, so the multiplies accumulated. Fixed! Now we have some additional problems, but these may be more manageable:

GPT-2 Agents

• Continuing to ingest tweets:

• 3:30 Meeting

How things are going

GPT-2 Agents

• Ingesting more data. It may take a while

GOES

• Continuing with Replayer
• Got the points moving based on the spreadsheet. I need to label them. It looks pretty straightforward to use 3D positions? I may have to use Billboards to keep things pointing at the camera

Book

• Look into Prof. Kristin Du Mez (Calvin College – @kkdumez)’s book (Jesus and John Wayne)?
• More writing
• Meeting with Michelle. Came up with an interesting tangent about programming/deprogramming wrt software development and cults

GPT-2 Agents

• Adding an optional split regex to parse_substrings. Here’s how I wound up doing it. This turns out to be slightly trickey, because the matches don’t include the text we’re interested in, so we have a step that splits out all the individual texts. We also throw away the text that leads to the first line and the last line of text since both can be assumed to be incomplete

split_regex = re.compile(r"(?=On [a-zA-Z]+ of [0-9]+,)")
split_iter = split_regex.finditer(tweet_str)
start = 0
tweet_list = []
for s in split_iter:
    t = tweet_str[start: s.start()]
    tweet_list.append(t)
    start = s.start()

• Shimei’s presentation went well!
• Work on translation

GOES

• Start on playback of the vehicle and reference frames

The Next Decade Could Be Even Worse

• A historian believes he has discovered iron laws that predict the rise and fall of societies. He has bad news.

GPT-2 Agents

• Tried Sim’s model, it’s very nice!
• Created a base class for creating and parsing tweets
• Found a regex that will find any text between two tokens. Thanks, stackoverflow!
• Here’s an example. I need to look into how large the meta information should be before it starts affecting the trajectory

On July of 2020, @MikenzieCromwell (screen name "Mikenzie Cromwell", 838 followers) posted a tweet from Boston, MA. They were using Twitter Web App. The post had 0 replies, 0 quotes, 1 retweets, and 3 favorites. "An example of the importance of the #mentalhealth community's response to #COVID19 is being featured in the @WorldBank survey. Check out the latest #MentalHealthResponse survey data on the state of mental health services in the wake of the pandemic. https://t.co/9qrq4G4XJi" 

GOES

• More Replayer
• Got the vertex manipulation! It’s hard to get at it though the geometry, but if you just save the LineSegs object,

ls = LineSegs(name)
self.prim_dict[name] = ls

• you can manipulate that directly

ls:LineSegs = self.lp.get_LineSeg("test_line")
ls.setVertex(1, x, y, z)

• Meeting with Vadim at 10:00. We found some pretty bad code that sets the torques on the reaction wheels

Book

• Write and send letters – done!
• More Moby Dick

I did something bad with my data yesterday. This is the correct version (I hope)

GPT-2 Agents

• Generating some results for Friday
• Splitting the results on the probes. It looks like the second tweet in a series is better formed. That kind of makes sense, because the second tweet is based on the first. That leads to an interesting idea. Maybe we should try building chains of text using the result from the previous
• Generating text with 1000 chars and parsing it, throwing away the first and last element in the list. I can also parse out the tweet, location, and sentiment:

[1]: In December of 2019, @svsvzz (21046 followers, 21784 following) posted a retweet, mentioning @ArticleSpot. It was sent from Saudi Arabia. @svsvzz wrote: "RT @ArticleSpot New update # Comment_study is coming..". The retweet was categorized as "Neutral". 
     Location = Saudi Arabia
     Sentiment = Neutral
     Tweet = RT @ArticleSpot New update # Comment_study is coming..
 [2]: In December of 2019, @HussainALhamad (2340 followers, 29 following) posted a retweet, mentioning @ejazah_ksa. It was sent from Riyadh, Saudi Arabia. @HussainALhamad wrote: "RT @ejazah_ksa Poll: Do you support #Suspension of studying in #Riyadh tomorrow, Monday? If you support (Retweet) If you do not support (Like)". The retweet was categorized as "Positive". 
     Location = Riyadh, Saudi Arabia
     Sentiment = Positive
     Tweet = RT @ejazah_ksa Poll: Do you support #Suspension of studying in #Riyadh tomorrow, Monday? If you support (Retweet) If you do not support (Like)
 [3]: In December of 2019, @mahfouz_nour (11 followers, 57 following) posted a tweet. She wrote: "♪ And the rest of the news about a news that the study was suspended in the study ♪ ♪ And God bless you ♪ ♪ Now ♪". The tweet was categorized as "Negative". 
     Location = None
     Sentiment = Negative
     Tweet = ♪ And the rest of the news about a news that the study was suspended in the study ♪ ♪ And God bless you ♪ ♪ Now ♪
 [4]: In December of 2019, @tansh99huda99 (1211 followers, 519 following) posted a retweet, mentioning @HashKSA. @tansh99huda99 wrote: "RT @HashKSA # comments on Monday at all schools for students in #Dahan, and the decision does not include teachers' and teachers' levels.". The retweet was categorized as "Neutral". 
     Location = None
     Sentiment = Neutral
     Tweet = RT @HashKSA # comments on Monday at all schools for students in #Dahan, and the decision does not include teachers' and teachers' levels.

Created some slides. I think they look pretty good:

America Is a Lot Sicker Than We Wanted to Believe

• Nearly half of the voters have seen Trump in all of his splendor—his infantile tirades, his disastrous and lethal policies, his contempt for democracy in all its forms—and they decided that they wanted more of it.

Added some code that makes it easier to compare countries and states and produced an animated GIF. I’m more concerned about Maryland now!

Book

• Letter to Stuart Kauffman
• Letter to Frans de Waal

GPT-2 Agents

• Created an animated GIF of English, Chinese, and Arabic countries for the Friday presentation
• DB-based topic text extraction
• 3:30 Meeting

GOES

• Work on Replayer

Went down to DC yesterday. So weird to see the White House behind multiple sets of walls, like a US base in Afghanistan

Dentist at 3:00

GPT-2 Agents

• Working on generating a new normalized data set. It needs to be mush smaller to get results by the end of the week. Done. It takes a couple of passes through the data to get totals needed for percentages, but it seems to be working well
• Restarted training
• Topic extraction from Tweet content

GOES

• Started working on 3D view of what’s going on with the two frames. I think I’m just going to have to start over with a a smaller codebase though, if Vadim can’t find what’s going on in his code.
• 1:30 Meeting with Vadim

I went on a long ride yesterday and the Trump signs that I’ve seen since 2015 are largely gone. The “Don’t tread on me” flags were still there. I wonder if it’s Trump fatigue?

GPT-2 Agents

• I had a brief flicker of a power failure that killed the training. I need to figure out how to restart from a checkpoint
• Need to normalize tweets by month and also subsample
• Topic extraction directly from DB

Took off yesterday and had a great time enjoying the weather and riding up to Gettysburg. And not thinking too much

There is also this

And later, this

GOES

• 10:00 Meeting with Vadim. If there is no progress, I’m going to build a display that plays back the recorded data so we can see what’s going on. He can’t make it, so postponed to Monday. I’ll start on the app after my other paperwork

NESDIS

• More paperwork

GPT-2 agents

• Respond to JF’s email and set up a meeting – done
• Hopefully finish the current training run.
• Adjust the code so that a query for the count of each month is made, and the results are normalized over each month. That means a query that gets the count and min/max row_ids
• Work on NMF and LDA topics for MER tweets. Do topics for the whole dataset (just the tweets), then select the ones that seem to make sense, and use them on the model-generated text. Create wordclouds from the topic labels
  • Got the NMF and LDA code into separate files. Need to change the get_content() method to pull directly from the db

Book

• Meeting with micelle

Did not want to wake up to this (Washington Post website as of 6:10 EST):

And the refrigerator was too narrow and too tall. Had to redo the cabinets. Sheesh.

GOES

• GVSETS continues today

GPT-2 Agents

• Working on the queries to build the test-train set. I’d like to start training against the medium model today
• I am doing things like this today: re.compile(r'(.?)(\<.?>)|$’)
• Created the training sets

Election day! Cue ominous music

Listening to Drinking with Historians episode with Drew McKevitt, and this came up:

..when when this came up with my students so my students they’re just an awesome group. I had this is two years ago i would say you know probably 18 of them probably 15 of them identified as like gun rights people right so they were they they owned guns themselves or they were all in favor of gun rights. And like every time the AR-15 came up and they’re just like “that’s this is just stupid culture war like that’s why people buy the gun.” These are like these are students who themselves are like gun rights people or gun owners and like that’s why people buy it it’s a culture war thing. I mean look at that guy in St Louis right who who pulled out his AR-15 and was aiming it at the protesters. Like what the hell is that guy doing with that gun? It’s total culture war stuff. It’s you know as we’d say like virtue signaling or something like that and that that’s what it is it’s saying. Like this it’s it’s it’s voting by buying a gun.”

https://www.youtube.com/watch?v=iAcKv03GfdE&list=PLvTW8ARBmW029eOiTQA8GSVBKmSGHXpIp&index=12

GOES

• GVSETS, starts at 9:00
• Agendas

Book

• Adding some content from Chimpanzee Politics and working with the Moby Dick section. I think that’s coming together now. Need to talk about how Starbuck continues to have his own thoughts, but goes along with the consensus of the ship, even though he suspects that this will kill him

GPT-2 Agents

• Meeting with Sim and Shimei. We’re going to focus on getting some preliminary results for the 13th
• Looked at the results for the medium model and all agree that it looks much better. Going to train the