Yet-another-efficientdet-pytorch: Cls loss & Total loss too big

[ChuyiZhong]

Hi all,
I am training it on my own dataset with one class, but the cls loss and total loss seem strange to me. How to get this right?
Thanks in advance.

[zylo117]

what's the training parameter and command? Try update the latest code and train with --head_only True

[ChuyiZhong]

what's the training parameter and command? Try update the latest code and train with --head_only True

python train.py -c 4 -p crowdhuman --batch_size 8 --lr 1e-5 --num_epochs 100 \
--load_weights /home/fudan/zhongchuyi/Yet-Another-EfficientDet-Pytorch/weights/efficientdet-d4.pth

[zylo117]

I think there is something wrong with the dataset. Can you run the tutorial? Try training on the shape dataset. If cls loss drop to less than 10 in a few epochs, the code is fine. It's take about 5 minutes.

[ggaziv]

@ChuyiZhong see https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch/pull/252

[sourabhyadav]

Hi all,
I am training it on my own dataset with one class, but the cls loss and total loss seem strange to me. How to get this right?
Thanks in advance.

I am facing a similar issue. Class loss is in range of hundreds after 2 epochs.

[zylo117]

It's not normal. It should drop to under 1.0 real soon.
Please check your dataset.
Can you get a good result from the tutorial on your server?

[sourabhyadav]

It's not normal. It should drop to under 1.0 real soon.
Please check your dataset.
Can you get a good result from the tutorial on your server?

Here is the results after 10 iterations:

Step: 12709. Epoch: 9/10. Iteration: 1271/1271. Cls loss: 0.46582. Reg loss: 0.02957. Total loss: 0.49539: 100%|█| 1271/1271 [05:45<00:00,  3.68it/
Val. Epoch: 9/10. Classification loss: 0.53922. Regression loss: 0.03294. Total loss: 0.57216

For 1-class model. I have cross-checked my category_id = 1.

The command used for training is:

python3 train.py -c 1 -p pc_loockdown --data_path /home/object_det/datasets/  --batch_size 8 --lr 1e-5 --num_epochs 10 --head_only True --load_weights weights/efficientdet-d1.pth

@zylo117 Does it look normal?

[zylo117]

Yes, now it is.
BTW, did you update the latest code. In the previous version, head_only may not work as expected.

[sourabhyadav]

Yes I cloned today's repo:

One doubt still remaining:
Why my training classification loss started with 1000?

[Info] loaded weights: efficientdet-d1.pth, resuming checkpoint from step: 0
[Info] freezed backbone
Step: 499. Epoch: 0/10. Iteration: 500/1271. Cls loss: 1132.17944. Reg loss: 0.06527. Total loss: 1132.24475:  39%|▍| 499/1271 [02:16<03:11,  4.04icheckpoint...
Step: 999. Epoch: 0/10. Iteration: 1000/1271. Cls loss: 866.44366. Reg loss: 0.02792. Total loss: 866.47156:  79%|▊| 999/1271 [04:27<00:56,  4.78itcheckpoint...
Step: 1270. Epoch: 0/10. Iteration: 1271/1271. Cls loss: 901.25433. Reg loss: 0.02943. Total loss: 901.28375: 100%|█| 1271/1271 [05:37<00:00,  3.76
Val. Epoch: 0/10. Classification loss: 1058.99283. Regression loss: 0.03893. Total loss: 1059.03176
Step: 1499. Epoch: 1/10. Iteration: 229/1271. Cls loss: 693.96698. Reg loss: 0.03183. Total loss: 693.99878:  18%|▏| 228/1271 [01:05<04:42,  3.69itcheckpoint...
Step: 1999. Epoch: 1/10. Iteration: 729/1271. Cls loss: 616.92957. Reg loss: 0.03471. Total loss: 616.96429:  57%|▌| 728/1271 [03:21<02:08,  4.22itcheckpoint...
Step: 2499. Epoch: 1/10. Iteration: 1229/1271. Cls loss: 294.32889. Reg loss: 0.05017. Total loss: 294.37906:  97%|▉| 1228/1271 [05:36<00:11,  3.77checkpoint...
Step: 2541. Epoch: 1/10. Iteration: 1271/1271. Cls loss: 367.30658. Reg loss: 0.03857. Total loss: 367.34515: 100%|█| 1271/1271 [05:46<00:00,  3.67
Val. Epoch: 1/10. Classification loss: 385.38531. Regression loss: 0.03736. Total loss: 385.42267
Step: 2999. Epoch: 2/10. Iteration: 458/1271. Cls loss: 228.76134. Reg loss: 0.03491. Total loss: 228.79625:  36%|▎| 457/1271 [02:05<03:08,  4.32itcheckpoint...
Step: 3499. Epoch: 2/10. Iteration: 958/1271. Cls loss: 126.56357. Reg loss: 0.03671. Total loss: 126.60027:  75%|▊| 957/1271 [04:23<01:10,  4.43itcheckpoint...
Step: 3812. Epoch: 2/10. Iteration: 1271/1271. Cls loss: 92.13811. Reg loss: 0.03446. Total loss: 92.17256: 100%|█| 1271/1271 [05:44<00:00,  3.68it
Val. Epoch: 2/10. Classification loss: 115.62846. Regression loss: 0.03631. Total loss: 115.66477

@zylo117 Is this normal behaviour for this repo?

[zylo117]

It is. It might be caused by a lower lr.
For your case, I think 1e-5 is a little bit too low.
But considering you are doing great so far, don't change it.

[sourabhyadav]

@zylo117 Thanks for proactive replies.
I did the inference on my custom testset I find lower AP than pre-trained network? Pre-trained (80 class) person AP = 64.69% and fine-tuned (1-class) person AP = 30.26%.

Basically the person in the dataset is in lookdown or top-down view. I have only 10k training images. I trained with the above command.

Why this would have happened? My intuition is since it is the person class but from a different view it should have improved at least better than pre-trained model?

[zylo117]

if it's not overfitting like your logs say. it's underfitting.
Maybe it's not converged.
Maybe increase learning rate and keep training

[sourabhyadav]

Yeah, after 100 epochs AP results were improved.
Thanks for awesome repo.

[lzh18628137361]

Yeah, after 100 epochs AP results were improved.
Thanks for awesome repo.

after 100 epochs ,loss = ? ,thanks

[sourabhyadav]

Yeah, after 100 epochs AP results were improved.
Thanks for awesome repo.

after 100 epochs ,loss = ? ,thanks

@lzh18628137361 approx after 50-60 epochs the losses are in the range of: classification loss: 0.00256, Reg loss: 0.0018 .

[yhenon]

This issue is caused by the initialization of the classifier layer. In practice, it is desirable to initialize the classifier layer such that it predicts 0.01 for each box. Since the large majority of boxes are negative, this leads to a much lower loss.
Add the following to line 396 of efficientdet/model.py:

        self.header.pointwise_conv.conv.weight.data.fill_(0)
        self.header.pointwise_conv.conv.bias.data.fill_(-4.59)

because sigmoid(-4.59) = 0.01
This reduced the loss at epoch 0 from 50 to 2.