activatedgeek/LeNet-5

Good packaging! Tested on Ubuntu 22.10:

git clone https://github.com/activatedgeek/LeNet-5
cd LeNet-5
git checkout 95b55a838f9d90536fd3b303cede12cf8b5da47f
virtualenv -p python3 .venv
. .venv/bin/activate

# Their requirements.txt uses >= and some == are incompatible with our Ubuntu.
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 \
;

time python run.py

This throws a billion exceptions because we didn't start the visdom server, but never mind that.

The scrip does a fixed 15 epochs.

Output on P51:

real    2m10.262s
user    11m9.771s
sys     0m26.368s

The run also produces a lenet.onnx ONNX file, which is pretty neat, and allows us for example to visualize it on Netron:

Figure 1. 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)`
This doesn't seem to conform to the original LeNet-5 however?

activatedgeek/LeNet-5 use ONNX for inference

Now let's try and use the trained ONNX file for inference on some manually drawn images on GIMP:

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!

Figure 1. Number 9 drawn with mouse on GIMP by Ciro Santilli (2023)

We can try the code adapted from thenewstack.io/tutorial-using-a-pre-trained-onnx-model-for-inferencing/ at python/onnx_cheat/infer_mnist.py:

cd python/onnx_cheat
./infer_mnist.py lenet.onnx infer_mnist_9.png

and it works pretty well! The protram outputs:

as desired.

We can also try with images directly from Extract MNIST images.

for f in /home/ciro/git/mnist_png/out/testing/1/*.png; do echo $f; infer.py $f ; done

and the accuracy is great as expected.

activatedgeek/LeNet-5 run on GPU

By 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:

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")
EOF

and leads to a faster runtime, with less user as now we are spending more time on the GPU than CPU:

real    1m27.829s
user    4m37.266s
sys     0m27.562s

Residual neural network (2015, ResNet)

Interesting layer skip architecture thing.

Apparently destroyed ImageNet 2015 and became very very famous as such.

ResNet v1 vs v1.5

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 smallperformance drawback (~5% imgs/sec).

Convolutional neural network (CNN)

CNN convolution kernels are also learnt

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:

nn.Conv2d(1, 6, kernel_size=(5, 5))

as used for example at: activatedgeek/LeNet-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.

YOLO model

Object detection model.

You can get some really sweet pre-trained versions of this, typically trained on the COCO dataset.

Trained artificial neural network

Deep learning

Deep learning is the name artificial neural networks basically converged to in the 2010s/2020s.

It is a bit of an unfortunate as it suggests something like "deep understanding" and even reminds one of AGI, which it almost certainly will not attain on its own. But at least it sounds good.

Backpropagation

Video 1. What is backpropagation really doing? by 3Blue1Brown (2017) Source. Good hand-wave intuition, but does not describe the exact algorithm.

MLperf

mlcommons.org/en/ Their homepage is not amazingly organized, but it does the job.

Benchmark focused on deep learning. It has two parts:

training: produces a trained network
inference: uses the trained network

Furthermore, a specific network model is specified for each benchmark in the closed category: so it goes beyond just specifying the dataset.

Results can be seen e.g. at:

training: mlcommons.org/en/training-normal-21/
inference: mlcommons.org/en/inference-datacenter-21/

And there are also separate repositories for each:

E.g. on mlcommons.org/en/training-normal-21/ we can see what the the benchmarks are:

Dataset	Model
ImageNet	ResNet
KiTS19	3D U-Net
OpenImages	RetinaNet
COCO dataset	Mask R-CNN
LibriSpeech	RNN-T
Wikipedia	BERT
1TB Clickthrough	DLRM
Go	MiniGo

MLperf v2.1 ResNet

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 --accuracy

Last line of output on P51, which appears to contain the benchmark results

TestScenario.SingleStream qps=58.85, mean=0.0138, time=0.136, acc=62.500%, queries=8, tiles=50.0:0.0129,80.0:0.0137,90.0:0.0155,95.0:0.0171,99.0:0.0184,99.9:0.0187

where presumably qps means queries per second, and is the main results we are interested in, the more the better.

Running:

tools/make_fake_imagenet.sh

produces a tiny ImageNet subset with 8 images under fake_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 285

where the numbers are the category indices from ImageNet1k. At gist.github.com/yrevar/942d3a0ac09ec9e5eb3a see e.g.:

817: 'sports car, sport car',
89: 'sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita',

and so on, so they are coherent with the image names. By quickly looking at the script we see that it just downloads from Wikimedia and manually creates the file.

TODO 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:

./run_local.sh onnxruntime mobilenet cpu --accuracy

fails immediately with:

No such file or directory: '/path/to/coco/val2017-300/val_map.txt

The more plausible looking:

./run_local.sh onnxruntime mobilenet cpu --accuracy --dataset coco-300

first takes a while to preprocess something most likely, which it does only one, and then fails:

Traceback (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!

Run MLperf v2.1 ResNet on Imagenette

Let's run on this Imagenet10 subset, Imagenette.

First ensure that you get the dummy test data run working as per MLperf v2.1 ResNet.

Next, in the imagenette2 directory, first let's create a 224x224 scaled version of the inputs as required by the benchmark at mlcommons.org/en/inference-datacenter-21/:

#!/usr/bin/env bash
rm -rf val224x224
mkdir -p val224x224
for syndir in val/*: do
  syn="$(dirname $syndir)"
  for img in "$syndir"/*; do
    convert "$img" -resize 224x224 "val224x224/$syn/$(basename "$img")"
  done
done

and then let's create the val_map.txt file to match the format expected by MLPerf:

#!/usr/bin/env bash
wget https://gist.githubusercontent.com/aaronpolhamus/964a4411c0906315deb9f4a3723aac57/raw/aa66dd9dbf6b56649fa3fab83659b2acbf3cbfd1/map_clsloc.txt
i=0
rm -f val_map.txt
while IFS="" read -r p || [ -n "$p" ]; do
  synset="$(printf '%s\n' "$p" | cut -d ' ' -f1)"
  if [ -d "val224x224/$synset" ]; then
    for f in "val224x224/$synset/"*; do
      echo "$f $i" >> val_map.txt
    done
  fi
  i=$((i + 1))
done < <( sort map_clsloc.txt )

then back on the mlperf directory we download our model:

wget https://zenodo.org/record/4735647/files/resnet50_v1.onnx

and finally run!

DATA_DIR=/mnt/sda3/data/imagenet/imagenette2 time ./run_local.sh onnxruntime resnet50 cpu --accuracy

which gives on P51:

TestScenario.SingleStream qps=164.06, mean=0.0267, time=23.924, acc=87.134%, queries=3925, tiles=50.0:0.0264,80.0:0.0275,90.0:0.0287,95.0:0.0306,99.0:0.0401,99.9:0.0464

where qps presumably means "querries per second". And the time results:

446.78user 33.97system 2:47.51elapsed 286%CPU (0avgtext+0avgdata 964728maxresident)k

The time=23.924 is much smaller than the time executable because of some lengthy pre-loading (TODO not sure what that means) that gets done every time:

INFO:imagenet:loaded 3925 images, cache=0, took=52.6sec
INFO:main:starting TestScenario.SingleStream

Let's try on the GPU now:

DATA_DIR=/mnt/sda3/data/imagenet/imagenette2 time ./run_local.sh onnxruntime resnet50 gpu --accuracy

which gives:

TestScenario.SingleStream qps=130.91, mean=0.0287, time=29.983, acc=90.395%, queries=3925, tiles=50.0:0.0265,80.0:0.0285,90.0:0.0405,95.0:0.0425,99.0:0.0490,99.9:0.0512
455.00user 4.96system 1:59.43elapsed 385%CPU (0avgtext+0avgdata 975080maxresident)k

TODO lower qps on GPU!

Some examples:

activatedgeek/LeNet-5 produces a trained .onnx from PyTorch
MLperf v2.1 ResNet can use .onnx as a pre-trained model

NumPy does not automatically use the GPU for it: stackoverflow.com/questions/49605231/does-numpy-automatically-detect-and-use-gpu, and PyTorch is one of the most notable compatible implementations, as it uses the same memory structure as NumPy arrays.

Sample runs on P51 to observe the GPU speedup:

$ time ./matmul.py g 10000 1000 10000 100
real    0m22.980s
user    0m22.679s
sys     0m1.129s
$ time ./matmul.py c 10000 1000 10000 100
real    1m9.924s
user    4m16.213s
sys     0m17.293s

cd python/pytorch
wget -O resnet_demo_in.jpg https://upload.wikimedia.org/wikipedia/commons/thumb/6/60/Rooster_portrait2.jpg/400px-Rooster_portrait2.jpg
./resnet_demo.py resnet_demo_in.jpg resnet_demo_out.jpg

This first downloads the model, which is currently 167 MB.

We know it is COCO because of the docs: pytorch.org/vision/0.13/models/generated/torchvision.models.detection.fasterrcnn_resnet50_fpn_v2.html which explains that

FasterRCNN_ResNet50_FPN_V2_Weights.DEFAULT

is an alias for:

FasterRCNN_ResNet50_FPN_V2_Weights.COCO_V1

The runtime is relatively slow on P51, about 4.7s.

After it finishes, the program prints the recognized classes:

['bird', 'banana']

so we get the expected bird, but also the more intriguing banana.

By looking at the output image with bounding boxes, we understand where the banana came from!

Figure 2. python/pytorch/resnet_demo_out.jpg. The beak was of course a banana, not a beak!

Neural network

Artificial neural network (ANN)

ANN model (ANN architecture)

LeNet (1998, LeNet-5)

LeNet implementation

activatedgeek/LeNet-5

activatedgeek/LeNet-5 use ONNX for inference

activatedgeek/LeNet-5 run on GPU

Residual neural network (2015, ResNet)

ResNet implementation

ResNet variant

ResNet v1 vs v1.5

Convolutional neural network (CNN)

CNN convolution kernels are also learnt

YOLO model

Trained artificial neural network

Deep learning

Backpropagation

Epoch and batch size

Epoch (deep learning)

Batch size (deep learning)

Learning rate

Deep learning benchmark

MLperf

MLperf v2.1 ResNet

Run MLperf v2.1 ResNet on Imagenette

Deep learning is mostly matrix multiplication

Deep learning framework

ONNX

Netron

PyTorch

python/pytorch/matmul.py

PyTorch model

torchvision

torchvision ResNet

TensorFlow

TensorFlow quantum

 Ancestors

Discussion (0)

Discussion (0) Subscribe (1)

Discussion (0)