If I reduce the image size Will the model become smaller?

[leeyunhome]

❔Question

Hello, @glenn-jocher

!python train.py --img 416 --batch 3 --epochs 500
If I reduce the image size
Will the model become smaller?

It was small when I experimented, but I don't know exactly, so I ask.

Thank you.

Additional context

[glenn-jocher]

@leeyunhome no, the model size is not a function of --img-size, they are independent.

[leeyunhome]

@leeyunhome no, the model size is not a function of --img-size, they are independent.

Hello,

When the size of the .pt file is small, I have experienced a faster inference speed.
If the image-size is smaller, the size of the .pt file is also smaller

The unit is MB.

Thank you.

[glenn-jocher]

@leeyunhome reduced image sizes will lead to reduced inference time. Model size is constant and not a function of any inference parameters.

[leeyunhome]

@leeyunhome reduced image sizes will lead to reduced inference time. Model size is constant and not a function of any inference parameters.

Hello,

In addition to resizing the image, what affects the inference spped
Is it just the use of spcecialized engines like tensorRT?

Thank you.

[glenn-jocher]

@leeyunhome using smaller models, smaller images, and batched inference will all speed up your inference time per image. Exporting to TensorRT should also speed up inference.

[leeyunhome]

@leeyunhome using smaller models, smaller images, and batched inference will all speed up your inference time per image. Exporting to TensorRT should also speed up inference.

@glenn-jocher
The smallest model is yolov5s
Batch inference means that you bundle multiple images and process them in batches?

[glenn-jocher]

@leeyunhome yes. If you use batch size of 16 for example then you should see faster speeds per image than running batch size 1 in a for loop for each image.

[leeyunhome]

@leeyunhome yes. If you use batch size of 16 for example then you should see faster speeds per image than running batch size 1 in a for loop for each image.

@glenn-jocher

When training, I set the batch size to 16
The batch size given as a factor affects the learning speed, right?

When inferring, do I have to set the batch size separately?

[glenn-jocher]

@leeyunhome training results are robust to variations in batch size (#2377) detect.py runs at --batch 1 for the most part. You can explore batch size speedups in test.py:

python test.py --data coco128.yaml --weights yolov5s.pt --batch 1
python test.py --data coco128.yaml --weights yolov5s.pt --batch 16
python test.py --data coco128.yaml --weights yolov5s.pt --batch 64