Can't achieve widerface val score shown in the README when training from scratch. about libfacedetection.train

[xwu6614555]

def adjust_learning_rate_poly(optimizer, iteration, max_iter):
"""Sets the learning rate
# Adapted from PyTorch Imagenet example:
# https://github.com/pytorch/examples/blob/master/imagenet/main.py
"""
lr = initial_lr * ( 1 - (iteration / max_iter)) * ( 1 - (iteration / max_iter))
if ( lr < 1.0e-7 ):
lr = 1.0e-7
return lr

learning_rate is not adjusted at all!
add these codes before return lr：
for param_group in optimizer.param_groups:
param_group['lr'] = lr
Maybe this can solve the problem

[waries]

def adjust_learning_rate_poly(optimizer, iteration, max_iter):
"""Sets the learning rate

Adapted from PyTorch Imagenet example:

https://github.com/pytorch/examples/blob/master/imagenet/main.py

"""
lr = initial_lr * ( 1 - (iteration / max_iter)) * ( 1 - (iteration / max_iter))
if ( lr < 1.0e-7 ):
lr = 1.0e-7
return lr

learning_rate is not adjusted at all!
add these codes before return lr：
for param_group in optimizer.param_groups:
param_group['lr'] = lr
Maybe this can solve the problem

Thanks for pointing that out, I'll try that and post the new result later.

[waries]

After modifying the lr code, the final result with scales=[1.], confidence_threshold=0.3 is
Easy Val AP: 0.8285190221051939
Medium Val AP: 0.7929441126273071
Hard Val AP: 0.5663795815749137

[ShiqiYu]

the learning rate is adapted according to epoches, not iterations

[ShiqiYu]

the learning rate is adapted according to epoches, not iterations

After modifying the lr code, the final result with scales=[1.], confidence_threshold=0.3 is
Easy Val AP: 0.8285190221051939
Medium Val AP: 0.7929441126273071
Hard Val AP: 0.5663795815749137

If you increase batch size to 256, and then fine-tune the mode, you will get a better result.

[waries]

the learning rate is adapted according to epoches, not iterations

After modifying the lr code, the final result with scales=[1.], confidence_threshold=0.3 is
Easy Val AP: 0.8285190221051939
Medium Val AP: 0.7929441126273071
Hard Val AP: 0.5663795815749137

If you increase batch size to 256, and then fine-tune the mode, you will get a better result.

Thank you for your guidance, I'll try it and post the new result. About fine-tune, does the initial lr need to be set to a smaller value?

[ShiqiYu]

You can use the same lr, but start from the 200th epoch.

[waries]

You can use the same lr, but start from the 200th epoch.

Thank you so much for your guidence！

[waries]

After fine-tune, the new result is
Easy Val AP: 0.8321736396427939
Medium Val AP: 0.7988401596758341
Hard Val AP: 0.579757106591874
I see professor Yu update the new result yesterday. So even though my new result is better than the original score which is tested on the model without fine-tune, there still exist a huge differences in the final result between the fine-tune model I trained and the official fine-tune model.
I train the first edition model with batchsize 8 and 500 epochs, then use the final model to fine-tune with batch size 256 and start at epoch 200/500. I'm not sure which step was wrong，maybe the best way is to use pretrained model directly, LOL.

You can use the same lr, but start from the 200th epoch.