How can I train the YOLOX and Feature aggregation module seperately?

[hahapt]

Hello, I'm working on if I can replace the YOLOX with YOLOv5, so I'm finding how can I train the YOLOX and Feature aggregation module seperately?

[YuHengsss]

For training YOLOX, you can configure files in exps/yolov/yoloxs_vid.py, download the coco pretrain model and run the script in tools/train.py to train YOLOX. Then, YOLOV, you can simply run: python tools/vid_train.py -f exps/yolov/yolov_s.py -c weights/[path to your yolox model] --fp16

[YuHengsss]

Hello, I'm working on if I can replace the YOLOX with YOLOv5, so I'm finding how can I train the YOLOX and Feature aggregation module seperately?

For more details on training YOLOX, please refer to https://github.com/Megvii-BaseDetection/YOLOX

[hahapt]

So I can train YOLOX with tools/train.py , and then train the feature aggregation module with python tools/vid_train.py -f exps/yolov/yolov_s.py -c weights/[path to your yolox model] --fp16? So how can I change the training strategies like ''When integrating the feature aggrega- tion module into the base detectors, we fine-tune them for 150K iterations with batch size of 16 on a single 2080Ti GPU''?

[YuHengsss]

So I can train YOLOX with tools/train.py , and then train the feature aggregation module with python tools/vid_train.py -f exps/yolov/yolov_s.py -c weights/[path to your yolox model] --fp16? So how can I change the training strategies like ''When integrating the feature aggrega- tion module into the base detectors, we fine-tune them for 150K iterations with batch size of 16 on a single 2080Ti GPU''?

Sorry, we find the coco annotation for training has the wrong path and we have updated it. For training base detector yolox, you just make a "annotations" directory in the dataset path and copy the annotation files to it or change the code in coco.py. Then, run the training script. Thanks for helping us debug.

[YuHengsss]

So I can train YOLOX with tools/train.py , and then train the feature aggregation module with python tools/vid_train.py -f exps/yolov/yolov_s.py -c weights/[path to your yolox model] --fp16? So how can I change the training strategies like ''When integrating the feature aggrega- tion module into the base detectors, we fine-tune them for 150K iterations with batch size of 16 on a single 2080Ti GPU''?

Hello, I'm working on if I can replace the YOLOX with YOLOv5, so I'm finding how can I train the YOLOX and Feature aggregation module seperately?

YOLOv5 is an anchor-based detector, which means you have to change the feature selection strategy. For changing the training process, please refer to our code for more details.

[hahapt]

So I can train YOLOX with tools/train.py , and then train the feature aggregation module with python tools/vid_train.py -f exps/yolov/yolov_s.py -c weights/[path to your yolox model] --fp16? So how can I change the training strategies like ''When integrating the feature aggrega- tion module into the base detectors, we fine-tune them for 150K iterations with batch size of 16 on a single 2080Ti GPU''?

Hello, I'm working on if I can replace the YOLOX with YOLOv5, so I'm finding how can I train the YOLOX and Feature aggregation module seperately?

YOLOv5 is an anchor-based detector, which means you have to change the feature selection strategy. For changing the training process, please refer to our code for more details.

Sorry I'm a little confuse. I think the idea of aggregate features after prediction head can be helpful, but why anchor-based detector need to change the feature selection strategy?

[YuHengsss]

So I can train YOLOX with tools/train.py , and then train the feature aggregation module with python tools/vid_train.py -f exps/yolov/yolov_s.py -c weights/[path to your yolox model] --fp16? So how can I change the training strategies like ''When integrating the feature aggrega- tion module into the base detectors, we fine-tune them for 150K iterations with batch size of 16 on a single 2080Ti GPU''?

Hello, I'm working on if I can replace the YOLOX with YOLOv5, so I'm finding how can I train the YOLOX and Feature aggregation module seperately?

YOLOv5 is an anchor-based detector, which means you have to change the feature selection strategy. For changing the training process, please refer to our code for more details.

Sorry I'm a little confuse. I think the idea of aggregate features after prediction head can be helpful, but why anchor-based detector need to change the feature selection strategy?

Because multiple bounding boxes may correspond to the same feature. We have not tried this practice and look forward to yours.

[hahapt]

Fine, I found it's not easy to directly try it on yolov5. So can you share the code based on ppyoloe?