I spent many long nights preparing this material for a visual introduction to optimization in Deep Learning, ranging from 1st-order methods, 2nd-order and Natural Gradient (and approximations of it such as K-FAC). Sharing the PDF (easier to download):
Ministry of Health in Brazil is adopting a stupid visual strategy: reducing the font size of COVID-19 deaths and increasing the font size of recovered cases. Every day there is a new visual strategy to disguise the death numbers.
99% of recent models:
We trained this new cool model that is chatgpt-like!
Reality:
1) Meta trained it, you don't have license for it;
2) The dataset you used is distilled from OpenAI;
3) You did a fine-tune of the model, you don't have to change its name because of this;
4)
I'm working on a ML forecast system for flooding in South of Brazil, the blue mask shows the simulated flooding, which is matching surprisingly well the regions that actually flooded (the satellite image), next step is to train the stream gauge prediction model with precipitation
Just sharing ~100 slides about PyTorch 2 internals focusing on recent innovations (Dynamo, Inductor, and ExecuTorch).
I had a lot of fun preparing this and hope you'll enjoy it. I'm planning to record it soon.
PDF:
Slideshare:
This is the most terrifying image I have seen, I lived in this city for around 8 years in Brazil, this is a true color image from Sentinel-2 satellite showing today and a few days ago, it shows the flooding basically on an entire part of the city. This *never* happened before.
Everyone is talking about the GPT-2, but nobody asked the model if he wants to be released. Since
@Thom_Wolf
just released the interface on pytorch_pretrained_bert, I decided to ask the model about what he thinks.
Just decided to make a thread to show how many decisions and pre-processing steps you need to train large language models (LLMs) such as LLaMA. This is mainly based on CommonCrawl (CC) dataset and the pipeline used in LLaMA to generate its dataset. 1/n
Social irresponsibility: even after trying to make the author remove the tweet,
@bollembd
denied. Now I would like to share the answer from the original authors of the study DEBUNKING the statement made by
@bollemdb
. Be responsible and READ studies the way they should be folks.
Para os brasileiros interessados, estamos criando um fórum de Machine Learning em português para trocar ideias, mostrar projetos e auxiliar quem precisa na área de Data Science/Machine Learning/Deep Learning no Brasil. O site é: todos são bem vindos 🤗
I just posted the slides from the yesterday night presentation "PyTorch under the hood" at the
@PyDataMTL
in Montreal: thank you all for coming and thanks
@alexkimxyz
and
@mariaKhalusova
for organizing this amazing event.
Predicted daily deaths by COVID-19 for Brazil tomorrow: 782 (between 455-907 w/ HPD 90%). Many cities in Brazil seem to be walking into an imminent lockdown. Stay home, have no regrets.
Just released the v0.2 of MedicalTorch, a PyTorch framework for medical imaging. Available in the comfort if your python pip store. Changelog and docs at: , thanks to all contributors of this release ! 🤘
I wasn't able to update the R(t) estimates for Brazil today because the Ministry of Health in Brazil changed the format of the file from CSV to EXCEL (yes, you read it well) and didn't provide historical data anymore. I think it is a mix of incompetence with a sordid purpose.
Here is my theory: Google will deploy Gemini Nano in edge devices with a 1B model using an internal version of TensorFlow Lite focused on LLMs. They will ship updates or fine-tuned models (for specific tasks) using LoRA. I think they call that model internally as "ULM-1B". 1/n
Just release the first version of Episuite, an open-source suite of tools for epidemiology in Python. Documentation at: . This is mainly a collection of tools I've been using for some analysis for COVID. Stay tuned, more docs and features to come 🙃
Just updated the plots with hospitalizations, race, age and outcomes for the outbreak in Porto Alegre/RS/Brazil.
Hard not to see the social differences:
Really nice initiative from NHS in the UK: you can have a pack of 7 rapid antigen tests delivered to your home, all without any cost (I mean, you pay taxes lol), but it is really handy and you can test yourself every 2 weeks or so.
I just made available for download a CSV with R(t) estimates for each date and for each state in Brazil at: . There are also plots for Google mobility data and Facebook symptom surveys. This is daily updated after the updates from the Ministry of Health.
[pt-br] Como o ninguém (impressionantemente nem o governo) está postando as imagens de hoje (11 de Maio) de satélite (provavelmente porque eles acham que o céu está todo encoberto?), estou postando estas imagens impressionantes () do Sentinel-1 que utiliza
These are COVID-19 confirmed cases in ICUs at the south of Brasil (Rio Grande do Sul). Government adopted a method that they called "controlled distancing" (free translation) that they said to be "prioritizing life". I call this method "shit is happening under our nose".
This is very important, and to add into it, one pattern that I adopted on PyTorch was to create custom transforms that are added at the end of the pre-processing pipeline, to check for data, format, min and max values, custom statistics, etc. It helps a lot to catch issues.
The unambiguously correct place to examine your training data is immediately before it feeds into the network. Take the raw x,y batch tuple, ship it back to CPU, unrender, visualize. V often catches bugs with data augmentation, label preprocessing, samplers, collation, etcetc.
"SoftTriple Loss: Deep Metric Learning Without Triplet Sampling" (), -> "minimizing normalized SoftMax loss with the smooth term λ is equivalent to optimizing a smoothed triplet loss.", a very useful perspective on softmax, no more cumbersome sampling.
Brazil is a time bomb, all elements for the tragedy are there: a president that tells everyone not to isolate, lack of testing capacity, widespread of misinformation, the feeling that the country is different (weather, density, etc) than others and poverty.
If there is a talk you should watch from NeurIPS, it is the
@EmtiyazKhan
talk on Deep Learning with Bayesian principles ().
@EmtiyazKhan
and his co-authors build a beautiful bridge from Bayesian world to Deep Learning and high-order methods.
The fact that now in Brazil we have a fragmentation of data sources for daily COVID deaths and cases, including the government itself, third parties or independent media counts, is a clear win of the government, where the average person now says "I don't trust any of them". 1/n
I made an animation of the bayesian ICU occupancy forecast for Portugal since March 20th. Note how uncertainty behaves when new evidence is incorporated into the modelling.
These are the libs being shipped no in AICore:
libfile_defrag_jni.so
libtartarus_core.so
libtflite_llm_jni.so
libulm1b_ggml_jni.so
These are some interesting source filenames in the libtflite_llm_jni library:
This doesn't make much sense. It is not that LLaMA 30B only needs 4GB of memory, you're comparing loading it completely into memory and userspace buffers vs loading it on demand through page faults into the page cache. How is this comparison fair ?
This is why CS fundamentals continue to be crucial: LLaMA 30B only needs 4gb of memory if we use mmap().
Not sure why this works but one reason could be that 30B weights are sparse. Thus, lazy loading the fraction of needed weights reduces memory usage.
Interesting plot from "A Survey of Large Language Models" (). Splits look the same, but make no mistake, LLaMA did a very wise decision into dropping from CC what wasn't used as reference, that certainly improved the quality by a large margin.
We have a lot of misinformation in Brazil, but we have good scientists working to bring information to general audience. Unfortunately, many threads often get lost. I'm working on a Q&A engine (using NLP sentence encoders) on tweets from Brazilian scientists. Still in progress.
Cool work from ICML 2022: "A General Recipe for Likelihood-free Bayesian Optimization", by Jiaming Song and co-authors. . I wasn't even aware that classifiers were giving such good results when compared to GPs for BO.
Just published an analysis of the delay between symptom onset and confirmation for COVID-19 in some states in Brazil. More information for the analysis at: .
#COVID19
#coronavirus
#covidbr
Take one minute today to say thanks to people who are working tirelessly in this fight against misinformation, intolerance and pseudoscience. They could've just ignored, but they decided to care and show that science prevails. 1/2
Impressive accomplishment:"... In our experiments we showed that DKS allows deep networks without skip connections or normalization layers to be trained at similar speeds to ResNets on Imagenet, assuming the use of K-FAC or Shampoo." - , by J Martens, et al
Social irresponsibility: even after trying to make the author remove the tweet,
@bollembd
denied. Now I would like to share the answer from the original authors of the study DEBUNKING the statement made by
@bollemdb
. Be responsible and READ studies the way they should be folks.
Just like many researchers that are working with COVID-19 data in Brazil, I won't be able to updated all the analysis and estimates for the estimation of R(t) after this *data censorship* from the government. No more incidence data for states, nothing. We're all in the dark now.
Predicted cases for COVID-19 in Brazil by the previous model was between 236 and 290, the confirmed cases today was 290. Below, updated bayesian forecast model for next days and an estimate of the growth factor and the early basic reproductive number.
@OsmarTerra
O mais interessante é que o Canadá, EUA, Europa e a OMS discordam do Osmar Terra, talvez porque não é o Osmar que está tendo que empilhar corpos para ganhar a imunidade de rebanho.
22 entities in Porto Alegre/Brazil wants to open: shopping malls, all stores, bars and restaurants. So what is the plan now ? Open to close next week ? This is so dumb, that will make them lose even more money in the following weeks, there is *now* 90% occupancy in ICUs. 1/4
GPT-4 vs Google Bard, the question:
"Write a function in Python that will draw the Google logo using only the turtle module."
The one w/ colored circles is GPT-4.
This is a plot showing the symptom onset vs confirmation date for COVID-19 cases in the south of Brazil (Rio Grande do Sul). More than 450 cases with more than 30 days of delay between symptom onset and confirmation. This is situation that we have in many states of Brazil.
I made an artistic work for the people using COVID-19 ICU occupancy as a metric in Porto Alegre/RS to say that they are optimistic. Top panel: mobility for top cities exporting patients to Porto Alegre. Bottom panel: confirmed COVID-19 ICU occupancy in Porto Alegre.
"Eu trabalho com dados, não com projeções matemáticas" - Osmar Terra. Em seguida mostrou um gráfico sem eixos com anotação em canetinha para se explicar quanto à afirmação que fez de que morreriam no total 800 pessoas de covid no Brasil. Sem vergonha.
I found it *incredible* that there are *nothing*, not a single journalist, reporting that coincidentally, on the same day of the football match in Rio Grande do Sul (grenal), we had the *worse mobility* median value since the day the outbreak started. Am I alone in this ?
South of Brazil, ICUs growing out of control due to COVID-19. The 3rd city in the ranking of confirmed cases. Yes, we had a football match. And yes, we had a lot of people gathering together to celebrate. Brazil on his finest moments. Fair play is only fair when you make money.
Save the date! Our first meetup will take place on February 25th.
We'll be joined by
@tarantulae
who's going to talk about PyTorch under the hood. Check out the details here:
Social irresponsibility: even after trying to make the author remove the tweet,
@bollembd
denied. Now I would like to share the answer from the original authors of the study DEBUNKING the statement made by
@bollemdb
. Be responsible and READ studies the way they should be folks.
Reports for COVID-19 today in Portugal: the virus continues the exponential growth, no signs of the effectiveness of the adopted measures. I'm also releasing the first R0 estimate in Portugal for the time being, it has mean R0=3.8 (1.7-6.1, HPD 94%).
Como um gaúcho com família no RS, é extremamente lamentável ver essa situação se encaminhando para um colapso do sistema de saúde. Mesmo depois de meses de trabalho estimando da melhor forma que pude (método em ) as incertezas ... 1/n
Just made a COVID-19 ICU occupation simulation for Porto Alegre/RS/Brazil using length of stay from confirmed patients (same as described here: ). I started from a occupation regime similar to what we have today, and replayed ... 1/4
Updated forecast for Portugal COVID-19 cases with data from yesterday night. Y-axis shows the highest posterior density (HPD) for 94% credibility interval. There was a change on the testing criteria yesterday as well, so let's follow.
They are also shipping together with tensorflow lite llm the ggml library, which is probably for CPU inference while tflite uses other delegates for edge tpu/google tensor. They are also using a libtartarus that I don't know much about (yet 😅). 2/n
I'll be giving the talk "Gradient-based optimization" with an introduction to optimization methods for Deep Learning (first-order, second-order and some approximations). It will be hosted online (in Portuguese) at the Machine Learning Porto Alegre Meetup ()
Updated analysis for Portugal using data released today. With new data we see a reduction of the growth coefficient and R0. The model is also putting the trend below the baseline with more evidence. It is known that there will be an increase of the testing capacity, ... 1/2
Inspired by this plot, I just made a couple for the 8th cities in RS/Brazil with the most number of cases. Interesting to note that in Passo Fundo there seem to have more cases in older age bins around the dates when a nursing home was found with infections.
Wow! Excellent heat map of the age distribution of new cases in Florida.
X-axis = time, y-axis = age, color = new cases.
The brightest part has a triangular shape, indicating that what started as growth in younger adults has expanded into older ages.
Credit:
@zorinaq
Remember the news about the UFRGS study telling about the COVID-19 "peak" in Porto Alegre/RS/Brazil. Prediction was to have 300 people in ICUs by the end of August. *Today* we have 316. When you don't take uncertainty into account, reality teaches you about it later.
Pesquisadores da UFRGS fazem estimativas de leitos para pico da pandemia em Porto Alegre. No final de agosto, a capital gaúcha terá cerca de 27 mil casos de covid-19, com 300 pessoas internadas em UTI e 200 necessitando de respiradores.
Ministry of Health in Brazil just released the new report today, and for our surprise there are *ZERO new cases* and *ZERO new deaths* reported from Rio Grande do Sul (RS). How is it possible that not a single journalist is talking about it ?
@celo_pri
@betepacheco_
@thassius
O motivo é para que os atendentes não tenham que esperar para pular para uma próxima chamada, o sistema liga para uma quantidade enorme de pessoas pra garantir que terá alguém que atendeu e está na linha pronto para o atendente, é realmente insuportável.
Here we go again, Ministry of Health released the new version of a dataset with national data, besides containing *personal data* from patients, it also contains *users and passwords* to access medical records of their COVID-19 PCR test results.
In the context of today's hackernews post about langchain by
@minimaxir
, and Sherjil being curious about what we use at :
1. As
@denisyarats
answered, we have never and will likely never use LangChain.
2. Reasons for it align with the HN post:
Updated models for SARS-CoV2 in Portugal. It is now very difficult to do analysis with the lack of transparency from government, especially on the testing procedures that seems to be changing (increasing) and some potential under-reporting in the past days.
ICUs in Porto Alegre/Brazil today. Do you remember when UFPel congratulated the population for the social distancing ? Or when the state was bragging that the outbreak was under control ? When EPICOVID19 showed zero positive cases on the last survey in Porto Alegre ? I do.
Just released EuclidesDB, a multi-model machine learning feature database server entirely written in C++ and tightly integrated with PyTorch. Documentation available at:
New blog post: "Visualizing network ensembles with bootstrap and randomized priors": , the animation below is the training of a 40-model ensemble with bootstrap and randomized priors.
This is COVID-10 incidence data, by date of symptom onset for Canoas/RS. When Osmar Terra (a well-known Brazilian comedian) sees this, he often says: look, we're past the peak. 1/n
Outra comparação usando SWIR do Sentinel-2, para se ter uma ideia da dimensão, a distância entre um ponto extremo ao outro da imagem é a mesma distância de Londres a Paris, ~300km.
This is a photo I took in NeurIPS 2018 in Montreal where Jakob Uszkoreit is presenting The Transformer and talking about self-attention, there were like 6 or fewer folks there watching. This was the only year NeurIPS wasn't giving mugs because they just had changed from NIPS.
As predicted yesterday,
@DGSaude
now confirmed 59 cases (the exactly posterior mean predicted by the forecast) for March 10th. I'll update the model with the new data tonight.
Updated forecast for Portugal COVID-19 cases with data from yesterday night. Y-axis shows the highest posterior density (HPD) for 94% credibility interval. There was a change on the testing criteria yesterday as well, so let's follow.
Daily update for the outbreak in Brazil:
▶️ Confirmed cases: 1.800.827 ⬆️ (+45.084)
▶️ Deaths: 70.398 ⬆️ (+1.214)
Downloadable updated R(t) per state, mobility data and ongoing symptom survey available at:
A lot of people criticized me for not re-fitting the model in the last days, well my model from Mar 13 predicted 642 confirmed for today, and guess what was confirmed today: 642. So again, I'll leave that thread here: , the point is not model fitting.
This is an estimation of deaths for COVID-19 in Brazil and with US data (shifted) plotted together (red). The model fit with data only from Brazil predicts almost exactly the deaths that happened on US. Brazil is on the same dangerous path as US, but without test kits.
#covidbr
Just updated the SARS-CoV-2 R(t) estimates, mobility data and symptom survey results for Brazil. Brazil approaching 1 million infections, on a period where many states started to relax interventions.
More information at:
This is the delay distribution of the data from SES-RS (south of Brazil). They are constantly adding patients that had their symptom onset like more than 60 days ago. It never ceases to amaze me how terrible is this whole process and information system they have. 1/2
Seeing some interesting cases of ML folks leaving big tech to start their own startups and realizing that they never implemented anything for ML training infra/platform before, and failing miserably at doing it because it is easy to take it for granted in big tech comfort zone.
I don't know why folks are worried about LLMs escaping to the world while you can just put them inside a docker container inside of another docker container, if they escape, they will never realize they are inside of yet another docker container. Follow me for more safety tips.
Given the surge of COVID-19 in ICUs at RS (south of Brazil), I added a new analysis of the R(t), by symptom onset and adjusting for right-censoring. I added some markers based on some important events during the course of the outbreak.
▶️ More info at:
For all states in Brazil: we really hope that you will follow the data sharing procedures adopted by
@SES_RS
that is sharing their data on a structured format with symptom dates, confirmation dates and incidence for cases and deaths. Congress need to create an independent system.
Brazilian government spent 3 weeks developing this stupid website () that contains nearly ZERO information about COVID tracking, on a country with a population of 210M people. No structured data, nothing. Media is also hiding the data for themselves.
Alegrete in the south of Brazil decided to do a lockdown on the weekend, from Friday night up to Monday morning. They were able to achieve mobility levels *never seen* since the start of the outbreak in the city for two consecutive days, quite an accomplishment from everyone.
AIs can serve us as tools, but eventually, when they are sufficiently advanced, it may become immoral to keep them subservient. What is a practical criterion for deciding when an AI should be set free?