how to load custom model for inference

[e101sg]

I am new to PyTorch and training for custom object detection. After model created , trying to load from local folder.

#model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True)
        
model = torch.hub.load('/Users/macmini04/Documents/yolov5','custom', path='best.pt',force_reload=True,source='local', pretrained =Flase)

The above code does not create bounding box on the output video file. I am following this code below and modify it, to load local model. Not sure, what i am missing. Any thoughts Welcome 👍 .
https://github.com/akash-agni/Real-Time-Object-Detection/blob/main/Drone_Human_Detection_Model.py

[glenn-jocher]

@e101sg you have a typo in your arguments, Flase.

In any case, I don't know why you would attempt to pass pretrained=False if you are expecting trained results.

[e101sg]

oh! it is mistake. i have corrected later but still not detection.

[glenn-jocher]

@e101sg just follow the YOLOv5 PyTorch Hub tutorial:

YOLOv5 Tutorials

• Train Custom Data  🚀 RECOMMENDED
• Tips for Best Training Results  ☘️ RECOMMENDED
• Weights & Biases Logging  🌟 NEW
• Supervisely Ecosystem  🌟 NEW
• Multi-GPU Training
• PyTorch Hub  ⭐ NEW
• TorchScript, ONNX, CoreML Export 🚀
• Test-Time Augmentation (TTA)
• Model Ensembling
• Model Pruning/Sparsity
• Hyperparameter Evolution
• Transfer Learning with Frozen Layers  ⭐ NEW
• TensorRT Deployment

[Abhinay-Sadineni]

oh! it is mistake. i have corrected later but still not detection.

is this resolved?

[Abhinay-Sadineni]

oh! it is mistake. i have corrected later but still not detection.

is this resolved?

[ghost]

The following code works for me.

import torch
model = torch.hub.load('.', 'custom', path='/path/to/yolov5/runs/train/exp5/weights/best.pt', source='local') 
# Image
img = '/path/to/test/image/25.jpg'
# Inference
results = model(img)
# Results, change the flowing to: results.show()
results.show()  # or .show(), .save(), .crop(), .pandas(), etc

[tudoanh]

Can I use yolov5 without yolov5 director? because while I am deploying model then hubconf.py and experimental.py files of the folder shows high severity warning.

Of course. You may try to fetch a pretrained model from github like this:

model = torch.hub.load("ultralytics/yolov5", "yolov5s", pretrained=True)
ckpt = torch.load(model_path)
model.load_state_dict(ckpt['model'].state_dict())

[kessmith]

model = torch.hub.load('.', 'custom', path='/path/to/yolov5/runs/train/exp5/weights/best.pt', source='local') This is not working. Got error FileNotFoundError: [Errno 2] No such file or directory: '.\hubconf.py'

The first param ("." in this case) should be point to your local yolov5 directory

Do I need to point to a Yolov5 directory to get this code to work?

[zkytony]

@tudoanh wait, how to do this for a custom model then? I want to load a pretrained custom model saved in a local .pt file.

Previously, torch.hub.load('ultralytics/yolov5', 'custom', path='path/to/model.pt') would just work. But now, I am getting all sorts of errors trying to make it work.

[ammarsaf]

Your code is so helpful @Matrix246, thanks! I just want to know, based on the last line that results.show() # or .show(), .save(), .crop(), .pandas(), etc, is there any option that I can save the bbox text from the output?

[BGAKSHAYA]

@Ammar-Azman The below code gives the bounding box coordinates, class label and the confidence score for the first image.
results.pandas().xyxy[0]

Sample Output:

          xmin       ymin        xmax        ymax  confidence  class  name
0   720.201294   4.749833  722.742065  256.000000    0.984483      0  dog
1   739.679504   8.926133  742.159546  256.000000    0.953222      0  dog

For xywh format, you can use results.pandas().xywh[0]

[Brunettow]

I don't want to use custom. However this doesn't work either
model = torch.hub.load('/home/project/yolov5', 'yolov5x6', path='/home/project/yolov5/best.pt', source='local')

[pinnintipraneethkumar]

Hi, @Brunettow you are using it in a wrong way insted of 'yolov5x6' use 'custom' it will work without any problem. you are trying to laod a custom model so you should use custom.

model = torch.hub.load('/home/project/yolov5', 'custom', path='/home/project/yolov5/best.pt', source='local')