Is it possible to provide the pretrain weight for deeplabv3 and FCN?
right now when i do:
model = torchvision.models.segmentation.fcn_resnet50(pretrained=True)
I receive this:
NotImplementedError: pretrained fcn_resnet50_coco is not supported as of now
Hi,
We do provide pre-trained weights, but for now only for ResNet 101 backbone.
Could you try that one instead?
@fmassa
Thanks for your reply.
I tried it. Resnet101 works fine, however, I cannot use it due to memory limitation.
Do you think that it would be possible to provide the weights for resnet 50 as well?
I could look into it after ECCV deadline, but I won't have bandwidth before that unfortunately :-/
I understand. Thank you @fmassa
@fmassa Just a gentle reminder regarding the pre-trained models for FCN50 :)
@isalirezag, if you are, like me, using a 4GB card and cannot start training because indeed calculating the gradients for the whole network is too much, you can train only the classifier part or the last layers of the backbone, that will work with 4GB.
@mthrok could you have a look into this?
I just launched the trainings for both fcn_resnet50 and deeplabv3_resnet50.
Let's keep the issue open until the new models are uploaded to torchvision
We got the following result;
| Model | Mean IoU | Global Accuracy |
| ------------- | ------------- | ------------- |
| fcn_resnet50 | 60.5 | 91.4 |
| deeplabv3_resnet50 | 66.4 | 92.4 |
We are preparing for the model release.
@mthrok can you tell me where i can find the pretrained models? right now I get an error when I want to load the pretrained version. I'm using the recent torch and torchvision.

@seyeeet you need a nightly version of torchvision for that, it is not available in the 0.6.0 release
@fmassa is there a way to install torchvision_nightly in conda?
conda install torchvision_nightly wont work
@seyeeet installation instructions are in https://pytorch.org/get-started/locally/
conda install pytorch torchvision -c pytorch-nightly
Hi @mthrok , thanks for the pretrained models. What testing script did you use to get mIoU 66.4 with deeplabv3_resnet50? I used this and got mIoU 65.3.
Hi @haichaoyu
I used the same function/script and the evaluation was performed while I was training the model.
You can see the training logs bellow; (the numbers are at the very bottom)
I used the most recent release version of PyTorch and Torchvision that were available in April 2020, so I believe they were Pytorch 1.4.0 and Torchvision 0.5.0.
@mthrok Thanks. I changed DataParallel to single gpu and use batch size 1. The accuracy is 66.4 now.
Most helpful comment
I just launched the trainings for both
fcn_resnet50anddeeplabv3_resnet50.