Are the pertained weights frozen while training?

[prateekmalhotra-hover]

Hi! Great work! I just wanted to inquire in more detail whether, while training, you're freezing the old weights. Thanks!

[iperov]

of course no

[godsmokescrack]

I think the pretrained weights should be frozen (at least at first) or some of the advantage they give you will be erased. For instance, the images below show training/validation loss curves (on some small segmentation dataset) for three different scenarios. The best performance was given by freezing the pretrained knowledge.

1. Frozen weights on pretrained vgg11 encoder:

2. Unfrozen weights on pretrained encoder:

3. Randomly initialized (pretrained=False):

[rrryan2016]

Has the author already provided pretrained weights?