ModuleNotFoundError: No module named 'models.yolo' on custom trained weight. Is it a bug?

[dilwolf]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

After training the Yolov5 model on a custom dataset with yolov5/train.py in the repository for multi-gpu and loss_weight purposes, I am able to do inference with yolov5/detect.py in the repository. However, when I use the fine tuned weight with short code for customization, model is returning ModuleNotFoundError:. Below is my short Python script and error logging:

inference.py:

from ultralytics import YOLO
model = YOLO("models/best.pt")
results = model.predict(source="videos/video.mp4", show=True, save=True)

Error>>>>>:

WARNING ⚠️ models/best.pt appears to require 'models.yolo', which is not in ultralytics requirements.
AutoInstall will run now for 'models.yolo' but this feature will be removed in the future.
Recommend fixes are to train a new model using the latest 'ultralytics' package or to run a command with an official YOLOv8 model, i.e. 'yolo predict model=yolov8n.pt'
requirements: Ultralytics requirement ['models.yolo'] not found, attempting AutoUpdate...
ERROR: Could not find a version that satisfies the requirement models.yolo (from versions: none)
ERROR: No matching distribution found for models.yolo
Retry 1/2 failed: Command 'pip install --no-cache "models.yolo" ' returned non-zero exit status 1.
requirements: AutoUpdate success ✅ 2.7s, installed 1 package: ['models.yolo']
requirements: ⚠️ Restart runtime or rerun command for updates to take effect

Traceback (most recent call last):
File "/home/imes-server2/anaconda3/envs/dilwolf/lib/python3.10/site-packages/ultralytics/nn/tasks.py", line 732, in torch_safe_load
ckpt = torch.load(file, map_location="cpu")
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/torch/serialization.py", line 1025, in load
return _load(opened_zipfile,
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/torch/serialization.py", line 1446, in _load
result = unpickler.load()
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/torch/serialization.py", line 1439, in find_class
return super().find_class(mod_name, name)
ModuleNotFoundError: No module named 'models.yolo'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):

File "/home/user/Dilwolf/football_analysis/inference.py", line 3, in
model = YOLO('models/best_weighted.pt')
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/ultralytics/models/yolo/model.py", line 23, in init
super().init(model=model, task=task, verbose=verbose)
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/ultralytics/engine/model.py", line 151, in init
self._load(model, task=task)
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/ultralytics/engine/model.py", line 240, in _load
self.model, self.ckpt = attempt_load_one_weight(weights)
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/ultralytics/nn/tasks.py", line 806, in attempt_load_one_weight
ckpt, weight = torch_safe_load(weight) # load ckpt
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/ultralytics/nn/tasks.py", line 752, in torch_safe_load
ckpt = torch.load(file, map_location="cpu")
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/torch/serialization.py", line 1025, in load
return _load(opened_zipfile,
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/torch/serialization.py", line 1446, in _load
result = unpickler.load()
File "/home/user/anaconda3/envs/dilwolf/lib/python3.10/site-packages/torch/serialization.py", line 1439, in find_class
return super().find_class(mod_name, name)
ModuleNotFoundError: No module named 'models.yolo'

Tried restarting runtime and other installations, did not work....

Additional

When I trained with !yolo task=detect mode=train model=yolov5x.pt data=data.yaml epochs=100 imgsz=640 command on jupyter notebook, I am able to run fine tuned weight with my inference.py. Current ultralystics==8.2.11 I also tried with little bit previous version of ultralytics but it did not help.

Furthermore, is it possible use mult-gpu, when I train with !yolo task=detect ..... on jupter notebook? I tried !yolo -m torch.distributed.run nproc_per_node=2 train.py batch=64 data=data.yaml --weights yolov5s.pt --device 0,1 but did not work?

Thanks in advance !!!

[glenn-jocher]

Hi there! 👋 It looks like you're encountering issues related to model package dependencies. From the error logs, it appears that the models.yolo module is being mistakenly required, which shouldn't be the case with the standard setup.

Please ensure you are using the latest version of YOLOv5 from the Ultralytics GitHub repository. If you've trained your model with a custom setup or modified version of YOLO, make sure all your imports and custom modules are correctly aligned with the Ultralytics library structures.

For the multi-GPU training query, ensure you’re using the latest updates of YOLOv5 and PyTorch with correct syntax:

python -m torch.distributed.run --nproc_per_node=2 train.py --img 640 --batch 32 --epochs 100 --data data.yaml --weights yolov5s.pt --device 0,1

If issues persist, try retracing your steps or consider training a fresh model with the official scripts to see if the issue is ongoing. For further detailed guidance, please check the docs.

[dilwolf]

Hey there! Thank you so much for your help and effort on this.

Today, I retrained my model again with fresh source code by cloning it after getting your response. I reinstalled requirements.txt in the official Yolov5 repository, which includes ultralytics==8.2.13. Again, it is returning the same error.

It may seem very childish, but being meticulous may solve our error. Let me briefly explain what I did again today.

I started cloning the official repo and fine tuned the pre-trained yolov5x model on my dataset. I trained using yolov5/train.py from yolov5 official repository. However, when I use the trained weight with my short inference.py script, it's giving ModuleNotFoundError. I am curious it should work normally with the YOLO module when I call from ultralytics import YOLO. Below is my inference script:

inference.py:

from ultralytics import YOLO
model = YOLO("models/best.pt")
results = model.predict(source="videos/video.mp4", show=True, save=True)

Any help or suggestion are really appreciated !!!

[glenn-jocher]

Hello! 👋 Thank you for your detailed follow-up and the efforts you’ve put into resolving this issue.

The error you're encountering is quite peculiar, especially after following all the correct steps. Just to ensure that we're on the exact same page, the YOLO class is intended to work directly with the pre-trained models provided or ones trained using Ultralytics official scripts without modification.

Given your situation, here's a simple thing you can try:

• Make sure that you are importing everything correctly and from the expected directories. Occasionally, if there are nested directories or misspellings, Python can get tripped up on imports.

Here's a slightly modified script that makes sure the weights are being pointed to from the correct relative path:

from ultralytics import YOLO

# Assuming your 'best.pt' is in the 'models' directory relative to your script
model = YOLO("./models/best.pt")
results = model.predict(source="videos/video.mp4", show=True, save=True)

Additionally, if this continues to fail, could you check the version of YOLOv5 directly in your environment to make certain it aligns with your operations? Sometimes, discrepancies in versioning can lead to unexpected issues.

If the problem persists, feel free to consult our docs or respond here with any updates or further details. We'll find a solution together! 🚀

[dilwolf]

Thank you so much for your help on this issue!

As I mentioned earlier, I am using yolov5/train.py by customizing cls_pw: [15, 0.1, 2] in data/hyps/hyp.scratch-low.yaml file for weighting loss for imbalanced dataset. I can easily use my fine-tuned weight with predict.py. However, I can not use my fine-tuned weight with YOLO class. The YOLO class just works fine with default pre-trained weight in the original repository.

Today, I tried with latest version ultralytics==8.2.16 but today it has got newtype of Error:

Traceback (most recent call last):
File "/home/dilwolf/my_projects/yolov5/inference.py", line 5, in
results = model.predict(source="input_videos/video.mp4", save=True)
File "/home/dilwolf/anaconda3/envs/yolo/lib/python3.10/site-packages/ultralytics/engine/model.py", line 446, in predict
self.predictor.setup_model(model=self.model, verbose=is_cli)
File "/home/dilwolf/anaconda3/envs/yolo/lib/python3.10/site-packages/ultralytics/engine/predictor.py", line 297, in setup_model
self.model = AutoBackend(
File "/home/dilwolf/anaconda3/envs/yolo/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "/home/dilwolf/anaconda3/envs/yolo/lib/python3.10/site-packages/ultralytics/nn/autobackend.py", line 144, in init
model = model.fuse(verbose=verbose)
TypeError: BaseModel.fuse() got an unexpected keyword argument 'verbose'

I am not using any kind of verbose function.

inference.py:

from ultralytics import YOLO

# Assuming your 'best.pt' is in the 'models' directory relative to your script
model = YOLO("./models/best.pt")
results = model.predict(source="input_videos/video.mp4", save=True)

[dilwolf]

Thank you so much for your help on this issue!

As I mentioned earlier, I am using yolov5/train.py by customizing cls_pw: [15, 0.1, 2] in data/hyps/hyp.scratch-low.yaml file for weighting loss for imbalanced dataset. I can easily use my fine-tuned weight with predict.py. However, I can not use my fine-tuned weight with YOLO class. The YOLO class just works fine with default pre-trained weight in the original repository.

Today, I tried with latest version ultralytics==8.2.16 but today it has got newtype of Error:

Traceback (most recent call last):
File "/home/dilwolf/my_projects/yolov5/inference.py", line 5, in
results = model.predict(source="input_videos/video.mp4", save=True)
File "/home/dilwolf/anaconda3/envs/yolo/lib/python3.10/site-packages/ultralytics/engine/model.py", line 446, in predict
self.predictor.setup_model(model=self.model, verbose=is_cli)
File "/home/dilwolf/anaconda3/envs/yolo/lib/python3.10/site-packages/ultralytics/engine/predictor.py", line 297, in setup_model
self.model = AutoBackend(
File "/home/dilwolf/anaconda3/envs/yolo/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "/home/dilwolf/anaconda3/envs/yolo/lib/python3.10/site-packages/ultralytics/nn/autobackend.py", line 144, in init
model = model.fuse(verbose=verbose)
TypeError: BaseModel.fuse() got an unexpected keyword argument 'verbose'

I am not using any kind of verbose function.

inference.py:

from ultralytics import YOLO

# Assuming your 'best.pt' is in the 'models' directory relative to your script
model = YOLO("./models/best.pt")
results = model.predict(source="input_videos/video.mp4", save=True)

I checked that I can not do this customization using yolo CLI. Because I can use fine-tuned weights using yolo CLI. Or is there a way to do the above weighted_loss custimization using YOLO CLI?

[glenn-jocher]

Hi there! 👋

It looks like you're running into a fairly specific error after customizing and training your model. The error being thrown when using your custom-trained model with the YOLO class might indicate an incompatibility or an issue with how the custom weights were saved or loaded.

Since the issue arises specifically with customized training configurations and not with the default pre-trained weights, this suggests that the customization itself (i.e., weighting in the loss) might be modifying the model architecture or saved state in a way that isn't fully compatible with the current YOLO class handling.

Regarding the error with unexpected keyword argument 'verbose', it seems like there might have been changes in the Ultralytics library that are causing this issue. Our team continuously updates the library, so checking the compatibility of different versions might help.

A temporary solution would be to use predict.py as it works without issues. However, pointing out these discrepancies is crucial for us to update and fix potential bugs in future releases.

For now, unfortunately, there isn't a direct way to implement this specific weighted_loss customization using the YOLO CLI without diving into the code and modifying the requisite parts as you have done.

Thanks for bringing this to attention, and we appreciate your input as it helps improve the tool! Let us know if there are any further details or changes, and we'll try to assist further! 🚀

[github-actions]

👋 Hello there! We wanted to give you a friendly reminder that this issue has not had any recent activity and may be closed soon, but don't worry - you can always reopen it if needed. If you still have any questions or concerns, please feel free to let us know how we can help.

For additional resources and information, please see the links below:

• Docs: https://docs.ultralytics.com
• HUB: https://hub.ultralytics.com
• Community: https://community.ultralytics.com

Feel free to inform us of any other issues you discover or feature requests that come to mind in the future. Pull Requests (PRs) are also always welcomed!

Thank you for your contributions to YOLO 🚀 and Vision AI ⭐

[fardinabbasi]

Description

I have trained a custom object detection model using the original YOLOv5 repository (https://github.com/ultralytics/yolov5). The training was successful and produced a best.pt file.

When I attempt to load and use this best.pt file within the newer ultralytics framework (specifically using from ultralytics import YOLO), I encounter an error. The error message indicates that the model requires a component or format that is not found in the ultralytics package.

Steps to Reproduce

1. Clone and set up YOLOv5 repository:

   git clone https://github.com/ultralytics/yolov5
   cd yolov5
   pip install -r requirements.txt

2. Train a custom model with YOLOv5:

   python train.py --img 640 --batch 16 --epochs 50 --data custom_data.yaml --weights yolov5s.pt --name my_yolov5_model

   This successfully generates runs/train/my_yolov5_model/weights/best.pt.

3. Attempt to load the model using the ultralytics package:

   from ultralytics import YOLO
   
   # This assumes 'best.pt' is the path to the trained model
   model = YOLO('path/to/my_yolov5_model/weights/best.pt') 

Expected Behavior

I expected the ultralytics package to be able to load and run inference on a best.pt model file trained by the original YOLOv5 repository, as ultralytics is a successor and provides support for various YOLO versions.

Current Behavior

The attempt to load the model results in a ModuleNotFoundError and a warning about missing requirements:
WARNING models/best.pt appears to require 'models.yolo', which is not in Ultralytics requirements. AutoInstall will run now for 'models.yolo' but this feature will be removed in the future. Recommend fixes are to train a new model using the latest 'ultralytics' package or to run a command with an official Ultralytics model, i.e. 'yolo predict model=yolo11n.pt'

requirements: Ultralytics requirement ['models.yolo'] not found, attempting AutoUpdate... ERROR: Could not find a version that satisfies the requirement models.yolo (from versions: none) ERROR: No matching distribution found for models.yolo WARNING Retry 1/2 failed: Command 'pip install --no-cache-dir "models.yolo" ' returned non-zero exit status 1.

Possible Solutions/Suggestions

• Could there be a converter or an official method to migrate best.pt files from the original YOLOv5 repo to be compatible with the ultralytics framework?

Thank you for your help.

[pderrenger]

@fardinabbasi Thanks for the clear repro. YOLOv5 repo checkpoints (.pt) are Python pickles that reference YOLOv5 modules (e.g., models.yolo), so loading them with the ultralytics package isn’t guaranteed and can raise ModuleNotFoundError. Use the same repo for inference, or export from YOLOv5 and load the exported artifact in ultralytics, for example: in yolov5 run python export.py --weights runs/train/my_yolov5_model/weights/best.pt --include onnx then in ultralytics from ultralytics import YOLO; model = YOLO('runs/train/my_yolov5_model/weights/best.onnx'); model.predict(...). There’s no official converter from YOLOv5 .pt to ultralytics PyTorch; for a unified API going forward, train directly with the ultralytics package (YOLO11 recommended).

[fardinabbasi]

Thanks for the clear explanation! However, the whole point is to do detection tracking. I want to use results = model.track(frame, **tracking_params) task from ultralytics.
My finetuned YOLOv5 model cannot be replaced. I don't know what is the best way to do detection tracking considering that the model is finetuned on the old YOLOv5 repo.