Executing model produces raw output

[bmharper]

When I run this code:

model = yolov5.load("yolov5m6.pt")
y = model(x)

Where x is a tensor with shape (8, 3, 1280, 1280)
Then y is a tensor with shape (8, 102000, 10)
It looks to me like I'm getting the raw output of the model, but the documentation says that one should be getting the post-processed output here.

[fcakyon]

Your model output sizes seem to be not normal. Which version are you using?

https://github.com/fcakyon/yolov5-pip#use-from-python

[bmharper]

I'm using the latest YOLOv5 master branch from https://github.com/ultralytics/yolov5 (commit 32a92185738c93e5f0b0f6971de0812cd6fd5f34)

Do I need to train using this pip package?

[bmharper]

My yolov5 pip version is 6.2.3, so I am using the latest.
My code looks exactly like the code in the readme, i.e.:

import yolov5

# load pretrained model
model = yolov5.load('yolov5s.pt')
#....
# perform inference
results = model(img)

But my results do not match what the readme describes.

[fcakyon]

i am trying but not able to recreate the issue, can you provide me some sample weight file so that I can recreate the issue?

[fcakyon]

continuous integration tests are also passing:

yolov5-pip/tests/test_yolov5.py

Line 148 in 3abe7cf

self.assertEqual(len(results.pred[0]), 4)

[fcakyon]

have you also tried 6.2.2? Is the same error present with older versions?

[bmharper]

My model is too big... I'll train a smaller one and upload it here. I'll also try 6.2.2 and let you know....

[bmharper]

Hi,
I've found the cause - it is these lines in common.py:

The key issue is that when you feed AutoShape a tensor, then it does not run the post-processing NMS etc. I'll query the canonical YOLOv5 repo for why this decision exists, or if they'd be open to a flag that controls this behaviour.

[fcakyon]

ok thanks for pointing that out 👍

[bmharper]

I have opened a PR ultralytics/yolov5#10139 which allows one to perform post-processing on Tensor inputs. I'll comment here again if that PR gets merged.

[fcakyon]

thanks for the update 👍