Pretrained models requested about pytorch-retinanet HOT 16 CLOSED

@yhenon Thanks for your project. Great work!
I wonder why SGD doesn't work as the original paper said? Adam works well.
Do you have any ideas?

from pytorch-retinanet.

I have the same issue here.
Looking for coco pre-trained model too.
Thanks.

from pytorch-retinanet.

Hi,
I'm glad to hear you find this repo useful. I haven't shared weights yet as it's not quite finalized. However, I do have weights that I should be able to share later today.
@sunset1995 the regression targets are normalized slightly differently in this repo VS the keras one. This one uses (0.1,0.1,0.2,0.2), whereas the keras one uses (0.2,0.2,0.2,0.2). So the weights for the final layer of the regression branch have to be scaled appropriately.

from pytorch-retinanet.

@yhenon thanks~
I also found keras version regression output is to shift x1, y1, x2, y2 instead of x, y, w, h. The visualization from few samples look reasonable now after this fix.
However, I still more prefer to use your weight to keep code cleaner.

Thanks again~

from pytorch-retinanet.

from pytorch-retinanet.

@yhenon I will open a new branch for it in this weekend.
Also, our team is waiting your pretrained weight :)

from pytorch-retinanet.

from pytorch-retinanet.

The results from few samples look prefect! (better than my keras version, so some fix is needed for my keras conversed model)
btw, it seem that coco_resnet_50_map_0_335.pt is saved from python2
I have to use below trick to load it from python3

from functools import partial
import pickle
pickle.load = partial(pickle.load, encoding="latin1")
pickle.Unpickler = partial(pickle.Unpickler, encoding="latin1")
retinanet = torch.load(parser.model, map_location=lambda storage, loc: storage, pickle_module=pickle)

from pytorch-retinanet.

I see this line:
anchors_nms_idx = nms(torch.cat([transformed_anchors, scores], dim=2)[0, :, :], 0.5)
) that should apply nms to prediction detections too, but i get results like this:
.

Do you have any idea why this is happening?

from pytorch-retinanet.

I've uploaded a model that has been serialized differently: https://drive.google.com/open?id=1yLmjq3JtXi841yXWBxst0coAgR26MNBS

You should be able to load it with:

retinanet = model.resnet50(num_classes=dataset_train.num_classes(),)
retinanet.load_state_dict(torch.load(PATH))

@sunset1995 This should fix the issues between python2/3

@lvaleriu this may fix the problem you reported, I'd appreciate if you could test it out.

from pytorch-retinanet.

Same thing for me.. I get multiple anchors/detections for the same object

from pytorch-retinanet.

@yhenon hi, yhenon, could you explain a little bit that why you use (0.1,0.1,0.2,0.2) to normalize the regression targets but not use the original regression targets? It's because the regression loss is too small to be fully optimized? But in the original paper, it seems that the author doesn't normalize the regression targets, so I don't know if it's necessary.
Thanks.

from pytorch-retinanet.

What do you mean with original regression targets ? If you compute the COCO regression mean and std values you will more or less get the 0.1 0.1 0.2 0.2 that is used. The original paper doesnt mention regression normalization but it is safe to assume they do. Also, as far as I know, Detectron (which is the closest to an official implementation) uses these regression values as well.

from pytorch-retinanet.

@hgaiser for the original regression targets, i mean the targets in this line. And the purpose of doing regression normalization is to amplify the regression loss and to make it can be fully optimized, do i understand it correctly?

from pytorch-retinanet.

@hgaiser is correct, I'll just add a few things.

• the original source for doing this is the fast-rcnn paper, where they
  state:

We normalize the ground-truth regression targets vi to have zero mean and unit variance. All experiments use λ = 1.

• these targets were computed over the whole COCO dataset.

the purpose of doing regression normalization is to amplify the regression
loss and to make it can be fully optimized, do i understand it correctly?

• partly, it also ensures that the 2 parts of the regression loss is balanced
  (the size and position)

• the main reason to do it is that it improves mAP a little.

from pytorch-retinanet.

@yhenon it's clear to me now, thanks for your explanation.

from pytorch-retinanet.