[Pre-trained models] What is the network architecture of the pretrained model?

[ghost]

Hello,

Thank you for your code. I want to apply your network directly on another dataset without having to retrain it. The available pretrained model corresponds to which architecture :

networks = {
'resnet101_baseline': get_resnet101_baseline,
'resnet101_base_oc_dsn': get_resnet101_base_oc_dsn,
'resnet101_pyramid_oc_dsn': get_resnet101_pyramid_oc_dsn,
'resnet101_asp_oc_dsn': get_resnet101_asp_oc_dsn,
}

Thank you for your time.

[TouqeerAhmad]

Hi, were you able to figure out which specific model is saved in the .pth file shared by the author. By looking at the .sh files, it seems like it should work with all of the architectures above.

However, when I load the .pth file and try to initialize the model, I get the following RuntimeError, for the 'baseline' model:

RuntimeError: Error(s) in loading state_dict for ResNet:
Missing key(s) in state_dict: "layer5.0.weight", "layer5.0.bias", "layer5.1.weight", "layer5.1.bias", "layer5.1.running_mean", "layer5.1.running_var", "layer6.weight", "layer6.bias".
Unexpected key(s) in state_dict: "fc.weight", "fc.bias".

Thanks,
Touqeer

[TouqeerAhmad]

And by looking further into the keys of the shared .pth file, I can clearly see there are no layer5 and layer6. Figure on the following link shows the actual keys in the .pth file:

https://drive.google.com/file/d/1sazPfvDDCPg3gQ9lRY3IE6tHmG7hT2O4/view?usp=sharing

Is only the resnet101 trained with ImageNet being shared and not essentially the model trained for parsing? If that is the case it should be clearly stated in the introduction.

[swjtulinxi]

我一直多都在关注你们的工作，感觉t想法挺好的，你们的文章我都在跟。
其中就是那个ocnet中的pramid_oc_block模块一直没看懂，到底是不是像pspnet中的金字塔pooling那样，先对输入x通过池化降采样成不同大小，然后再每一个后面跟base oc，但是我看你的代码似乎又不是那样，请指教一下，感谢。

[PkuRainBow]

是在2x2/3x3/6x6的不同划分的sub-regions上单独用base-oc，没有down-sample操作。比如对于2x2的划分，那么就是在4个region上分别用base-oc。我们论文的图可能不够清楚。

[PkuRainBow]

@MattToul Thanks for your interest in our work and we would like to release the related checkpoints in the other repo openseg.pytorch ASAP.

[PkuRainBow]

@TouqeerAhmad Currently, we have not released the checkpoints of the segmentation models. Sorry for the inconvenience and we will release the checkpoints once our work is accepted.

[swjtulinxi]

是在2x2/3x3/6x6的不同划分的sub-regions上单独用base-oc，没有down-sample操作。比如对于2x2的划分，那么就是在4个region上分别用base-oc。我们论文的图可能不够清楚。

你的意思就是在局部进行baseoc，那么你的2x2的4个region是不是就是4个像素，，那你们是在这四个像素内部进行baseoc还是将这4个像素与其他所有4个像素进行baseoc

[swjtulinxi]

我明白了，是不是就像是一个正方形分成4个等分，左上右上左下右下，这样就在4个区域进行baseoc避免在计算时一个像素要计算一整张图所有像素对他的影响，这样就可以起到降低分辨率的作用，而且还可以从不同感受野来计算。 感谢， 突然都很难想象这个代码你们是怎么写出来的，大佬啊

[PkuRainBow]

@swjtulinxi Please refer the code for pyramid-oc, we first divide the input to 2x2 / 3x3 / 6x6 sub-regions, then we apply the base-oc, and finally we concatenate their results.