coco_train_script.py params to train YOLO V7 W6 model close to the original model provided

[Alexey234432]

Hi all,

I'd like to train YOLO V7 W6 from scratch to be close accuracy wise (and have the same architecture in terms of anchors etc) to the provided converted model. Does anyone know which params should I provide to coco_train_script.py script for that? I understand that original performance might not be achievable but I want to have a "reasonable" start.

Thank you very much for help.

[leondgarse]

Just try a basic run with totally 16 epochs:

python3 ./coco_train_script.py -p adamw -b 16 -i 512 -D yolov7.YOLOV7_W6 --rescale_mode raw01 \
--lr_decay_steps 15 --lr_cooldown_steps 1 --lr_warmup_steps 3

I've just finished 3 warmup epochs, and the Iou=0.50:0.95 is 0.053 at this point. Still rather low though, just it's coverging.

Ya, COCO training still cannot match with original version...

[Alexey234432]

Thank you very much for advice. I am currently experimenting with training with the aim to achieve something similar to the baseline (so we could have confidence when training new custom detectors - that we are getting "maximum out of architecture")

In particular, I tried the following -
coco_train_script.py -p AdamW -b 16 -i 512 -D yolov7.YOLOV7_W6 --freeze_backbone_epochs 0 --eval_start_epoch 0 --disable_float16 --rescale_mode raw01 --lr_decay_steps 32 --lr_cooldown_steps 1 --lr_warmup_steps 3

It is near the end with current Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.238 on iteration 32 which is significantly lower then the baseline.

I am now planning further experiments and looking for things to try. I think it would be interesting to try with lr_decay_steps of 60 or 90. Also weight_decay is 0.02 by default - seems like it is quite big (ie here https://blog.paperspace.com/train-yolov7-custom-data/ it is much smaller)? Curious if you have any other ideas?

Thank you very much for all the great work and help.

[Alexey234432]

(Forgot to mention that initial try with --lr_decay_steps 15 led to worse results. Thank you!)

[Alexey234432]

Just wanted to update, tried the following python ./keras_cv_attention_models-main/coco_train_script.py \ -p AdamW -b 16 -i 512 -D yolov7.YOLOV7_W6 \ --freeze_backbone_epochs 0 --eval_start_epoch 0 --disable_float16 --weight_decay 0.0005 \ --rescale_mode raw01 --lr_decay_steps 90 --lr_cooldown_steps 1 --lr_warmup_steps 3

as a result on iteration 75 got to Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.346 without much changes in AP per each epoch

Now trying with 180 epochs and looking for any ideas with which params to play with

Thank you

[leondgarse]

Thanks for your detailed results, and it's a good reference! Sorry for the late reply, just back from a vacation.
For COCO training, currently it's actually just basically working. Things like the data augmentation and loss weights also not matching...
I may take more tests if some spare time.