Incoming links: Ubuntu 24.10
It trains the LeNet-5 neural network on the MNIST dataset from scratch, and afterwards you can give it newly hand-written digits 0 to 9 and it will hopefully recognize the digit for you.
Ciro Santilli created a small fork of this repo at lenet adding better automation for:
- extracting MNIST images as PNG
- ONNX CLI inference taking any image files as input
- a Python
tkinterGUI that lets you draw and see inference live - running on GPU
Install on Ubuntu 24.10 with:We use our own
sudo apt install protobuf-compiler
git clone https://github.com/activatedgeek/LeNet-5
cd LeNet-5
git checkout 95b55a838f9d90536fd3b303cede12cf8b5da47f
virtualenv -p python3 .venv
. .venv/bin/activate
pip install \
Pillow==6.2.0 \
numpy==1.24.2 \
onnx==1.13.1 \
torch==2.0.0 \
torchvision==0.15.1 \
visdom==0.2.4 \
;pip install because their requirements.txt uses >= instead of == making it random if things will work or not.On Ubuntu 22.10 it was instead:
pip install
Pillow==6.2.0 \
numpy==1.26.4 \
onnx==1.17.0 torch==2.6.0 \
torchvision==0.21.0 \
visdom==0.2.4 \
;Then run with:This script:
python run.pyIt throws a billion exceptions because we didn't start the Visdom server, but everything works nevertheless, we just don't get a visualization of the training.
The terminal outputs lines such as:
Train - Epoch 1, Batch: 0, Loss: 2.311587
Train - Epoch 1, Batch: 10, Loss: 2.067062
Train - Epoch 1, Batch: 20, Loss: 0.959845
...
Train - Epoch 1, Batch: 230, Loss: 0.071796
Test Avg. Loss: 0.000112, Accuracy: 0.967500
...
Train - Epoch 15, Batch: 230, Loss: 0.010040
Test Avg. Loss: 0.000038, Accuracy: 0.989300One of the benefits of the ONNX output is that we can nicely visualize the neural network on Netron:
Netron visualization of the activatedgeek/LeNet-5 ONNX output
. From this we can see the bifurcation on the computational graph as done in the code at:output = self.c1(img)
x = self.c2_1(output)
output = self.c2_2(output)
output += x
output = self.c3(output)Homepage: aihabitat.org/
The thing was definitely built by researchers. How to cite first, actually working later! And docs are just generally awkward.
This has happened a few times a day on Ubuntu 24.10 and Chromium 133. It has also been happening in previous versions of Ubuntu and Chromium.
As Ciro Santilli starts typing on the omnibox, sometimes the window freezes and the dreaded "is not responding" window shows up.
The only somewhat similar reports that Ciro Santilli could find as of 2025:
This section describes our attempts at compiling the Linux kernel for Ubuntu so as to use the exact patches and build configuration as used for a given Ubuntu release. The same toolchain would also be ideal, but perhaps this would require a Linux distribution buildable from source.
canonical-kteam-docs.readthedocs-hosted.com/en/public/how-to/build-kernel.html seems promising it says that for Ubuntu 24.04 and above you should do the following which was tested on Ubuntu 24.10:
sudo cp /etc/apt/sources.list /etc/apt/sources.list~
sudo sed -Ei 's/^# deb-src /deb-src /' /etc/apt/sources.list
sudo apt-get update
sudo apt build-dep -y linux linux-image-unsigned-$(uname -r)
sudo apt install -y fakeroot llvm libncurses-dev dwarves
apt source linux-image-unsigned-$(uname -r)
~/tmp/ubuntu/linux-6.11.0
cd linux-6.11.0
chmod a+x debian/rules
chmod a+x debian/scripts/*
chmod a+x debian/scripts/misc/*
fakeroot debian/rules clean
fakeroot debian/rules binaryThe build is extremely slow compared to a build of a more embedded and specifically targeted minimal kernel, and took about 2 hours on P14s. Their philosophy is likely to enable as many drivers as possible so that a single download will work for everyone. Which makes sense, fair enough. It would be cute though if there was a smarter way. Oh well.
linux-6.11.0/debian/build/build-genericOpenAI Gym development by OpenAI ceased in 2021, and the Farama Foundation not for profit took up maintenance of it.
gymnasium==1.1.1 just worked on Ubuntu 24.10 testing with the hello world gym/random_control.py:just works and opens a game window on my desktop.
sudo apt install swig
cd gym
virtualenv -p python3
. .venv/bin/activate
pip install -r requirements-python-3-12.txt
./random_control.py
Lunar Lander environment of Farama Gymnasium with random controls
. This example just passes random commands to the ship so don't expect wonders. The cool thing about it though is that you can open any environment with it e.g.
./random_control.py CarRacing-v3Manual control is extremely useful to get an intuition about the problem. You will notice immediately that controlling the ship is extremely difficult.
Lunar Lander environment of Farama Gymnasium with manual control
. We don't know if it is realistic, but what is certain is that this is definitely not designed to be a fun video game!A good strategy is to land anywhere very slowly and then inch yourself towards the landing pad.
- the legs of the lander are short and soft, and you're not supposed to hit the body on ground, so you have to go very slow
- the thrusters are quite weak and inertia management is super important
- the ground is very slippery
The documentation for it is available at: gymnasium.farama.org/environments/box2d/lunar_lander/ The agent input is described as:so it is a fundamentally flawed robot training example as global x and y coordinates are precisely known.
The state is an 8-dimensional vector: the coordinates of the lander in x & y, its linear velocities in x & y, its angle, its angular velocity, and two booleans that represent whether each leg is in contact with the ground or not.
Variation in the scenario comes from:
To actually play the games manually with keyboard, you need to define your own keybindings with gymnasium.utils.play.play. Feature request for default keybindings: github.com/Farama-Foundation/Gymnasium/discussions/1330
There is no C API, you have to go through Python: github.com/Farama-Foundation/Gymnasium/discussions/1181. Shame.
They have video recording support, minimal ex stackoverflow.com/questions/77042526/how-to-record-and-save-video-of-gym-environment/79514542#79514542
This is a small fork of activatedgeek/LeNet-5 by Ciro Santilli adding better integration and automation for:
Install on Ubuntu 24.10:
sudo apt install protobuf-compiler
cd lenet
virtualenv -p python3 .venv
. .venv/bin/activate
pip install -r requirements-python-3-12.txtDownload and extract MNIST train, test accuracy, and generate the ONNX Extract MNIST images as PNG:Infer some individual images using the ONNX:Draw on a GUI and see live inference using the ONNX:TODO: the following are missing for this to work:
lenet.onnx:./train.py./extract_pngs.py./infer.py data/MNIST/png/test/0/*.png./draw.py- start a background task. This we know how to do: stackoverflow.com/questions/1198262/tkinter-locks-python-when-an-icon-is-loaded-and-tk-mainloop-is-in-a-thread/79502287#79502287
- get bytes from the canvas: all methods are ugly: stackoverflow.com/questions/9886274/how-can-i-convert-canvas-content-to-an-image
Let's create some test data like this:
time psql tmp -c 'DROP TABLE IF EXISTS fts;'
time psql tmp -c 'CREATE TABLE fts(s TEXT, i INTEGER);'
time psql tmp <<'EOF'
INSERT INTO fts SELECT
i::text || ' ' ||
(i * 2 )::text || ' ' ||
(i * 5 )::text || ' ' ||
(i * 7 )::text || ' ' ||
(i * 11 )::text || ' ' ||
(i * 13 )::text || ' ' ||
(i * 17 )::text || ' ' ||
(i * 23 )::text || ' ' ||
(i * 29 )::text || ' ' ||
(i * 31 )::text
,
i % 100
FROM generate_series(1::bigint, 100000000::bigint) AS s(i);
EOFThe creation time was 2m13s, and the final size was:
table_name | pg_size_pretty | pg_total_relation_size
------------------+----------------+------------------------
fts | 13 GB | 14067326976This test data will be simple to predict what each line contains so we can make educated queries, while also posing some difficulty to the RDMS. As per:the first columns look like:
time psql tmp -c 'SELECT * FROM fts LIMIT 10;' s | i
-------------------------------------+----
1 2 5 7 11 13 17 23 29 31 | 1
2 4 10 14 22 26 34 46 58 62 | 2
3 6 15 21 33 39 51 69 87 93 | 3
4 8 20 28 44 52 68 92 116 124 | 4
5 10 25 35 55 65 85 115 145 155 | 5
6 12 30 42 66 78 102 138 174 186 | 6
7 14 35 49 77 91 119 161 203 217 | 7
8 16 40 56 88 104 136 184 232 248 | 8
9 18 45 63 99 117 153 207 261 279 | 9
10 20 50 70 110 130 170 230 290 310 | 10We aimed to create a test table of size around 10 GB, as in practice it is around that order of size that index speedups start to become very obvious on a SSD-based system.
Before we create the index, let's see if our non-indexed queries are slow enough for our tests:which gives:so it should be enough to observe the index speedup.
time psql tmp -c "SELECT * FROM fts WHERE s LIKE '% 50000000 %';" s | i
---------------------------------------------------------------------------------------------------+---
10000000 20000000 50000000 70000000 110000000 130000000 170000000 230000000 290000000 310000000 | 0
25000000 50000000 125000000 175000000 275000000 325000000 425000000 575000000 725000000 775000000 | 0
(2 rows)
real 0m11.758s
user 0m0.017s
sys 0m0.008sNow let's create the index. First we create a generated column that splits the strings with These commands took 8m51s and 40m8s and the DB size went up about 5x:
to_tsvector, and then we index that split column:time psql tmp <<'EOF'
ALTER TABLE fts ADD COLUMN s_ts tsvector
GENERATED ALWAYS AS (to_tsvector('english', s)) STORED;
EOF
time psql tmp -c 'CREATE INDEX s_ts_gin_idx ON fts USING GIN (s_ts);' table_name | pg_size_pretty | pg_total_relation_size
------------------+----------------+------------------------
fts | 69 GB | 74487758848And finally let's try out the index:which "instantly" gives us in 0m0.129s:so the index worked!
time psql tmp -c "SELECT s, i FROM fts WHERE s_ts @@ to_tsquery('english', '50000000');" s | i
-------------------------------------------------------------------------------------------------------+---
10000000 20000000 50000000 70000000 110000000 130000000 170000000 230000000 290000000 310000000 | 0
25000000 50000000 125000000 175000000 275000000 325000000 425000000 575000000 725000000 775000000 | 0
50000000 100000000 250000000 350000000 550000000 650000000 850000000 1150000000 1450000000 1550000000 | 0Another important use case is to search for prefixes of words, e.g. as you'd want in a simple autocompletion system. This can be achieved by adding This finishes in the same amount of time, and gives:so now we have cool hits such as
:* at the end of the search term as in:time psql tmp -c "SELECT s, i FROM fts WHERE s_ts @@ to_tsquery('english', '50000000:*');" s | i
-----------------------------------------------------------------------------------------------------------+----
10000000 20000000 50000000 70000000 110000000 130000000 170000000 230000000 290000000 310000000 | 0
38461539 76923078 192307695 269230773 423076929 500000007 653846163 884615397 1115384631 1192307709 | 39
45454546 90909092 227272730 318181822 500000006 590909098 772727282 1045454558 1318181834 1409090926 | 46
50000000 100000000 250000000 350000000 550000000 650000000 850000000 1150000000 1450000000 1550000000 | 0
71428572 142857144 357142860 500000004 785714292 928571436 1214285724 1642857156 2071428588 2214285732 | 72
100000000 200000000 500000000 700000000 1100000000 1300000000 1700000000 2300000000 2900000000 3100000000 | 0
29411765 58823530 147058825 205882355 323529415 382352945 500000005 676470595 852941185 911764715 | 65
25000000 50000000 125000000 175000000 275000000 325000000 425000000 575000000 725000000 775000000 | 0500000000, 500000004, 500000005, 500000007 and 500000006. The syntax is also mentioned at:Next we can also try some other queries with multiple terms. Text must contain two words with gives:
&:time psql tmp -c "SELECT s, i FROM fts WHERE s_ts @@ to_tsquery('english', '50000000 & 175000000');" s | i
-------------------------------------------------------------------------------------------------------+---
25000000 50000000 125000000 175000000 275000000 325000000 425000000 575000000 725000000 775000000 | 0Text can contain either word with gives:
|:time psql tmp -c "SELECT s, i FROM fts WHERE s_ts @@ to_tsquery('english', '50000000 | 175000000');" s | i
---------------------------------------------------------------------------------------------------------+---
10000000 20000000 50000000 70000000 110000000 130000000 170000000 230000000 290000000 310000000 | 0
50000000 100000000 250000000 350000000 550000000 650000000 850000000 1150000000 1450000000 1550000000 | 0
87500000 175000000 437500000 612500000 962500000 1137500000 1487500000 2012500000 2537500000 2712500000 | 0
25000000 50000000 125000000 175000000 275000000 325000000 425000000 575000000 725000000 775000000 | 0
35000000 70000000 175000000 245000000 385000000 455000000 595000000 805000000 1015000000 1085000000 | 0Text can contain the given words sequentially:gives:
time psql tmp -c "SELECT s, i FROM fts WHERE s_ts @@ to_tsquery('english', '50000000 <-> 125000000 <-> 175000000');" s | i
-------------------------------------------------------------------------------------------------------+---
25000000 50000000 125000000 175000000 275000000 325000000 425000000 575000000 725000000 775000000 | 0We can also inspect how words were split by simply doing a
SELECT * again: s | i | s_ts
----------------------------+---+----------------------------------------------------------------------
1 2 5 7 11 13 17 23 29 31 | 1 | '1':1 '11':5 '13':6 '17':7 '2':2 '23':8 '29':9 '31':10 '5':3 '7':4
2 4 10 14 22 26 34 46 58 62 | 2 | '10':3 '14':4 '2':1 '22':5 '26':6 '34':7 '4':2 '46':8 '58':9 '62':10
3 6 15 21 33 39 51 69 87 93 | 3 | '15':3 '21':4 '3':1 '33':5 '39':6 '51':7 '6':2 '69':8 '87':9 '93':10Let's check if the index updates automatically when we do an insert and if insertion seems to have been significantly slowed down by the index:finishes in:so performance is OK. Presumably, the insertion time is proportional to the number of tokens, doing one logarithmic operation per token, so indexing short chunks of text like titles is easy. And then let's find it:which finds it with:so we are all good. Unfortunately, accurate performance benchmarking is a bit harder than that, as the index by default first collects a certain number of updates into memory into the "pending list", before actually inserting them all at once after a certain mass is reached, as documented at: www.postgresql.org/docs/17/gin.html#GIN-IMPLEMENTATION. We are not going that deep today.
time psql tmp -c "INSERT INTO fts VALUES ('abcd efgh', 99)"real 0m0.043s
user 0m0.014s
sys 0m0.010stime psql tmp -c "SELECT s, i FROM fts WHERE s_ts @@ to_tsquery('english', 'efgh');" s | i
-----------+----
abcd efgh | 99The next thing that we need to understand is how gives:so we understand some of the heuristic normalizations:
to_tsvector tokenizes strings for the english language. For example running:psql -c "select to_tsvector('english', 'A Dog run runs fast faster two Cats: b c to from 1 é befhyph-afthyph.')"'1':13
'afthyph':17
'b':9
'befhyph':16
'befhyph-afthyph':15
'c':10
'cat':8
'dog':2
'fast':5
'faster':6
'run':3,4
'two':7
'é':14- prepositions like
toandfromare gone. These are called stopwords as documented at: www.postgresql.org/docs/17/textsearch-controls.html#TEXTSEARCH-PARSING-DOCUMENTS - words are lowercased and singularized, e.g.
Catsbecomescat - hyphenated words are stored both in separate components and in the full hyphenated form:
'afthyph':17'befhyph':16'befhyph-afthyph':15
The full list of languages available can be obtained with:On Ubuntu 24.10, the list contains major world languages, plus the special gives:so we understand that it is similar to
psql -c '\dF'simple configuration such that:psql -c "select to_tsvector('simple', 'A Dog run runs fast faster two Cats: b c to from 1 é befhyph-afthyph.')"'1':13
'a':1
'afthyph':17
'b':9
'befhyph':16
'befhyph-afthyph':15
'c':10
'cats':8
'dog':2
'fast':5
'faster':6
'from':12
'run':3
'runs':4
'to':11
'two':7
'é':14english but it does not:- seem to have any stopwords
- do singularization normalization
From the query side of things, if the query is going to be open to end users on a web interface, we need to understand giving:To avoid such errors, we can use:Bibliography:
to_tsquery better. The issue is that to_tsquery is quite brutal and happily throws errors for common things users might do e.g. spaces:select to_tsquery('english', 'abc def');ERROR: syntax error in tsquery: "abc def"plainto_tsquery: ANDs everythingwebsearch_to_tsquery: supportsANDwith spaces, OR withor, word negation with-wordand concatenation with"my word". But it unfortunately does not support prefixing, which is what everyone and their mother wants for autocomplete: stackoverflow.com/questions/14103880/escaping-special-characters-in-to-tsquery#comment78452351_41804957
- stackoverflow.com/questions/16020164/psqlexception-error-syntax-error-in-tsquery/16020565#16020565
- stackoverflow.com/questions/14103880/escaping-special-characters-in-to-tsquery
- www.reddit.com/r/rails/comments/1cmloa6/sanitizing_a_search_phrase_when_using_to_tsvector/
- stackoverflow.com/questions/6735881/to-tsquery-validation
- dba.stackexchange.com/questions/135030/filtering-special-characters-in-to-tsquery
Also posted at:
- www.reddit.com/r/PostgreSQL/comments/12yld1o/comment/m3l5nkv/ "Is it worth using Postgres' builtin full-text search or should I go straight to Elastic?", high top Google result for "PostgreSQL full text search" as of 2024. Random, but it's there.
Updates Post OurBigBook job search round 2025 Updated 2025-05-29 +Created 2025-05-07 2025-05-29
I shouldn't be doing this on funded OurBigBook time which is until the end of May, but I was getting too nervous and decided to start a casual job search to test the waters.
In particular I want to see if I can get past the HR lady step without toning down my online profiles. If nothing works out for the next round I'll be hiding anything too spicy like:Another interesting point is to see if French companies are more likely to reply given that Ciro Santilli studied at École Polytechnique which the French worship.
- prominently seeking funding for OurBigBook on my LinkedIn profile
- CIA 2010 covert communication websites references. This will be my first job hunt since I have published that article. Wish me luck.
- gay Putin profile picture on Stack Overflow
Gay Putin, currently used in Ciro Santilli's Stack Overflow profile
. Ciro's profiles may be a bit too much for the HR ladies who reject his job applications on the spot. To be fair, perhaps not enough years of experience for certain applications and job hopping may have something to do with it too. But since they don't ever tell you anything not to get sued, we'll never know.I'm looking in particular either for:
- machine learning-adjacent jobs in companies that seem to be doing something that could further AGI, e.g. automatic code generation or robotics would be ideal
- quantum computing
- systems programming, which is what I actually have work experience with
I spent the last two weeks doing that:
- one week browsing everything of interest in London and Paris and sending applications to anything that seemed both relevant and interesting. Maintaining an application list at: Section "Job application by Ciro Santilli".
- one week on a very laborious but somewhat interesting take home exercise for Linux kernel engineer a Canonical, makers of Ubuntu.I had a week to finish 5 practical coding and packaging questions, and I tried to do everything as perfectly as possible, but I somewhat underestimated the amount of work and wait needed to do everything and didn't manage to finish question 4 and missed 5. Oops let's see how that goes.At least this had a few good outcomes for the Internet as I tried to document things as nicely as I could where they were missing from Google as usual:
- I re-tested Linux Kernel Module Cheat and made some small improvements. Things still worked from a Ubuntu 24.10 host (using Docker to Ubuntu 22.04), and I also checked that kernel 6.8 builds and GDB step debugs after adding the newly required config
CONFIG_DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT, also mentioned that at: Why are there no debug symbols in my vmlinux when using gdb with /proc/kcore? - I contributed some simple updates to github.com/martinezjavier/ldd3 getting it closer to work on Linux kernel v6.8. That repository aims to keep the venerable examples from Linux kernel module book LDD3 alive on newer kernels, and is a very good source for kernel module developers.
- How to compile a Linux kernel module?: wrote a quick Ciro-approved tutorial
- Dynamic array in Linux kernel module: I gave an educational example of a dynamic byte array (like std::string) using the kvmalloc family of allocators
- quickemu: this is a good emulator manager and I think I'll be using it for Ubuntu images when needed from now on. I wrote:
- How to run Ubuntu desktop on QEMU?: an introductory tutorial to the software as their README is not that good as is often the case. It's hard for project authors to predict what new users want or not. This is my second answer to this question, the previous one focusing on a more manual approach without third party helpers.
- How to share folder between guest/host? (Quickemu): I explained how to setup a 9p mount to share a directory between guest and host
- Error :: You must put some 'source' URIs in your sources.list: updated this answer for Ubuntu 24.04. This issue comes up when you want to do either of:which don't work by default, and my answer explains how to do it from the GUI and CLI. The CLI method is specially important for Docker images. Since Ubuntu doesn't offer a stable CLI method for this, the method breaks from time to time and we have to find the new config file to edit.
sudo apt build-dep sudo apt source - What is hardware enablement (HWE)?: I learned a bit better how Ubuntu structures its kernel releases for each Ubuntu release
Some of the main issues I had were:- compiling Linux kernel for Ubuntu is extremely slow. I was used to compiling for embedded system with Buildroot, which finishes in minutes, but for Ubuntu is hours, presumably because they enable as many drivers as possible to make a single ISO work on as many different computers as possible, which makes sense, but also makes development harder
- my QEMU setup for Ubuntu was not quite as streamlined and I relearned a few things and set up quickemu. By chance I had recently come across quickemu for testing OurBigBook on MacOS, but I had to learn a bit how to set it up reasonably too
- I re-tested Linux Kernel Module Cheat and made some small improvements. Things still worked from a Ubuntu 24.10 host (using Docker to Ubuntu 22.04), and I also checked that kernel 6.8 builds and GDB step debugs after adding the newly required config
