YOLOX on Kaggle - full custom model training notebook

[rkinas]

I created and published (probably first) custom model training for Kaggle competitions. Here you can find full pipeline: https://www.kaggle.com/remekkinas/yolox-full-training-pipeline-for-cots-dataset

[FateScript]

Good job :)

[rkinas]

From Epoch 17 ... to Epoch 56 ... no changes in IoUs stats .... :(

[GOATmessi7]

Hi, I've seen your training log and I think the IoU loss is high. Maybe you should close the mixup strategy and adjust the "no_aug_epochs" to a small number according to your max epochs. And we also support rectangle input size now, so you can keep the original 720x1280 resolution.
Hope this helps

[rkinas]

Now we have second part as well: https://www.kaggle.com/remekkinas/yolox-inference-on-kaggle-for-cots
It is time to ... build great model .... :)

[KeremTurgutlu]

I think not all rectangle resolutions work, for example with (720,1280) gives error:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-39-65fc47b58164> in <module>
----> 1 trainer.model.backbone(inp.cuda())

/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1100         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1101                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102             return forward_call(*input, **kwargs)
   1103         # Do not call functions when jit is used
   1104         full_backward_hooks, non_full_backward_hooks = [], []

/trainman-mount/trainman-k8s-storage-5db6d360-c9b6-4daf-8039-164efd47b892/cots/YOLOX/yolox/models/yolo_pafpn.py in forward(self, input)
     97         fpn_out0 = self.lateral_conv0(x0)  # 1024->512/32
     98         f_out0 = self.upsample(fpn_out0)  # 512/16
---> 99         f_out0 = torch.cat([f_out0, x1], 1)  # 512->1024/16
    100         f_out0 = self.C3_p4(f_out0)  # 1024->512/16
    101 

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 46 but got size 45 for tensor number 1 in the list.

however input size (800,1280) works fine - should be fine though since it will be padded:

[rkinas]

I am training now :) and experiment. We will see :)

[rkinas]

I know I know this is to early .... for full happines but ..... YoloX is the best scored notebook currently in this competition https://www.kaggle.com/c/tensorflow-great-barrier-reef/leaderboard I know that everything will change soon but now ... I can fully confirm that full Kaggle training (I update this notebook) and inference pipeline is ready. Hope now people will more often use YoloX in competitions.

[rkinas]

Now even better .... LB 0.507 😍😍😍 -> TOP11 and the best public notebook score!!! YoloX .... WoW

[dsantiago]

Good job, always making wonderful work on kaggle notebooks

[rkinas]

Thank you! :)

[rkinas]

Could anybody help me to build in Albumentations (it could be one function eg. Cutout) in training dataloader (hook for Albumentations)?

[rkinas]

I managed to write custom Albumentation Data Loader class :) Works fine!

[jasper-vaart]

Could you maybe show how you managed to get it working?

I tried adjusting TrainTransform (https://github.com/Megvii-BaseDetection/YOLOX/blob/6880e3999eb5cf83037e1818ee63d589384587bd/yolox/data/data_augment.py) by commenting out the flip and hsv augs; and calling albumentations instead.
That makes it trainable, however when I visualize the augmentations, the bounding boxes are all wrong.