sweetytian / efficientdet Goto Github PK

Thanks for your code. But after I replaced fpn with your bifpn in libra rcnn, I got lower map. Did you get higher map in faster rcnn?

how to confing necks in mmdetection?

when I run the example code

    >>> from mmdet.models import EfficientNet
    >>> import torch
    >>> self = EfficientNet(model_name='tf_efficientnet_b2', pretrained=False)
    >>> self.eval()
    >>> inputs = torch.rand(1,3,768,768)
    >>> level_outputs = self(inputs)

it is error for this:
outs.append(feature_map[i])
IndexError: index 2 is out of bounds for dimension 0 with size 1

RT.

Thanks for you share.

efficientnet中B0 backbone如下所示:

其中stage 2, stage 5并没有下采样，而在efficientdet中采用了最后5个stage, 都进行了下采样

请问这是什么原因?

I found it for a long time, but I didn't find out how to train. Is there no train code and I need to write it myself? Or I didn't find the train interface, If so, can you please give a train and valid tutorial? Thank you very much!

thank you for your code, and i want to run rightly them

but i can't find two files, Could you tell me how to get them?

Did you get a good results? I train the coco ,and get map=0.

I found your weighted Feature Fusion Module is try to new a tensor as weight, That will make this Tensor be a constant and will not in training.

That parameter may be got from global average polling or similar operation.

I use your config to train coco, and the loss convergences at about 1.6 and map of coco is 0...

I am debuging the issues....