After using Focal Loss, the network does not converge

[pprp]

f1_gamma=0.5
alpha=0.5/0.25

we get the error below:

WARNING: non-finite loss, ending training  tensor([9.14797,     nan, 0.00000,     nan], device='cuda:0')

After I set the parameters as:

f1_gamma=2
alpha=0.25

The network works but fails to converge.

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    76/272     4.97G      2.39  2.73e-06         0      2.39        82       416: 100%|██████████████████████████████████████| 55/55 [00:14<00:00,  3.77it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:04<00:00,  1.09s/it]
                 all       391       409    0.0278     0.139   0.00872    0.0464

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    77/272     4.97G      2.35  2.73e-06         0      2.35        83       416: 100%|██████████████████████████████████████| 55/55 [00:15<00:00,  3.59it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:04<00:00,  1.02s/it]
                 all       391       409    0.0463     0.169    0.0249    0.0727

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    78/272     4.97G      2.36  2.71e-06         0      2.36        83       416: 100%|██████████████████████████████████████| 55/55 [00:15<00:00,  3.58it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:05<00:00,  1.40s/it]
                 all       391       409    0.0199     0.147   0.00453    0.0351

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    79/272     4.97G      2.35  2.72e-06         0      2.35        84       416: 100%|██████████████████████████████████████| 55/55 [00:14<00:00,  3.74it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:05<00:00,  1.29s/it]
                 all       391       409    0.0146     0.132   0.00409    0.0262

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    80/272     4.97G      2.33  2.71e-06         0      2.33        85       416: 100%|██████████████████████████████████████| 55/55 [00:15<00:00,  3.66it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:04<00:00,  1.03s/it]
                 all       391       409    0.0613     0.152    0.0397    0.0873

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    81/272     4.97G      2.35  2.74e-06         0      2.35        83       416: 100%|██████████████████████████████████████| 55/55 [00:14<00:00,  3.68it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:05<00:00,  1.46s/it]
                 all       391       409    0.0137     0.112   0.00248    0.0244

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    82/272     4.97G      2.36  2.72e-06         0      2.36        80       416: 100%|██████████████████████████████████████| 55/55 [00:15<00:00,  3.65it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:05<00:00,  1.40s/it]
                 all       391       409    0.0159     0.115   0.00383    0.0279

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    83/272     4.97G      2.33  2.78e-06         0      2.33        77       416: 100%|██████████████████████████████████████| 55/55 [00:15<00:00,  3.59it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:05<00:00,  1.31s/it]
                 all       391       409    0.0288     0.174    0.0126    0.0495

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    84/272     4.97G      2.34  2.74e-06         0      2.34        99       416: 100%|██████████████████████████████████████| 55/55 [00:15<00:00,  3.59it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:05<00:00,  1.40s/it]
                 all       391       409    0.0225     0.147   0.00658     0.039

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    85/272     4.97G      2.34  2.73e-06         0      2.34        86       416: 100%|██████████████████████████████████████| 55/55 [00:15<00:00,  3.65it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:03<00:00,  1.03it/s]
                 all       391       409    0.0492     0.149    0.0127     0.074

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    86/272     4.97G      2.32  2.78e-06         0      2.32        79       416: 100%|██████████████████████████████████████| 55/55 [00:14<00:00,  3.78it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|████████████████████████████████████████| 4/4 [00:04<00:00,  1.04s/it]
                 all       391       409    0.0303     0.139   0.00757    0.0498

what's more, I only have one class and I use the command below:

python train.py --cfg cfg/yolov3-tiny.cfg --arc Fdefault 

[glenn-jocher]

@pprp there is about zero obj loss in your second example, so obviously the network will never learn obj this way.

[glenn-jocher]

@pprp also, if focal loss produces worse results, then clearly don't use it.

[pprp]

What should I do if i want use focal loss？

[glenn-jocher]

@pprp try different settings.

[pprp]

Thank you very much. I will try to fix this problem..

[glenn-jocher]

@pprp by the way, I was looking at the focal loss function. I think the reduction setting may need an update now that the loss reduction functions are set to sum rather than mean, so there may be a bug here that is our fault. I'll try to push an update today.

[glenn-jocher]

@pprp ok, the fix is done in 189c704

Can you git pull and try training again, starting from the default focal loss parameters?

[pprp]

Thanks for your reply, I will retrain tomorrow and inform you of the final result.

[pprp]

@glenn-jocher I try the fixed version but get the same problem.
I use your default focal loss parameters:

f1_gamma=0.5
alpha=1

if I use Fdefault, the network will get non-finite loss error.

if I use uFBCE, the network does not converge.

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
     9/272     4.98G      3.01     0.228         0      3.24        76       416: 100%|███████████████████████████████████████████████████████████████████████████| 76/76 [00:20<00:00,  3.78it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|█████████████████████████████████████████████████████████████████████████████| 4/4 [01:36<00:00, 24.12s/it]
                 all       391       409   0.00201     0.902   0.00185   0.00401

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    10/272     4.98G      2.99     0.219         0      3.21        71       416: 100%|███████████████████████████████████████████████████████████████████████████| 76/76 [00:21<00:00,  3.54it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|█████████████████████████████████████████████████████████████████████████████| 4/4 [01:35<00:00, 23.93s/it]
                 all       391       409   0.00203     0.914    0.0019   0.00406

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    11/272     4.98G       2.9     0.213         0      3.11        80       416: 100%|███████████████████████████████████████████████████████████████████████████| 76/76 [00:20<00:00,  3.70it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|█████████████████████████████████████████████████████████████████████████████| 4/4 [01:37<00:00, 24.30s/it]
                 all       391       409   0.00205     0.922   0.00195   0.00409

     Epoch   gpu_mem      GIoU       obj       cls     total   targets  img_size
    12/272     4.98G      2.83     0.193         0      3.03        86       416: 100%|███████████████████████████████████████████████████████████████████████████| 76/76 [00:20<00:00,  3.65it/s]
               Class    Images   Targets         P         R   mAP@0.5        F1: 100%|█████████████████████████████████████████████████████████████████████████████| 4/4 [01:36<00:00, 24.16s/it]
                 all       391       409   0.00203     0.912   0.00191   0.00405

[glenn-jocher]

@pprp ah ok. Well, it seems focal loss is not the best choice for your problem. I recommend you stick to the repo defaults (i.e. --arc default). They are the defaults for a reason.

[FranciscoReveriano]

From experience @pprp Focal Loss is usually not the best way to go. I don't know what you are training on. But I would recommend either increasing the img-size, lowering the initial learning rate by a magnitude of 10, or lowering the training IoU.

[pprp]

In my problem, I want to use focal loss to balance the positive samples and negative samples.

I have a question about lobj.

In compute_loss function:

BCEobj = nn.BCEWithLogitsLoss(pos_weight=ft([h['obj_pw']]), reduction=red)

giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False,
                            GIoU=True)  # giou computation

tobj[b, a, gj, gi] = giou.detach().type(tobj.dtype)

 lobj += BCEobj(pi[..., 4], tobj)

Can you tell me why to calculate the loss between the output and the giou? Does this have an effect on the focal loss?

[glenn-jocher]

@pprp this is experimental. I think we will revert back to the original formulation below, we are currently testing the effect of the change. Focal loss is independent of this though.

tobj[b, a, gj, gi] = 1.0

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove Stale label or comment or this will be closed in 5 days.