An IBM made/pushed term, but that matches Ciro Santilli's general view of how we should move forward AGI.
Instructions at:
Ubuntu 22.10 setup with tiny dummy manually generated ImageNet and run on ONNX:
sudo apt install pybind11-dev
git clone https://github.com/mlcommons/inference
cd inference
git checkout v2.1
virtualenv -p python3 .venv
. .venv/bin/activate
pip install numpy==1.24.2 pycocotools==2.0.6 onnxruntime==1.14.1 opencv-python==4.7.0.72 torch==1.13.1
cd loadgen
CFLAGS="-std=c++14" python setup.py develop
cd -
cd vision/classification_and_detection
python setup.py develop
wget -q https://zenodo.org/record/3157894/files/mobilenet_v1_1.0_224.onnx
export MODEL_DIR="$(pwd)"
export EXTRA_OPS='--time 10 --max-latency 0.2'
tools/make_fake_imagenet.sh
DATA_DIR="$(pwd)/fake_imagenet" ./run_local.sh onnxruntime mobilenet cpu --accuracyLast line of output on P51, which appears to contain the benchmark resultswhere presumably
TestScenario.SingleStream qps=58.85, mean=0.0138, time=0.136, acc=62.500%, queries=8, tiles=50.0:0.0129,80.0:0.0137,90.0:0.0155,95.0:0.0171,99.0:0.0184,99.9:0.0187qps means queries per second, and is the main results we are interested in, the more the better.Running:produces a tiny ImageNet subset with 8 images under
tools/make_fake_imagenet.shfake_imagenet/.fake_imagenet/val_map.txt contains:val/800px-Porsche_991_silver_IAA.jpg 817
val/512px-Cacatua_moluccensis_-Cincinnati_Zoo-8a.jpg 89
val/800px-Sardinian_Warbler.jpg 13
val/800px-7weeks_old.JPG 207
val/800px-20180630_Tesla_Model_S_70D_2015_midnight_blue_left_front.jpg 817
val/800px-Welsh_Springer_Spaniel.jpg 156
val/800px-Jammlich_crop.jpg 233
val/782px-Pumiforme.JPG 285TODO prepare and test on the actual ImageNet validation set, README says:
Prepare the imagenet dataset to come.
Since that one is undocumented, let's try the COCO dataset instead, which uses COCO 2017 and is also a bit smaller. Note that his is not part of MLperf anymore since v2.1, only ImageNet and open images are used. But still:
wget https://zenodo.org/record/4735652/files/ssd_mobilenet_v1_coco_2018_01_28.onnx
DATA_DIR_BASE=/mnt/data/coco
export DATA_DIR="${DATADIR_BASE}/val2017-300"
mkdir -p "$DATA_DIR_BASE"
cd "$DATA_DIR_BASE"
wget http://images.cocodataset.org/zips/val2017.zip
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
unzip val2017.zip
unzip annotations_trainval2017.zip
mv annotations val2017
cd -
cd "$(git-toplevel)"
python tools/upscale_coco/upscale_coco.py --inputs "$DATA_DIR_BASE" --outputs "$DATA_DIR" --size 300 300 --format png
cd -Now:fails immediately with:The more plausible looking:first takes a while to preprocess something most likely, which it does only one, and then fails:
./run_local.sh onnxruntime mobilenet cpu --accuracyNo such file or directory: '/path/to/coco/val2017-300/val_map.txt./run_local.sh onnxruntime mobilenet cpu --accuracy --dataset coco-300Traceback (most recent call last):
File "/home/ciro/git/inference/vision/classification_and_detection/python/main.py", line 596, in <module>
main()
File "/home/ciro/git/inference/vision/classification_and_detection/python/main.py", line 468, in main
ds = wanted_dataset(data_path=args.dataset_path,
File "/home/ciro/git/inference/vision/classification_and_detection/python/coco.py", line 115, in __init__
self.label_list = np.array(self.label_list)
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 2 dimensions. The detected shape was (5000, 2) + inhomogeneous part.TODO!
Interesting layer skip architecture thing.
Apparently destroyed ImageNet 2015 and became very very famous as such.
- torchvision ResNet
- MLperf v2.1 ResNet contains a pre-trained ResNet ONNX at zenodo.org/record/4735647/files/resnet50_v1.onnx for its inference benchmark. We've tested it at: Run MLperf v2.1 ResNet on Imagenette.
catalog.ngc.nvidia.com/orgs/nvidia/resources/resnet_50_v1_5_for_pytorch explains:
The difference between v1 and v1.5 is that, in the bottleneck blocks which requires downsampling, v1 has stride = 2 in the first 1x1 convolution, whereas v1.5 has stride = 2 in the 3x3 convolution.This difference makes ResNet50 v1.5 slightly more accurate (~0.5% top1) than v1, but comes with a small performance drawback (~5% imgs/sec).
CNN convolution kernels are not hardcoded. They are learnt and optimized via backpropagation. You just specify their size! Example in PyTorch you'd do just:as used for example at: activatedgeek/LeNet-5.
nn.Conv2d(1, 6, kernel_size=(5, 5))This can also be inferred from: stackoverflow.com/questions/55594969/how-to-visualise-filters-in-a-cnn-with-pytorch where we see that the kernels are not perfectly regular as you'd expected from something hand coded.
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)
Note that:
- the images must be drawn with white on black. If you use black on white, it the accuracy becomes terrible. This is a good very example of brittleness in AI systems!
- images must be converted to 32x32 for
lenet.onnx, as that is what training was done on. The training step converted the 28x28 images to 32x32 as the first thing it does before training even starts
We can try the code adapted from thenewstack.io/tutorial-using-a-pre-trained-onnx-model-for-inferencing/ at lenet/infer.py:and it works pretty well! The program outputs:as desired.
cd lenet
cp ~/git/LeNet-5/lenet.onnx .
wget -O 9.png https://raw.githubusercontent.com/cirosantilli/media/master/Digit_9_hand_drawn_by_Ciro_Santilli_on_GIMP_with_mouse_white_on_black.png
./infer.py 9.png9We can also try with images directly from Extract MNIST images.and the accuracy is great as expected.
infer_mnist.py lenet.onnx mnist_png/out/testing/1/*.pngBy default, the setup runs on CPU only, not GPU, as could be seen by running htop. But by the magic of PyTorch, modifying the program to run on the GPU is trivial:and leads to a faster runtime, with less
cat << EOF | patch
diff --git a/run.py b/run.py
index 104d363..20072d1 100644
--- a/run.py
+++ b/run.py
@@ -24,7 +24,8 @@ data_test = MNIST('./data/mnist',
data_train_loader = DataLoader(data_train, batch_size=256, shuffle=True, num_workers=8)
data_test_loader = DataLoader(data_test, batch_size=1024, num_workers=8)
-net = LeNet5()
+device = 'cuda'
+net = LeNet5().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=2e-3)
@@ -43,6 +44,8 @@ def train(epoch):
net.train()
loss_list, batch_list = [], []
for i, (images, labels) in enumerate(data_train_loader):
+ labels = labels.to(device)
+ images = images.to(device)
optimizer.zero_grad()
output = net(images)
@@ -71,6 +74,8 @@ def test():
total_correct = 0
avg_loss = 0.0
for i, (images, labels) in enumerate(data_test_loader):
+ labels = labels.to(device)
+ images = images.to(device)
output = net(images)
avg_loss += criterion(output, labels).sum()
pred = output.detach().max(1)[1]
@@ -84,7 +89,7 @@ def train_and_test(epoch):
train(epoch)
test()
- dummy_input = torch.randn(1, 1, 32, 32, requires_grad=True)
+ dummy_input = torch.randn(1, 1, 32, 32, requires_grad=True).to(device)
torch.onnx.export(net, dummy_input, "lenet.onnx")
onnx_model = onnx.load("lenet.onnx")
EOFuser as now we are spending more time on the GPU than CPU:real 1m27.829s
user 4m37.266s
sys 0m27.562s Pinned article: Introduction to the OurBigBook Project
Welcome to the OurBigBook Project! Our goal is to create the perfect publishing platform for STEM subjects, and get university-level students to write the best free STEM tutorials ever.
Everyone is welcome to create an account and play with the site: ourbigbook.com/go/register. We belive that students themselves can write amazing tutorials, but teachers are welcome too. You can write about anything you want, it doesn't have to be STEM or even educational. Silly test content is very welcome and you won't be penalized in any way. Just keep it legal!
Intro to OurBigBook
. Source. We have two killer features:
- topics: topics group articles by different users with the same title, e.g. here is the topic for the "Fundamental Theorem of Calculus" ourbigbook.com/go/topic/fundamental-theorem-of-calculusArticles of different users are sorted by upvote within each article page. This feature is a bit like:
- a Wikipedia where each user can have their own version of each article
- a Q&A website like Stack Overflow, where multiple people can give their views on a given topic, and the best ones are sorted by upvote. Except you don't need to wait for someone to ask first, and any topic goes, no matter how narrow or broad
This feature makes it possible for readers to find better explanations of any topic created by other writers. And it allows writers to create an explanation in a place that readers might actually find it.
Figure 1. Screenshot of the "Derivative" topic page. View it live at: ourbigbook.com/go/topic/derivativeVideo 2. OurBigBook Web topics demo. Source. - local editing: you can store all your personal knowledge base content locally in a plaintext markup format that can be edited locally and published either:This way you can be sure that even if OurBigBook.com were to go down one day (which we have no plans to do as it is quite cheap to host!), your content will still be perfectly readable as a static site.
- to OurBigBook.com to get awesome multi-user features like topics and likes
- as HTML files to a static website, which you can host yourself for free on many external providers like GitHub Pages, and remain in full control
Figure 3. Visual Studio Code extension installation.
Figure 4. Visual Studio Code extension tree navigation.
Figure 5. Web editor. You can also edit articles on the Web editor without installing anything locally.Video 3. Edit locally and publish demo. Source. This shows editing OurBigBook Markup and publishing it using the Visual Studio Code extension.Video 4. OurBigBook Visual Studio Code extension editing and navigation demo. Source. 
- Infinitely deep tables of contents:
All our software is open source and hosted at: github.com/ourbigbook/ourbigbook
Further documentation can be found at: docs.ourbigbook.com
Feel free to reach our to us for any help or suggestions: docs.ourbigbook.com/#contact