fregu856 / deeplabv3 Goto Github PK

Hi

I have a question about the models. In the name you specify which ResNet you are using (how many layers) but also "OS_X". Looking at the code it seems to me it has to do with the number of skip-connection blocks. Is that correct?

Thank you very much!

Hi!

Very nice repo! I'm currently trying to integrate your model into our framework (https://github.com/DIVA-DIA/DeepDIVA, feel free to check it out!). However, when I load the provided weights for deeplabv3 I get the following error:

	size mismatch for aspp.conv_1x1_4.weight: copying a param with shape torch.Size([20, 256, 1, 1]) from checkpoint, the shape in current model is torch.Size([8, 256, 1, 1]).
	size mismatch for aspp.conv_1x1_4.bias: copying a param with shape torch.Size([20]) from checkpoint, the shape in current model is torch.Size([8]). (deeplabv3.py:50)

I am using exactly the Resnet (ResNet18_OS8) and the ASPP (no bottleneck) that you are using in your code. Do you know what could be causing this?

Thank you very much already in advance.

Cheers,
Linda

Hi,

For a custom dataset,

• I changed the number of classes to 7 in the deeplabv3.py file (my data has 8 classes).
• Eliminated the class_wieghts file requirement in the train.py file
• Put my data & labels into the same folder structure

I get the following error in the line loss = loss_fn(outputs, label_imgs):
IndexError: Target 78 is out of bounds.

Can someone point out where I am going wrong?
Thanks

Hi, @fregu856 ,

Thanks for releasing such a useful package using pytorch. I had practiced on eval_on_val_for_metrics.py as guided, and obtain the same metrics output as yours. I'm concerning about the big gap between your result and the official deeplabv3+ result.
The class IOU of yours is 59.3, while the official deeplabv3+ is reported as 82.1.
Could you list the difference regarding your implementation? Is the provided pre-trained model model_13_2_2_2_epoch_580.pth a very preliminary training result?

THX!

hello!

i have a question

in this line, why you use 20classes ? not 19 classes

Hi,

Can you please add a license to your repo? Otherwise people will not be able to use your project.

Thanks

In get_item, you resize image to 50% of original dimensions. Then, you randomly scale downscaled image to 70-200%. In the case of upsamplig to 200%, information about original image has been lost. Wouldn't it be best to work with original image if upscaling by factor of <100% ?

when i use resnet50-152 as the backbone to evaluate on my datasets , the shape doesn't match?

i train my own module,but train loss decrease to 0.17 and stop decreasing,what about yours

I want to know make the 2K image to 1024, then use random crop to make it to 255*255 shape in the train dataloader.Won't this have a bad effect on the results of image segmentation?

Hi,

How can I input a random sidewalk image and output a segmented image based on this model, after cloning your repository?

Thanks,
Jacob

I don't find the weight file 'class_weights.pkl',can you tell me have to get it

Hi:

I evaluated your pretrained model on cityscapes dataset and find that the performance is not that good. I haven't trained deeplabv3 using your scripts, but before that, I wondered if this is the best model trained by your scripts?(I think that you must have also adjust many parameters to achieve current performance) Thank you very much!

I'm trying to produce deeplabv3 paper's performance recently, so this question is important to me :)

Hi , Thank you for this code repository.
Does this implementation support Multi Grid method as discussed in the paper?

Using a random cropping method, what is the difference between training and testing？