Use for custom dataset

[MZeinab]

Search before asking

• I have searched the Supervision issues and found no similar bug report.

Bug

Thanks, it was so amazing.
I have one question. I'll be so grateful if you could help:

1. I watched the video that explained how to train yolov5 with the custom dataset. I used that notebook and exported the best.pt file, containing model weights. My question is: how to use it for inference?

• I used model = torch.hub.load('ultralytics/yolov5', 'custom', path='best.pt') # PyTorch
  but It loads the original .pt file, not my custom one.

Environment

colab

Minimal Reproducible Example

No response

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!

[github-actions[bot]]

Hello there, thank you for opening an Issue ! 🙏🏻 The team was notified and they will get back to you asap.

[SkalskiP]

Hello there as it is not really a problem with Supervision, I'll convert that issue into discussion.

[SkalskiP]

@MZeinab did you try to pass a global path to the weights file?

[MZeinab]

what do you mean by global path?
I use google colab. I uploaded my custom weight file "best.pt" and used it in
torch.hub.load('ultralytics/yolov5', 'custom', path='best.pt') .
Anything more should I do?
I think the problem is with 'ultralytics/yolov5'.
Do you have any idea? I'll be so thankful if you could help.

[SkalskiP]

Hi, @MZeinab 👋🏻!

I think the problem is with ultralytics/yolov5.

It is possible, but before we create an issue on their repository, let's try to pass the global path. Let me try to explain. Let's try to imagine this directory structure (I just made it up to give you an example):

content/
├─ yolov5/
├─ weights/
│  ├─ best.pt

Now if you are inside the weights directory path to your best.pt can be expressed in the form of a global and relative paths.

• global /content/weights/best.pt
• relative best.pt

Problem is that when you are inside yolov5 directory relative path is no longer best.pt but ../best.pt. That's why it is always better to use global paths. They are longer but stay the same regardless of your active directory.

I can help you to resolve your global path to best.pt file if you upload a screenshot of your directory structure here. ;)

[MZeinab]

Hi @SkalskiP
I just wanted to let you know that I checked what you explained. I used the global path.

• This is my best.pt file:

file = "/content/best.pt" model_real = torch.load(file) model_real
The output is:

{'epoch': -1, 'best_fitness': None, 'model': Model(
(model): Sequential(
(0): Conv(
(conv): Conv2d(3, 32, kernel_size=(6, 6), stride=(2, 2), padding=(2, 2), bias=False)
(bn): BatchNorm2d(32, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU(inplace=True)
)
(1): Conv(
(conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=0.001, momentum=0.03, affine=True, track_running_stats=True)
(act): SiLU(inplace=True)
)
........... (I didn't add the remain of it)

• I compared it with the code I mentioned earlier :

model = torch.hub.load('ultralytics/yolov5', 'custom', path="/content/best.pt") model
the output is:

Using cache found in /root/.cache/torch/hub/ultralytics_yolov5_master
YOLOv5 🚀 2023-2-18 Python-3.8.10 torch-1.13.1+cu116 CUDA:0 (Tesla T4, 15110MiB)

Fusing layers...
Model summary: 213 layers, 7012822 parameters, 0 gradients, 15.8 GFLOPs
Adding AutoShape...
AutoShape(
(model): DetectMultiBackend(
(model): DetectionModel(
(model): Sequential(
(0): Conv(
(conv): Conv2d(3, 32, kernel_size=(6, 6), stride=(2, 2), padding=(2, 2))
(act): SiLU(inplace=True)
)
(1): Conv(
(conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
(act): SiLU(inplace=True)
)

you can see that the batch normalization layers are not in the custom model (model_real) and It seems that the code uses the original model of ultralytics

[SkalskiP]

torch.hub.load might load just selected layers. Did you try to run the actual inference?

[MZeinab]

yeah, I think so.
How can I use the best.pt file to convert it to model.
I used
`file = "/content/best.pt"

model_real = torch.load(file)`

and runned this part:

`import supervision as sv

extract video frame

generator = sv.get_video_frames_generator('/content/Multi_Ablated_25a25c_Exp21_6day.mp4')
iterator = iter(generator)
frame = next(iterator)

detect

results = model_real(frame, size=1280)
detections = sv.Detections.from_yolov5(results)

annotate

box_annotator = sv.BoxAnnotator(thickness=4, text_thickness=4, text_scale=2)
frame = box_annotator.annotate(scene=frame, detections=detections)

%matplotlib inline
sv.show_frame_in_notebook(frame, (16, 16))`

but It gets error:
'dict' object is not callable

so my question is how to define model_real based on my best.pt file to be callable?

[SkalskiP]

@MZeinab, but did you actually try to infer using torch.hub.load or just loaded?

If you want to load YOLOv5 model not via torch.hub.load I think your best choice is to use YOLOv5 pip package.

[MZeinab]

yes, I tried to infer the model defined with torch.hub.load.
but the detection result was not good,
It detects some bounding boxes that are not the desired object.

so after that when I compared this 2 .pt file (my custom model and the yolov5 acheived with torch.hub.load ) I found out that they are different.
and after that I defined my custom model with model_real = torch.load(file) and try to infer but I get error that 'dict' object is not callable