You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Thank you for your interest in our project! We are actually still in the development phase of the YOLO model. We have yet to match the accuracy of the paper when trained from scratch.
It is hard to tell what the issue is without looking at your build.py, but I am assuming that you are trying to train YOLOv3 from the pretrained weights given in the original paper.
We have a configuration for training YOLOv3, but I do not remember where the original weights came from. We will be able to give you more detailed information when we completely reproduce the paper and clean up the repository, but you can definitely use the weights from the author's website by changing csdarknet53 in the config to yolov3.
The build.py is the building instruction in the repo. I have attached an image with respect to that. . Also, I am trying to train (using Yolov4) from scratch on an input image with a different number of channels than the original paper
I would still recommend holding off on training with our model implementation since we are not done.
If you do want to train the model from scratch using our repo, you can make modifications to the yolov4.yaml experiment that we provide and use the command that I gave earlier to train. The weights for the CSPDarknet model that we start from can be downloaded here. You need to adjust the number of training steps and validation steps for your experiment. For most datasets that have a similar distribution as COCO, you can scale them by the number of images in the dataset. An example for how to use the model for inference is in the inference.ipynb file.
For now, if you want to have a model that matches the paper's accuracy, I would suggest training using darknet until we finish the YOLO model.
do we need to run the build.py to build the model before training or we can use your .yaml files to start the training?
When I run the build.py, I got the same error as mentioned above.
Hello,
I think the .yaml will be used to start the training. From the previousse
messages.
python3 -m yolo.train_vm --mode=train_and_eval --experiment=yolo_custom
--model_dir=checkpoints/yolov3
--config_file=yolo/configs/experiments/yolov3/yolov3.yaml
oh sorry, I did not come across this error. But I can remember that I
changed the file path of "--
*config_file=yolo/configs/experiments/yolov3/yolov3.yaml" *from the
python3 -m yolo.train_vm --mode=train_and_eval --experiment=yolo_custom
--model_dir=checkpoints/yolov3
--config_file=yolo/configs/experiments/yolov3/yolov3.yaml
I hope the authors reply soon for a solution
Activity
anivegesana commentedon Mar 31, 2021
Hi @Tomysinee,
Thank you for your interest in our project! We are actually still in the development phase of the YOLO model. We have yet to match the accuracy of the paper when trained from scratch.
It is hard to tell what the issue is without looking at your build.py, but I am assuming that you are trying to train YOLOv3 from the pretrained weights given in the original paper.
We have a configuration for training YOLOv3, but I do not remember where the original weights came from. We will be able to give you more detailed information when we completely reproduce the paper and clean up the repository, but you can definitely use the weights from the author's website by changing
csdarknet53in the config toyolov3.If you are trying to train YOLOv4 or something else, feel free to reply.
Tomysinee commentedon Mar 31, 2021
Thank you so much for your reply.
The build.py is the building instruction in the repo. I have attached an image with respect to that. . Also, I am trying to train (using Yolov4) from scratch on an input image with a different number of channels than the original paper
anivegesana commentedon Apr 1, 2021
Hi @Tomysinee,
I would still recommend holding off on training with our model implementation since we are not done.
If you do want to train the model from scratch using our repo, you can make modifications to the
yolov4.yamlexperiment that we provide and use the command that I gave earlier to train. The weights for the CSPDarknet model that we start from can be downloaded here. You need to adjust the number of training steps and validation steps for your experiment. For most datasets that have a similar distribution as COCO, you can scale them by the number of images in the dataset. An example for how to use the model for inference is in the inference.ipynb file.For now, if you want to have a model that matches the paper's accuracy, I would suggest training using darknet until we finish the YOLO model.
kalikhademi commentedon Jun 11, 2021
do we need to run the build.py to build the model before training or we can use your .yaml files to start the training?
When I run the build.py, I got the same error as mentioned above.
Tomysinee commentedon Jun 11, 2021
kalikhademi commentedon Jun 11, 2021
do we need to make changes in the .yaml file. I have used the train command mentioned above and it resulted in this error.
Tomysinee commentedon Jun 11, 2021