gene-chou / efpn-detectron2 Goto Github PK

Hello! First of all, I think your idea is very good. I read your blog, and I want to Recurrence results myself, but I didn't find the pretraining model. Can you provide the pre training model on Baidu cloud.

Which loss function should I use?

Can I use Ce loss?

def cross_entropy(input, target, *, reduction="mean", **kwargs):
"""
Same as torch.nn.functional.cross_entropy, but returns 0 (instead of nan)
for empty inputs.
"""
if target.numel() == 0 and reduction == "mean":
return input.sum() * 0.0 # connect the gradient
return F.cross_entropy(input, target, **kwargs)

Which loss function should I use?

Can I use Ce loss?

def cross_entropy(input, target, *, reduction="mean", **kwargs):
"""
Same as torch.nn.functional.cross_entropy, but returns 0 (instead of nan)
for empty inputs.
"""
if target.numel() == 0 and reduction == "mean":
return input.sum() * 0.0 # connect the gradient
return F.cross_entropy(input, target, **kwargs)

Nice work.
I am attracted by your work, and I am trying to replace my model's original FPN by EFPN, which is developed with MMdetection.
Is it able to realize it by mmdetection?THis module will be integrated into mmdetection?
Thanks.

Is this code applicable for running on custom datasets? Will there be any problem caused by the loss function?

Thanks for the repo，but i have a problem about training。
When I training with the faster_rcnn_X_101_32x8d_FPN_3x,it always out of the memory（TITAN X which i used only has 12G memory）
so i want to change it with faster_rcnn_R_101_FPN_3x
That‘s what i set
cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_101_FPN_3x.yaml"))
cfg.MODEL.WEIGHTS = ""
cfg.MODEL.RESNETS.NUM_GROUPS = 1 # 1 ==> ResNet; > 1 ==> ResNeXt
cfg.MODEL.RESNETS.WIDTH_PER_GROUP = 8
But it has the error of wrong input size：
RuntimeError: Given groups=1, weight of size [32, 512, 1, 1], expected input[2, 16, 208, 208] to have 512 channels, but got 16 channels instead
Am i setting wrong？
What should i modifiy