Performance issue for NYUv2

[kcs6568]

Thank you for your excellent contribution!

I have a question about the performance of MTAN with PCGrad.

In my experiment, the result of NYUv2 is very different from the PCGrad official performance (although I not used UW in this experiment)

(sorry for this image to very small)

Also, when I experimented to the resnet50-HPS without weighting method, the result is beyond the MTAN official performance.

I tried your official training command line in all cases and used the dataset in your Dropbox.

Is that result the right phenomenon?

Thank you in advance!

[Baijiong-Lin]

The SegNet is used as the backbone network in the original PCGrad paper. While we use the DeepLabV3+ (ResNet50 with dilated convolution + ASPP) as the backbone here, which achieves better performance than SegNet.

[kcs6568]

Thank you for your answer!

Is there SegNet architecture in this library?

I couldn't find it in the LibMTL model directory.

Is not supported in the current version?

Or, is the same architecture when I tried the supported command-line below for the original MTAN with SegNet? python train_nyu.py --weighting DWA --arch MTAN --gpu_id 5 --scheduler step --train_bs 4

[Baijiong-Lin]

LibMTL does not support the SegNet backbone network. MTAN is orthogonal to the backbone network, which means MTAN can combine with both SegNet and ResNet50.

[kcs6568]

Thank you!!