ucbdrive/dla

Code for the CVPR 2018 Oral Paper "Deep Layer Aggregation"

dla's Issues

Did you use pretrained models for training all fine-grained datasets?

In your paper, figure 6 showed all results for fine-grained datasets. However, when I tried to reproduce the results in bird dataset, I can only get 53% acc. Did you use any pre-trained model for training that? And can you post the code that can be helpful to reproduce your paper's result? Best regards.

Eigenvalue and vector for new dataset

Hi, I'm trying to use dla with my own dataset but I don't really understand the eigenvalues and eigenvectors. How are they calculated/what does "image pixel variations" mean?

Thank you!

The .project weight in '.pth' file seems mismatch with the model?

I've taken a look at the '.pth' file and I found there are two .project.weight when the depth of a tree is 2.
This is for dla34:


According to the figure in your paper and your code:

self.project comes after downsample and appears only once during a single level, but why here are two .project.weight in the pth?
Hope for your reply soon.

About Model Modules

Is the function in the code corresponding to the module in the picture like this? But after I finished writing, I found that the size of my model changed from 70M to about 7M in the original. However, the dLA34 in the paper is only 1.3M. Is there something wrong with my understanding？

Visualize the model

Here is how to visualize the model

from tensorboardX import SummaryWriter
import torch
from dla import dla34

model = dla34(pretrained=None)
image = torch.rand(2, 3, 224, 224)
x = model(image)
with SummaryWriter(comment='LinearInLinear') as w:
    w.add_graph(model, image, True)
model.eval()

Residual use in Tree

In the forward method of the Tree class in dla.py, the residual argument is always ignored (overriden line 209).
Is it what is wanted ?
If it is, then building project in the constructor and calling its forward method it line 209 is useless when levels > 1. Some computation could be saved.
Also bottom computation is not used if levels > 1 and level_root is false.

aggregation and BasicBlock do not work for all image sizes (which are passed in forward pass).

Are the pretrained model weights in RGB or BGR order?

Does the pre-trained DLA model expect input images in BGR or RGB channel order? Thanks.

dl.yf.io is down

The link http://dl.yf.io/dla/models/ to download pretrained models seems to be down.

initialization

there is a bug in DLA_UP that you forget initial parameters of proj in IDAUP?

License

Could you include a license for this repo? Thank you!

size mismatch when load dla60

Difference between IDA_up and DLA_up

The paper mentions DLA, IDA and HDA.

But how are DLAUp and IDAUp in your code different? Can you provide some reference please?

Thanks.

DLA-X46C Backbone

Hello,

Is this repo maintained? I had a few questions regarding the DLA-X46C backbone?

Thanks.

slides

hello, where can I find the slides of this paper?

bilinear initialization of deconv layer

Hi,
I am not quite sure whether the following initialization of deconv layer is a valid implementation of bilinear filter for all input feature channels:

def fill_up_weights(up):
    w = up.weight.data
    f = math.ceil(w.size(2) / 2)
    c = (2 * f - 1 - f % 2) / (2. * f)
    for i in range(w.size(2)):
        for j in range(w.size(3)):
            w[0, 0, i, j] = \
                (1 - math.fabs(i / f - c)) * (1 - math.fabs(j / f - c))
    for c in range(1, w.size(0)):
        w[c, 0, :, :] = w[0, 0, :, :]

It doesn't initialize w in dim 1 except for the first input channel.

A problem with pretrain dla model

Thanks for your great work.

Some users reported that the pretrained model links failed. Please help fix the problem?

Owen-Liuyuxuan/visualDet3D#73

Hello, I hope you can provide the pre-training model.Thank you

Hello, I want to apply your proposed model to other fields, and I need a pre-trained model on Image Net. My experimental equipment cannot support me to retrain a model. Could you please provide me with a pre-trained model?Thank you

dl.yf.io is down and cannot load models

Hi, as of February 9th, 2024. It seems that the server for http://dl.yf.io/dla/models is down again. Please check if possible to get it back.

the performance on object detection?

hello
do you have do some experiments to validate the effectiveness on object detection?