Real Time Inference on Raspberry Pi 4 and 5 (40 fps!)#

Prerequisites#

To follow this tutorial you’ll need a Raspberry Pi 4 or 5, a camera for it and all the other standard accessories.

• Raspberry Pi 4 Model B 2GB+

• Raspberry Pi Camera Module

• Heat sinks and Fan (optional but recommended)

• 5V 3A USB-C Power Supply

• SD card (at least 8gb)

• SD card read/writer

Raspberry Pi Setup#

PyTorch only provides pip packages for Arm 64bit (aarch64) so you’ll need to install a 64 bit version of the OS on your Raspberry Pi

You’ll need to install the official rpi-imager to install Rasbperry Pi OS.

32-bit Raspberry Pi OS will not work.

Installation will take at least a few minutes depending on your internet speed and sdcard speed. Once it’s done it should look like:

Time to put your sdcard in your Raspberry Pi, connect the camera and boot it up.

# This enables the extended features such as the camera.
start_x=1

# This needs to be at least 128M for the camera processing, if it's bigger you can just leave it as is.
gpu_mem=128

Installing PyTorch and picamera2#

PyTorch and all the other libraries we need have ARM 64-bit/aarch64 variants so you can just install them via pip and have it work like any other Linux system.

$ sudo apt install -y python3-picamera2 python3-libcamera
$ pip install torch torchvision --break-system-packages

We can now check that everything installed correctly:

$ python -c "import torch; print(torch.__version__)"

Video Capture#

Test the camera is working first, by running libcamera-hello in a terminal.

For video capture we’re going to be using picamera2 to capture the video frames.

The model we’re using (MobileNetV2) takes in image sizes of 224x224 so we can request that directly from picamera2 at 36fps. We’re targeting 30fps for the model but we request a slightly higher framerate than that so there’s always enough frames.

from picamera2 import Picamera2

picam2 = Picamera2()

# print available sensor modes
print(picam2.sensor_modes)

config = picam2.create_still_configuration(main={
    "size": (224, 224),
    "format": "BGR888",
}, display="main")
picam2.configure(config)
picam2.set_controls({"FrameRate": 36})
picam2.start()

To capture the frames we can call capture_image to return a PIL.Image object that we can use with PyTorch.

# read frame
image = picam2.capture_image("main")

# show frame for testing
image.show()

This data reading and processing takes about 3.5 ms.

Image Preprocessing#

We need to take the frames and transform them into the format the model expects. This is the same processing as you would do on any machine with the standard torchvision transforms.

from torchvision import transforms

preprocess = transforms.Compose([
    # convert the frame to a CHW torch tensor for training
    transforms.ToTensor(),
    # normalize the colors to the range that mobilenet_v2/3 expect
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
input_tensor = preprocess(image)
# The model can handle multiple images simultaneously so we need to add an
# empty dimension for the batch.
# [3, 224, 224] -> [1, 3, 224, 224]
input_batch = input_tensor.unsqueeze(0)

Model Choices#

There’s a number of models you can choose from to use with different performance characteristics. Not all models provide a qnnpack pretrained variant so for testing purposes you should chose one that does but if you train and quantize your own model you can use any of them.

We’re using mobilenet_v2 for this tutorial since it has good performance and accuracy.

Raspberry Pi 4 Benchmark Results:

MobileNetV2: Quantization and JIT#

For optimal performance we want a model that’s quantized and fused. Quantized means that it does the computation using int8 which is much more performant than the standard float32 math. Fused means that consecutive operations have been fused together into a more performant version where possible. Commonly things like activations (ReLU) can be merged into the layer before (Conv2d) during inference.

The aarch64 version of pytorch requires using the qnnpack engine.

import torch
torch.backends.quantized.engine = 'qnnpack'

For this example we’ll use a prequantized and fused version of MobileNetV2 that’s provided out of the box by torchvision.

from torchvision import models
net = models.quantization.mobilenet_v2(pretrained=True, quantize=True)

We then want to jit the model to reduce Python overhead and fuse any ops. Jit gives us ~30fps instead of ~20fps without it.

net = torch.jit.script(net)

Putting It Together#

We can now put all the pieces together and run it:

import time

import torch
from torchvision import models, transforms
from picamera2 import Picamera2

torch.backends.quantized.engine = 'qnnpack'

picam2 = Picamera2()

# print available sensor modes
print(picam2.sensor_modes)

config = picam2.create_still_configuration(main={
    "size": (224, 224),
    "format": "BGR888",
}, display="main")
picam2.configure(config)
picam2.set_controls({"FrameRate": 36})
picam2.start()

preprocess = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

net = models.quantization.mobilenet_v2(pretrained=True, quantize=True)
# jit model to take it from ~20fps to ~30fps
net = torch.jit.script(net)

started = time.time()
last_logged = time.time()
frame_count = 0

with torch.no_grad():
    while True:
        # read frame
        image = picam2.capture_image("main")


        # preprocess
        input_tensor = preprocess(image)

        # create a mini-batch as expected by the model
        input_batch = input_tensor.unsqueeze(0)

        # run model
        output = net(input_batch)
        # do something with output ...
        print(output.argmax())

        # log model performance
        frame_count += 1
        now = time.time()
        if now - last_logged > 1:
            print(f"{frame_count / (now-last_logged)} fps")
            last_logged = now
            frame_count = 0

Running it shows that we’re hovering at ~30 fps on a Raspberry Pi 4 and ~41 fps on a Raspberry Pi 5.

This is with all the default settings in Raspberry Pi OS. If you disabled the UI and all the other background services that are enabled by default it’s more performant and stable.

If we check htop we see that we have almost 100% utilization.

To verify that it’s working end to end we can compute the probabilities of the classes and use the ImageNet class labels to print the detections.

top = list(enumerate(output[0].softmax(dim=0)))
top.sort(key=lambda x: x[1], reverse=True)
for idx, val in top[:10]:
    print(f"{val.item()*100:.2f}% {classes[idx]}")

mobilenet_v3_large running in real time:

Detecting an orange:

Detecting a mug:

Troubleshooting: Performance#

PyTorch by default will use all of the cores available. If you have anything running in the background on the Raspberry Pi it may cause contention with the model inference causing latency spikes. To alleviate this you can reduce the number of threads which will reduce the peak latency at a small performance penalty.

torch.set_num_threads(2)

For shufflenet_v2_x1_5 using 2 threads instead of 4 threads increases best case latency to 72 ms from 60 ms but eliminates the latency spikes of 128 ms.

Next Steps#

You can create your own model or fine tune an existing one. If you fine tune on one of the models from torchvision.models.quantized most of the work to fuse and quantize has already been done for you so you can directly deploy with good performance on a Raspberry Pi.

See more:

• Quantization for more information on how to quantize and fuse your model.

• Transfer Learning Tutorial for how to use transfer learning to fine tune a pre-existing model to your dataset.